From bff8d0ce42db9207de8d0c880e30c2daf036750c Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 09:27:50 -0500 Subject: [PATCH 01/10] Fix swinir arch name --- .../{swinir_arch.py => swinir_model_arch.py} | 1734 ++++++++--------- 1 file changed, 867 insertions(+), 867 deletions(-) rename modules/{swinir_arch.py => swinir_model_arch.py} (97%) diff --git a/modules/swinir_arch.py b/modules/swinir_model_arch.py similarity index 97% rename from modules/swinir_arch.py rename to modules/swinir_model_arch.py index a5eb9a36..461fb354 100644 --- a/modules/swinir_arch.py +++ b/modules/swinir_model_arch.py @@ -1,867 +1,867 @@ -# ----------------------------------------------------------------------------------- -# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257 -# Originally Written by Ze Liu, Modified by Jingyun Liang. -# ----------------------------------------------------------------------------------- - -import math -import torch -import torch.nn as nn -import torch.nn.functional as F -import torch.utils.checkpoint as checkpoint -from timm.models.layers import DropPath, to_2tuple, trunc_normal_ - - -class Mlp(nn.Module): - def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): - super().__init__() - out_features = out_features or in_features - hidden_features = hidden_features or in_features - self.fc1 = nn.Linear(in_features, hidden_features) - self.act = act_layer() - self.fc2 = nn.Linear(hidden_features, out_features) - self.drop = nn.Dropout(drop) - - def forward(self, x): - x = self.fc1(x) - x = self.act(x) - x = self.drop(x) - x = self.fc2(x) - x = self.drop(x) - return x - - -def window_partition(x, window_size): - """ - Args: - x: (B, H, W, C) - window_size (int): window size - - Returns: - windows: (num_windows*B, window_size, window_size, C) - """ - B, H, W, C = x.shape - x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) - windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) - return windows - - -def window_reverse(windows, window_size, H, W): - """ - Args: - windows: (num_windows*B, window_size, window_size, C) - window_size (int): Window size - H (int): Height of image - W (int): Width of image - - Returns: - x: (B, H, W, C) - """ - B = int(windows.shape[0] / (H * W / window_size / window_size)) - x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) - x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) - return x - - -class WindowAttention(nn.Module): - r""" Window based multi-head self attention (W-MSA) module with relative position bias. - It supports both of shifted and non-shifted window. - - Args: - dim (int): Number of input channels. - window_size (tuple[int]): The height and width of the window. - num_heads (int): Number of attention heads. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set - attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 - proj_drop (float, optional): Dropout ratio of output. Default: 0.0 - """ - - def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): - - super().__init__() - self.dim = dim - self.window_size = window_size # Wh, Ww - self.num_heads = num_heads - head_dim = dim // num_heads - self.scale = qk_scale or head_dim ** -0.5 - - # define a parameter table of relative position bias - self.relative_position_bias_table = nn.Parameter( - torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH - - # get pair-wise relative position index for each token inside the window - coords_h = torch.arange(self.window_size[0]) - coords_w = torch.arange(self.window_size[1]) - coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww - coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww - relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww - relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 - relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 - relative_coords[:, :, 1] += self.window_size[1] - 1 - relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 - relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww - self.register_buffer("relative_position_index", relative_position_index) - - self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) - self.attn_drop = nn.Dropout(attn_drop) - self.proj = nn.Linear(dim, dim) - - self.proj_drop = nn.Dropout(proj_drop) - - trunc_normal_(self.relative_position_bias_table, std=.02) - self.softmax = nn.Softmax(dim=-1) - - def forward(self, x, mask=None): - """ - Args: - x: input features with shape of (num_windows*B, N, C) - mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None - """ - B_, N, C = x.shape - qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) - q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) - - q = q * self.scale - attn = (q @ k.transpose(-2, -1)) - - relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( - self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH - relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww - attn = attn + relative_position_bias.unsqueeze(0) - - if mask is not None: - nW = mask.shape[0] - attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) - attn = attn.view(-1, self.num_heads, N, N) - attn = self.softmax(attn) - else: - attn = self.softmax(attn) - - attn = self.attn_drop(attn) - - x = (attn @ v).transpose(1, 2).reshape(B_, N, C) - x = self.proj(x) - x = self.proj_drop(x) - return x - - def extra_repr(self) -> str: - return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}' - - def flops(self, N): - # calculate flops for 1 window with token length of N - flops = 0 - # qkv = self.qkv(x) - flops += N * self.dim * 3 * self.dim - # attn = (q @ k.transpose(-2, -1)) - flops += self.num_heads * N * (self.dim // self.num_heads) * N - # x = (attn @ v) - flops += self.num_heads * N * N * (self.dim // self.num_heads) - # x = self.proj(x) - flops += N * self.dim * self.dim - return flops - - -class SwinTransformerBlock(nn.Module): - r""" Swin Transformer Block. - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resulotion. - num_heads (int): Number of attention heads. - window_size (int): Window size. - shift_size (int): Shift size for SW-MSA. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float, optional): Stochastic depth rate. Default: 0.0 - act_layer (nn.Module, optional): Activation layer. Default: nn.GELU - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0, - mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., - act_layer=nn.GELU, norm_layer=nn.LayerNorm): - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.num_heads = num_heads - self.window_size = window_size - self.shift_size = shift_size - self.mlp_ratio = mlp_ratio - if min(self.input_resolution) <= self.window_size: - # if window size is larger than input resolution, we don't partition windows - self.shift_size = 0 - self.window_size = min(self.input_resolution) - assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" - - self.norm1 = norm_layer(dim) - self.attn = WindowAttention( - dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, - qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) - - self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() - self.norm2 = norm_layer(dim) - mlp_hidden_dim = int(dim * mlp_ratio) - self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) - - if self.shift_size > 0: - attn_mask = self.calculate_mask(self.input_resolution) - else: - attn_mask = None - - self.register_buffer("attn_mask", attn_mask) - - def calculate_mask(self, x_size): - # calculate attention mask for SW-MSA - H, W = x_size - img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1 - h_slices = (slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None)) - w_slices = (slice(0, -self.window_size), - slice(-self.window_size, -self.shift_size), - slice(-self.shift_size, None)) - cnt = 0 - for h in h_slices: - for w in w_slices: - img_mask[:, h, w, :] = cnt - cnt += 1 - - mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 - mask_windows = mask_windows.view(-1, self.window_size * self.window_size) - attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) - attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) - - return attn_mask - - def forward(self, x, x_size): - H, W = x_size - B, L, C = x.shape - # assert L == H * W, "input feature has wrong size" - - shortcut = x - x = self.norm1(x) - x = x.view(B, H, W, C) - - # cyclic shift - if self.shift_size > 0: - shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) - else: - shifted_x = x - - # partition windows - x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C - x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C - - # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size - if self.input_resolution == x_size: - attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C - else: - attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device)) - - # merge windows - attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) - shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C - - # reverse cyclic shift - if self.shift_size > 0: - x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) - else: - x = shifted_x - x = x.view(B, H * W, C) - - # FFN - x = shortcut + self.drop_path(x) - x = x + self.drop_path(self.mlp(self.norm2(x))) - - return x - - def extra_repr(self) -> str: - return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \ - f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}" - - def flops(self): - flops = 0 - H, W = self.input_resolution - # norm1 - flops += self.dim * H * W - # W-MSA/SW-MSA - nW = H * W / self.window_size / self.window_size - flops += nW * self.attn.flops(self.window_size * self.window_size) - # mlp - flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio - # norm2 - flops += self.dim * H * W - return flops - - -class PatchMerging(nn.Module): - r""" Patch Merging Layer. - - Args: - input_resolution (tuple[int]): Resolution of input feature. - dim (int): Number of input channels. - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - """ - - def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): - super().__init__() - self.input_resolution = input_resolution - self.dim = dim - self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) - self.norm = norm_layer(4 * dim) - - def forward(self, x): - """ - x: B, H*W, C - """ - H, W = self.input_resolution - B, L, C = x.shape - assert L == H * W, "input feature has wrong size" - assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." - - x = x.view(B, H, W, C) - - x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C - x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C - x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C - x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C - x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C - x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C - - x = self.norm(x) - x = self.reduction(x) - - return x - - def extra_repr(self) -> str: - return f"input_resolution={self.input_resolution}, dim={self.dim}" - - def flops(self): - H, W = self.input_resolution - flops = H * W * self.dim - flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim - return flops - - -class BasicLayer(nn.Module): - """ A basic Swin Transformer layer for one stage. - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - """ - - def __init__(self, dim, input_resolution, depth, num_heads, window_size, - mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., - drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False): - - super().__init__() - self.dim = dim - self.input_resolution = input_resolution - self.depth = depth - self.use_checkpoint = use_checkpoint - - # build blocks - self.blocks = nn.ModuleList([ - SwinTransformerBlock(dim=dim, input_resolution=input_resolution, - num_heads=num_heads, window_size=window_size, - shift_size=0 if (i % 2 == 0) else window_size // 2, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, qk_scale=qk_scale, - drop=drop, attn_drop=attn_drop, - drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, - norm_layer=norm_layer) - for i in range(depth)]) - - # patch merging layer - if downsample is not None: - self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer) - else: - self.downsample = None - - def forward(self, x, x_size): - for blk in self.blocks: - if self.use_checkpoint: - x = checkpoint.checkpoint(blk, x, x_size) - else: - x = blk(x, x_size) - if self.downsample is not None: - x = self.downsample(x) - return x - - def extra_repr(self) -> str: - return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" - - def flops(self): - flops = 0 - for blk in self.blocks: - flops += blk.flops() - if self.downsample is not None: - flops += self.downsample.flops() - return flops - - -class RSTB(nn.Module): - """Residual Swin Transformer Block (RSTB). - - Args: - dim (int): Number of input channels. - input_resolution (tuple[int]): Input resolution. - depth (int): Number of blocks. - num_heads (int): Number of attention heads. - window_size (int): Local window size. - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. - qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. - drop (float, optional): Dropout rate. Default: 0.0 - attn_drop (float, optional): Attention dropout rate. Default: 0.0 - drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 - norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm - downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. - img_size: Input image size. - patch_size: Patch size. - resi_connection: The convolutional block before residual connection. - """ - - def __init__(self, dim, input_resolution, depth, num_heads, window_size, - mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., - drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False, - img_size=224, patch_size=4, resi_connection='1conv'): - super(RSTB, self).__init__() - - self.dim = dim - self.input_resolution = input_resolution - - self.residual_group = BasicLayer(dim=dim, - input_resolution=input_resolution, - depth=depth, - num_heads=num_heads, - window_size=window_size, - mlp_ratio=mlp_ratio, - qkv_bias=qkv_bias, qk_scale=qk_scale, - drop=drop, attn_drop=attn_drop, - drop_path=drop_path, - norm_layer=norm_layer, - downsample=downsample, - use_checkpoint=use_checkpoint) - - if resi_connection == '1conv': - self.conv = nn.Conv2d(dim, dim, 3, 1, 1) - elif resi_connection == '3conv': - # to save parameters and memory - self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(dim // 4, dim, 3, 1, 1)) - - self.patch_embed = PatchEmbed( - img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, - norm_layer=None) - - self.patch_unembed = PatchUnEmbed( - img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, - norm_layer=None) - - def forward(self, x, x_size): - return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x - - def flops(self): - flops = 0 - flops += self.residual_group.flops() - H, W = self.input_resolution - flops += H * W * self.dim * self.dim * 9 - flops += self.patch_embed.flops() - flops += self.patch_unembed.flops() - - return flops - - -class PatchEmbed(nn.Module): - r""" Image to Patch Embedding - - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - if norm_layer is not None: - self.norm = norm_layer(embed_dim) - else: - self.norm = None - - def forward(self, x): - x = x.flatten(2).transpose(1, 2) # B Ph*Pw C - if self.norm is not None: - x = self.norm(x) - return x - - def flops(self): - flops = 0 - H, W = self.img_size - if self.norm is not None: - flops += H * W * self.embed_dim - return flops - - -class PatchUnEmbed(nn.Module): - r""" Image to Patch Unembedding - - Args: - img_size (int): Image size. Default: 224. - patch_size (int): Patch token size. Default: 4. - in_chans (int): Number of input image channels. Default: 3. - embed_dim (int): Number of linear projection output channels. Default: 96. - norm_layer (nn.Module, optional): Normalization layer. Default: None - """ - - def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): - super().__init__() - img_size = to_2tuple(img_size) - patch_size = to_2tuple(patch_size) - patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] - self.img_size = img_size - self.patch_size = patch_size - self.patches_resolution = patches_resolution - self.num_patches = patches_resolution[0] * patches_resolution[1] - - self.in_chans = in_chans - self.embed_dim = embed_dim - - def forward(self, x, x_size): - B, HW, C = x.shape - x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C - return x - - def flops(self): - flops = 0 - return flops - - -class Upsample(nn.Sequential): - """Upsample module. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - """ - - def __init__(self, scale, num_feat): - m = [] - if (scale & (scale - 1)) == 0: # scale = 2^n - for _ in range(int(math.log(scale, 2))): - m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(2)) - elif scale == 3: - m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) - m.append(nn.PixelShuffle(3)) - else: - raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.') - super(Upsample, self).__init__(*m) - - -class UpsampleOneStep(nn.Sequential): - """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) - Used in lightweight SR to save parameters. - - Args: - scale (int): Scale factor. Supported scales: 2^n and 3. - num_feat (int): Channel number of intermediate features. - - """ - - def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): - self.num_feat = num_feat - self.input_resolution = input_resolution - m = [] - m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1)) - m.append(nn.PixelShuffle(scale)) - super(UpsampleOneStep, self).__init__(*m) - - def flops(self): - H, W = self.input_resolution - flops = H * W * self.num_feat * 3 * 9 - return flops - - -class SwinIR(nn.Module): - r""" SwinIR - A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer. - - Args: - img_size (int | tuple(int)): Input image size. Default 64 - patch_size (int | tuple(int)): Patch size. Default: 1 - in_chans (int): Number of input image channels. Default: 3 - embed_dim (int): Patch embedding dimension. Default: 96 - depths (tuple(int)): Depth of each Swin Transformer layer. - num_heads (tuple(int)): Number of attention heads in different layers. - window_size (int): Window size. Default: 7 - mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 - qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True - qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None - drop_rate (float): Dropout rate. Default: 0 - attn_drop_rate (float): Attention dropout rate. Default: 0 - drop_path_rate (float): Stochastic depth rate. Default: 0.1 - norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. - ape (bool): If True, add absolute position embedding to the patch embedding. Default: False - patch_norm (bool): If True, add normalization after patch embedding. Default: True - use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False - upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction - img_range: Image range. 1. or 255. - upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None - resi_connection: The convolutional block before residual connection. '1conv'/'3conv' - """ - - def __init__(self, img_size=64, patch_size=1, in_chans=3, - embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6], - window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None, - drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1, - norm_layer=nn.LayerNorm, ape=False, patch_norm=True, - use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv', - **kwargs): - super(SwinIR, self).__init__() - num_in_ch = in_chans - num_out_ch = in_chans - num_feat = 64 - self.img_range = img_range - if in_chans == 3: - rgb_mean = (0.4488, 0.4371, 0.4040) - self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) - else: - self.mean = torch.zeros(1, 1, 1, 1) - self.upscale = upscale - self.upsampler = upsampler - self.window_size = window_size - - ##################################################################################################### - ################################### 1, shallow feature extraction ################################### - self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) - - ##################################################################################################### - ################################### 2, deep feature extraction ###################################### - self.num_layers = len(depths) - self.embed_dim = embed_dim - self.ape = ape - self.patch_norm = patch_norm - self.num_features = embed_dim - self.mlp_ratio = mlp_ratio - - # split image into non-overlapping patches - self.patch_embed = PatchEmbed( - img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None) - num_patches = self.patch_embed.num_patches - patches_resolution = self.patch_embed.patches_resolution - self.patches_resolution = patches_resolution - - # merge non-overlapping patches into image - self.patch_unembed = PatchUnEmbed( - img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, - norm_layer=norm_layer if self.patch_norm else None) - - # absolute position embedding - if self.ape: - self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) - trunc_normal_(self.absolute_pos_embed, std=.02) - - self.pos_drop = nn.Dropout(p=drop_rate) - - # stochastic depth - dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule - - # build Residual Swin Transformer blocks (RSTB) - self.layers = nn.ModuleList() - for i_layer in range(self.num_layers): - layer = RSTB(dim=embed_dim, - input_resolution=(patches_resolution[0], - patches_resolution[1]), - depth=depths[i_layer], - num_heads=num_heads[i_layer], - window_size=window_size, - mlp_ratio=self.mlp_ratio, - qkv_bias=qkv_bias, qk_scale=qk_scale, - drop=drop_rate, attn_drop=attn_drop_rate, - drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results - norm_layer=norm_layer, - downsample=None, - use_checkpoint=use_checkpoint, - img_size=img_size, - patch_size=patch_size, - resi_connection=resi_connection - - ) - self.layers.append(layer) - self.norm = norm_layer(self.num_features) - - # build the last conv layer in deep feature extraction - if resi_connection == '1conv': - self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) - elif resi_connection == '3conv': - # to save parameters and memory - self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), - nn.LeakyReLU(negative_slope=0.2, inplace=True), - nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1)) - - ##################################################################################################### - ################################ 3, high quality image reconstruction ################################ - if self.upsampler == 'pixelshuffle': - # for classical SR - self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), - nn.LeakyReLU(inplace=True)) - self.upsample = Upsample(upscale, num_feat) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - elif self.upsampler == 'pixelshuffledirect': - # for lightweight SR (to save parameters) - self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch, - (patches_resolution[0], patches_resolution[1])) - elif self.upsampler == 'nearest+conv': - # for real-world SR (less artifacts) - self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), - nn.LeakyReLU(inplace=True)) - self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - if self.upscale == 4: - self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1) - self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) - self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) - else: - # for image denoising and JPEG compression artifact reduction - self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1) - - self.apply(self._init_weights) - - def _init_weights(self, m): - if isinstance(m, nn.Linear): - trunc_normal_(m.weight, std=.02) - if isinstance(m, nn.Linear) and m.bias is not None: - nn.init.constant_(m.bias, 0) - elif isinstance(m, nn.LayerNorm): - nn.init.constant_(m.bias, 0) - nn.init.constant_(m.weight, 1.0) - - @torch.jit.ignore - def no_weight_decay(self): - return {'absolute_pos_embed'} - - @torch.jit.ignore - def no_weight_decay_keywords(self): - return {'relative_position_bias_table'} - - def check_image_size(self, x): - _, _, h, w = x.size() - mod_pad_h = (self.window_size - h % self.window_size) % self.window_size - mod_pad_w = (self.window_size - w % self.window_size) % self.window_size - x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') - return x - - def forward_features(self, x): - x_size = (x.shape[2], x.shape[3]) - x = self.patch_embed(x) - if self.ape: - x = x + self.absolute_pos_embed - x = self.pos_drop(x) - - for layer in self.layers: - x = layer(x, x_size) - - x = self.norm(x) # B L C - x = self.patch_unembed(x, x_size) - - return x - - def forward(self, x): - H, W = x.shape[2:] - x = self.check_image_size(x) - - self.mean = self.mean.type_as(x) - x = (x - self.mean) * self.img_range - - if self.upsampler == 'pixelshuffle': - # for classical SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.conv_last(self.upsample(x)) - elif self.upsampler == 'pixelshuffledirect': - # for lightweight SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.upsample(x) - elif self.upsampler == 'nearest+conv': - # for real-world SR - x = self.conv_first(x) - x = self.conv_after_body(self.forward_features(x)) + x - x = self.conv_before_upsample(x) - x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) - if self.upscale == 4: - x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) - x = self.conv_last(self.lrelu(self.conv_hr(x))) - else: - # for image denoising and JPEG compression artifact reduction - x_first = self.conv_first(x) - res = self.conv_after_body(self.forward_features(x_first)) + x_first - x = x + self.conv_last(res) - - x = x / self.img_range + self.mean - - return x[:, :, :H*self.upscale, :W*self.upscale] - - def flops(self): - flops = 0 - H, W = self.patches_resolution - flops += H * W * 3 * self.embed_dim * 9 - flops += self.patch_embed.flops() - for i, layer in enumerate(self.layers): - flops += layer.flops() - flops += H * W * 3 * self.embed_dim * self.embed_dim - flops += self.upsample.flops() - return flops - - -if __name__ == '__main__': - upscale = 4 - window_size = 8 - height = (1024 // upscale // window_size + 1) * window_size - width = (720 // upscale // window_size + 1) * window_size - model = SwinIR(upscale=2, img_size=(height, width), - window_size=window_size, img_range=1., depths=[6, 6, 6, 6], - embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect') - print(model) - print(height, width, model.flops() / 1e9) - - x = torch.randn((1, 3, height, width)) - x = model(x) - print(x.shape) +# ----------------------------------------------------------------------------------- +# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257 +# Originally Written by Ze Liu, Modified by Jingyun Liang. +# ----------------------------------------------------------------------------------- + +import math +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint as checkpoint +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ + + +class Mlp(nn.Module): + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +def window_partition(x, window_size): + """ + Args: + x: (B, H, W, C) + window_size (int): window size + + Returns: + windows: (num_windows*B, window_size, window_size, C) + """ + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + """ + Args: + windows: (num_windows*B, window_size, window_size, C) + window_size (int): Window size + H (int): Height of image + W (int): Width of image + + Returns: + x: (B, H, W, C) + """ + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + r""" Window based multi-head self attention (W-MSA) module with relative position bias. + It supports both of shifted and non-shifted window. + + Args: + dim (int): Number of input channels. + window_size (tuple[int]): The height and width of the window. + num_heads (int): Number of attention heads. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 + proj_drop (float, optional): Dropout ratio of output. Default: 0.0 + """ + + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + # define a parameter table of relative position bias + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + + self.proj_drop = nn.Dropout(proj_drop) + + trunc_normal_(self.relative_position_bias_table, std=.02) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + """ + Args: + x: input features with shape of (num_windows*B, N, C) + mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None + """ + B_, N, C = x.shape + qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q * self.scale + attn = (q @ k.transpose(-2, -1)) + + relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + def extra_repr(self) -> str: + return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}' + + def flops(self, N): + # calculate flops for 1 window with token length of N + flops = 0 + # qkv = self.qkv(x) + flops += N * self.dim * 3 * self.dim + # attn = (q @ k.transpose(-2, -1)) + flops += self.num_heads * N * (self.dim // self.num_heads) * N + # x = (attn @ v) + flops += self.num_heads * N * N * (self.dim // self.num_heads) + # x = self.proj(x) + flops += N * self.dim * self.dim + return flops + + +class SwinTransformerBlock(nn.Module): + r""" Swin Transformer Block. + + Args: + dim (int): Number of input channels. + input_resolution (tuple[int]): Input resulotion. + num_heads (int): Number of attention heads. + window_size (int): Window size. + shift_size (int): Shift size for SW-MSA. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float, optional): Stochastic depth rate. Default: 0.0 + act_layer (nn.Module, optional): Activation layer. Default: nn.GELU + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.input_resolution = input_resolution + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + if min(self.input_resolution) <= self.window_size: + # if window size is larger than input resolution, we don't partition windows + self.shift_size = 0 + self.window_size = min(self.input_resolution) + assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" + + self.norm1 = norm_layer(dim) + self.attn = WindowAttention( + dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + if self.shift_size > 0: + attn_mask = self.calculate_mask(self.input_resolution) + else: + attn_mask = None + + self.register_buffer("attn_mask", attn_mask) + + def calculate_mask(self, x_size): + # calculate attention mask for SW-MSA + H, W = x_size + img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + return attn_mask + + def forward(self, x, x_size): + H, W = x_size + B, L, C = x.shape + # assert L == H * W, "input feature has wrong size" + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + # cyclic shift + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + else: + shifted_x = x + + # partition windows + x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C + + # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size + if self.input_resolution == x_size: + attn_windows = self.attn(x_windows, mask=self.attn_mask) # nW*B, window_size*window_size, C + else: + attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device)) + + # merge windows + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C + + # reverse cyclic shift + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + x = x.view(B, H * W, C) + + # FFN + x = shortcut + self.drop_path(x) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + def extra_repr(self) -> str: + return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \ + f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}" + + def flops(self): + flops = 0 + H, W = self.input_resolution + # norm1 + flops += self.dim * H * W + # W-MSA/SW-MSA + nW = H * W / self.window_size / self.window_size + flops += nW * self.attn.flops(self.window_size * self.window_size) + # mlp + flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio + # norm2 + flops += self.dim * H * W + return flops + + +class PatchMerging(nn.Module): + r""" Patch Merging Layer. + + Args: + input_resolution (tuple[int]): Resolution of input feature. + dim (int): Number of input channels. + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm): + super().__init__() + self.input_resolution = input_resolution + self.dim = dim + self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) + self.norm = norm_layer(4 * dim) + + def forward(self, x): + """ + x: B, H*W, C + """ + H, W = self.input_resolution + B, L, C = x.shape + assert L == H * W, "input feature has wrong size" + assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even." + + x = x.view(B, H, W, C) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + def extra_repr(self) -> str: + return f"input_resolution={self.input_resolution}, dim={self.dim}" + + def flops(self): + H, W = self.input_resolution + flops = H * W * self.dim + flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim + return flops + + +class BasicLayer(nn.Module): + """ A basic Swin Transformer layer for one stage. + + Args: + dim (int): Number of input channels. + input_resolution (tuple[int]): Input resolution. + depth (int): Number of blocks. + num_heads (int): Number of attention heads. + window_size (int): Local window size. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, dim, input_resolution, depth, num_heads, window_size, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False): + + super().__init__() + self.dim = dim + self.input_resolution = input_resolution + self.depth = depth + self.use_checkpoint = use_checkpoint + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock(dim=dim, input_resolution=input_resolution, + num_heads=num_heads, window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, qk_scale=qk_scale, + drop=drop, attn_drop=attn_drop, + drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, + norm_layer=norm_layer) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer) + else: + self.downsample = None + + def forward(self, x, x_size): + for blk in self.blocks: + if self.use_checkpoint: + x = checkpoint.checkpoint(blk, x, x_size) + else: + x = blk(x, x_size) + if self.downsample is not None: + x = self.downsample(x) + return x + + def extra_repr(self) -> str: + return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}" + + def flops(self): + flops = 0 + for blk in self.blocks: + flops += blk.flops() + if self.downsample is not None: + flops += self.downsample.flops() + return flops + + +class RSTB(nn.Module): + """Residual Swin Transformer Block (RSTB). + + Args: + dim (int): Number of input channels. + input_resolution (tuple[int]): Input resolution. + depth (int): Number of blocks. + num_heads (int): Number of attention heads. + window_size (int): Local window size. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + img_size: Input image size. + patch_size: Patch size. + resi_connection: The convolutional block before residual connection. + """ + + def __init__(self, dim, input_resolution, depth, num_heads, window_size, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., + drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False, + img_size=224, patch_size=4, resi_connection='1conv'): + super(RSTB, self).__init__() + + self.dim = dim + self.input_resolution = input_resolution + + self.residual_group = BasicLayer(dim=dim, + input_resolution=input_resolution, + depth=depth, + num_heads=num_heads, + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, qk_scale=qk_scale, + drop=drop, attn_drop=attn_drop, + drop_path=drop_path, + norm_layer=norm_layer, + downsample=downsample, + use_checkpoint=use_checkpoint) + + if resi_connection == '1conv': + self.conv = nn.Conv2d(dim, dim, 3, 1, 1) + elif resi_connection == '3conv': + # to save parameters and memory + self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True), + nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), + nn.LeakyReLU(negative_slope=0.2, inplace=True), + nn.Conv2d(dim // 4, dim, 3, 1, 1)) + + self.patch_embed = PatchEmbed( + img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, + norm_layer=None) + + self.patch_unembed = PatchUnEmbed( + img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, + norm_layer=None) + + def forward(self, x, x_size): + return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x + + def flops(self): + flops = 0 + flops += self.residual_group.flops() + H, W = self.input_resolution + flops += H * W * self.dim * self.dim * 9 + flops += self.patch_embed.flops() + flops += self.patch_unembed.flops() + + return flops + + +class PatchEmbed(nn.Module): + r""" Image to Patch Embedding + + Args: + img_size (int): Image size. Default: 224. + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] + self.img_size = img_size + self.patch_size = patch_size + self.patches_resolution = patches_resolution + self.num_patches = patches_resolution[0] * patches_resolution[1] + + self.in_chans = in_chans + self.embed_dim = embed_dim + + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + x = x.flatten(2).transpose(1, 2) # B Ph*Pw C + if self.norm is not None: + x = self.norm(x) + return x + + def flops(self): + flops = 0 + H, W = self.img_size + if self.norm is not None: + flops += H * W * self.embed_dim + return flops + + +class PatchUnEmbed(nn.Module): + r""" Image to Patch Unembedding + + Args: + img_size (int): Image size. Default: 224. + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + img_size = to_2tuple(img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] + self.img_size = img_size + self.patch_size = patch_size + self.patches_resolution = patches_resolution + self.num_patches = patches_resolution[0] * patches_resolution[1] + + self.in_chans = in_chans + self.embed_dim = embed_dim + + def forward(self, x, x_size): + B, HW, C = x.shape + x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1]) # B Ph*Pw C + return x + + def flops(self): + flops = 0 + return flops + + +class Upsample(nn.Sequential): + """Upsample module. + + Args: + scale (int): Scale factor. Supported scales: 2^n and 3. + num_feat (int): Channel number of intermediate features. + """ + + def __init__(self, scale, num_feat): + m = [] + if (scale & (scale - 1)) == 0: # scale = 2^n + for _ in range(int(math.log(scale, 2))): + m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1)) + m.append(nn.PixelShuffle(2)) + elif scale == 3: + m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1)) + m.append(nn.PixelShuffle(3)) + else: + raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.') + super(Upsample, self).__init__(*m) + + +class UpsampleOneStep(nn.Sequential): + """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle) + Used in lightweight SR to save parameters. + + Args: + scale (int): Scale factor. Supported scales: 2^n and 3. + num_feat (int): Channel number of intermediate features. + + """ + + def __init__(self, scale, num_feat, num_out_ch, input_resolution=None): + self.num_feat = num_feat + self.input_resolution = input_resolution + m = [] + m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1)) + m.append(nn.PixelShuffle(scale)) + super(UpsampleOneStep, self).__init__(*m) + + def flops(self): + H, W = self.input_resolution + flops = H * W * self.num_feat * 3 * 9 + return flops + + +class SwinIR(nn.Module): + r""" SwinIR + A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer. + + Args: + img_size (int | tuple(int)): Input image size. Default 64 + patch_size (int | tuple(int)): Patch size. Default: 1 + in_chans (int): Number of input image channels. Default: 3 + embed_dim (int): Patch embedding dimension. Default: 96 + depths (tuple(int)): Depth of each Swin Transformer layer. + num_heads (tuple(int)): Number of attention heads in different layers. + window_size (int): Window size. Default: 7 + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4 + qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None + drop_rate (float): Dropout rate. Default: 0 + attn_drop_rate (float): Attention dropout rate. Default: 0 + drop_path_rate (float): Stochastic depth rate. Default: 0.1 + norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. + ape (bool): If True, add absolute position embedding to the patch embedding. Default: False + patch_norm (bool): If True, add normalization after patch embedding. Default: True + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False + upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction + img_range: Image range. 1. or 255. + upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None + resi_connection: The convolutional block before residual connection. '1conv'/'3conv' + """ + + def __init__(self, img_size=64, patch_size=1, in_chans=3, + embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6], + window_size=7, mlp_ratio=4., qkv_bias=True, qk_scale=None, + drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1, + norm_layer=nn.LayerNorm, ape=False, patch_norm=True, + use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv', + **kwargs): + super(SwinIR, self).__init__() + num_in_ch = in_chans + num_out_ch = in_chans + num_feat = 64 + self.img_range = img_range + if in_chans == 3: + rgb_mean = (0.4488, 0.4371, 0.4040) + self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1) + else: + self.mean = torch.zeros(1, 1, 1, 1) + self.upscale = upscale + self.upsampler = upsampler + self.window_size = window_size + + ##################################################################################################### + ################################### 1, shallow feature extraction ################################### + self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1) + + ##################################################################################################### + ################################### 2, deep feature extraction ###################################### + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.ape = ape + self.patch_norm = patch_norm + self.num_features = embed_dim + self.mlp_ratio = mlp_ratio + + # split image into non-overlapping patches + self.patch_embed = PatchEmbed( + img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + num_patches = self.patch_embed.num_patches + patches_resolution = self.patch_embed.patches_resolution + self.patches_resolution = patches_resolution + + # merge non-overlapping patches into image + self.patch_unembed = PatchUnEmbed( + img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + + # absolute position embedding + if self.ape: + self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim)) + trunc_normal_(self.absolute_pos_embed, std=.02) + + self.pos_drop = nn.Dropout(p=drop_rate) + + # stochastic depth + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + + # build Residual Swin Transformer blocks (RSTB) + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = RSTB(dim=embed_dim, + input_resolution=(patches_resolution[0], + patches_resolution[1]), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=self.mlp_ratio, + qkv_bias=qkv_bias, qk_scale=qk_scale, + drop=drop_rate, attn_drop=attn_drop_rate, + drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], # no impact on SR results + norm_layer=norm_layer, + downsample=None, + use_checkpoint=use_checkpoint, + img_size=img_size, + patch_size=patch_size, + resi_connection=resi_connection + + ) + self.layers.append(layer) + self.norm = norm_layer(self.num_features) + + # build the last conv layer in deep feature extraction + if resi_connection == '1conv': + self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1) + elif resi_connection == '3conv': + # to save parameters and memory + self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), + nn.LeakyReLU(negative_slope=0.2, inplace=True), + nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), + nn.LeakyReLU(negative_slope=0.2, inplace=True), + nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1)) + + ##################################################################################################### + ################################ 3, high quality image reconstruction ################################ + if self.upsampler == 'pixelshuffle': + # for classical SR + self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), + nn.LeakyReLU(inplace=True)) + self.upsample = Upsample(upscale, num_feat) + self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) + elif self.upsampler == 'pixelshuffledirect': + # for lightweight SR (to save parameters) + self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch, + (patches_resolution[0], patches_resolution[1])) + elif self.upsampler == 'nearest+conv': + # for real-world SR (less artifacts) + self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1), + nn.LeakyReLU(inplace=True)) + self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) + if self.upscale == 4: + self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1) + self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1) + self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1) + self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) + else: + # for image denoising and JPEG compression artifact reduction + self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1) + + self.apply(self._init_weights) + + def _init_weights(self, m): + if isinstance(m, nn.Linear): + trunc_normal_(m.weight, std=.02) + if isinstance(m, nn.Linear) and m.bias is not None: + nn.init.constant_(m.bias, 0) + elif isinstance(m, nn.LayerNorm): + nn.init.constant_(m.bias, 0) + nn.init.constant_(m.weight, 1.0) + + @torch.jit.ignore + def no_weight_decay(self): + return {'absolute_pos_embed'} + + @torch.jit.ignore + def no_weight_decay_keywords(self): + return {'relative_position_bias_table'} + + def check_image_size(self, x): + _, _, h, w = x.size() + mod_pad_h = (self.window_size - h % self.window_size) % self.window_size + mod_pad_w = (self.window_size - w % self.window_size) % self.window_size + x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect') + return x + + def forward_features(self, x): + x_size = (x.shape[2], x.shape[3]) + x = self.patch_embed(x) + if self.ape: + x = x + self.absolute_pos_embed + x = self.pos_drop(x) + + for layer in self.layers: + x = layer(x, x_size) + + x = self.norm(x) # B L C + x = self.patch_unembed(x, x_size) + + return x + + def forward(self, x): + H, W = x.shape[2:] + x = self.check_image_size(x) + + self.mean = self.mean.type_as(x) + x = (x - self.mean) * self.img_range + + if self.upsampler == 'pixelshuffle': + # for classical SR + x = self.conv_first(x) + x = self.conv_after_body(self.forward_features(x)) + x + x = self.conv_before_upsample(x) + x = self.conv_last(self.upsample(x)) + elif self.upsampler == 'pixelshuffledirect': + # for lightweight SR + x = self.conv_first(x) + x = self.conv_after_body(self.forward_features(x)) + x + x = self.upsample(x) + elif self.upsampler == 'nearest+conv': + # for real-world SR + x = self.conv_first(x) + x = self.conv_after_body(self.forward_features(x)) + x + x = self.conv_before_upsample(x) + x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) + if self.upscale == 4: + x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest'))) + x = self.conv_last(self.lrelu(self.conv_hr(x))) + else: + # for image denoising and JPEG compression artifact reduction + x_first = self.conv_first(x) + res = self.conv_after_body(self.forward_features(x_first)) + x_first + x = x + self.conv_last(res) + + x = x / self.img_range + self.mean + + return x[:, :, :H*self.upscale, :W*self.upscale] + + def flops(self): + flops = 0 + H, W = self.patches_resolution + flops += H * W * 3 * self.embed_dim * 9 + flops += self.patch_embed.flops() + for i, layer in enumerate(self.layers): + flops += layer.flops() + flops += H * W * 3 * self.embed_dim * self.embed_dim + flops += self.upsample.flops() + return flops + + +if __name__ == '__main__': + upscale = 4 + window_size = 8 + height = (1024 // upscale // window_size + 1) * window_size + width = (720 // upscale // window_size + 1) * window_size + model = SwinIR(upscale=2, img_size=(height, width), + window_size=window_size, img_range=1., depths=[6, 6, 6, 6], + embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect') + print(model) + print(height, width, model.flops() / 1e9) + + x = torch.randn((1, 3, height, width)) + x = model(x) + print(x.shape) From bfb7f15d46048f27338eeac3a591a5943d03c5f1 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 09:29:22 -0500 Subject: [PATCH 02/10] Rename swinir -> swinir_model --- modules/{swinir.py => swinir_model.py} | 246 ++++++++++++------------- 1 file changed, 123 insertions(+), 123 deletions(-) rename modules/{swinir.py => swinir_model.py} (96%) diff --git a/modules/swinir.py b/modules/swinir_model.py similarity index 96% rename from modules/swinir.py rename to modules/swinir_model.py index 8c534495..e86d0789 100644 --- a/modules/swinir.py +++ b/modules/swinir_model.py @@ -1,123 +1,123 @@ -import sys -import traceback -import cv2 -import os -import contextlib -import numpy as np -from PIL import Image -import torch -import modules.images -from modules.shared import cmd_opts, opts, device -from modules.swinir_arch import SwinIR as net - -precision_scope = ( - torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext -) - - -def load_model(filename, scale=4): - model = net( - upscale=scale, - in_chans=3, - img_size=64, - window_size=8, - img_range=1.0, - depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], - embed_dim=240, - num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8], - mlp_ratio=2, - upsampler="nearest+conv", - resi_connection="3conv", - ) - - pretrained_model = torch.load(filename) - model.load_state_dict(pretrained_model["params_ema"], strict=True) - if not cmd_opts.no_half: - model = model.half() - return model - - -def load_models(dirname): - for file in os.listdir(dirname): - path = os.path.join(dirname, file) - model_name, extension = os.path.splitext(file) - - if extension != ".pt" and extension != ".pth": - continue - - try: - modules.shared.sd_upscalers.append(UpscalerSwin(path, model_name)) - except Exception: - print(f"Error loading SwinIR model: {path}", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) - - -def upscale( - img, - model, - tile=opts.SWIN_tile, - tile_overlap=opts.SWIN_tile_overlap, - window_size=8, - scale=4, -): - img = np.array(img) - img = img[:, :, ::-1] - img = np.moveaxis(img, 2, 0) / 255 - img = torch.from_numpy(img).float() - img = img.unsqueeze(0).to(device) - with torch.no_grad(), precision_scope("cuda"): - _, _, h_old, w_old = img.size() - h_pad = (h_old // window_size + 1) * window_size - h_old - w_pad = (w_old // window_size + 1) * window_size - w_old - img = torch.cat([img, torch.flip(img, [2])], 2)[:, :, : h_old + h_pad, :] - img = torch.cat([img, torch.flip(img, [3])], 3)[:, :, :, : w_old + w_pad] - output = inference(img, model, tile, tile_overlap, window_size, scale) - output = output[..., : h_old * scale, : w_old * scale] - output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy() - if output.ndim == 3: - output = np.transpose( - output[[2, 1, 0], :, :], (1, 2, 0) - ) # CHW-RGB to HCW-BGR - output = (output * 255.0).round().astype(np.uint8) # float32 to uint8 - return Image.fromarray(output, "RGB") - - -def inference(img, model, tile, tile_overlap, window_size, scale): - # test the image tile by tile - b, c, h, w = img.size() - tile = min(tile, h, w) - assert tile % window_size == 0, "tile size should be a multiple of window_size" - sf = scale - - stride = tile - tile_overlap - h_idx_list = list(range(0, h - tile, stride)) + [h - tile] - w_idx_list = list(range(0, w - tile, stride)) + [w - tile] - E = torch.zeros(b, c, h * sf, w * sf, dtype=torch.half, device=device).type_as(img) - W = torch.zeros_like(E, dtype=torch.half, device=device) - - for h_idx in h_idx_list: - for w_idx in w_idx_list: - in_patch = img[..., h_idx : h_idx + tile, w_idx : w_idx + tile] - out_patch = model(in_patch) - out_patch_mask = torch.ones_like(out_patch) - - E[ - ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf - ].add_(out_patch) - W[ - ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf - ].add_(out_patch_mask) - output = E.div_(W) - - return output - - -class UpscalerSwin(modules.images.Upscaler): - def __init__(self, filename, title): - self.name = title - self.model = load_model(filename) - - def do_upscale(self, img): - model = self.model.to(device) - img = upscale(img, model) - return img +import sys +import traceback +import cv2 +import os +import contextlib +import numpy as np +from PIL import Image +import torch +import modules.images +from modules.shared import cmd_opts, opts, device +from modules.swinir_arch import SwinIR as net + +precision_scope = ( + torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext +) + + +def load_model(filename, scale=4): + model = net( + upscale=scale, + in_chans=3, + img_size=64, + window_size=8, + img_range=1.0, + depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], + embed_dim=240, + num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8], + mlp_ratio=2, + upsampler="nearest+conv", + resi_connection="3conv", + ) + + pretrained_model = torch.load(filename) + model.load_state_dict(pretrained_model["params_ema"], strict=True) + if not cmd_opts.no_half: + model = model.half() + return model + + +def load_models(dirname): + for file in os.listdir(dirname): + path = os.path.join(dirname, file) + model_name, extension = os.path.splitext(file) + + if extension != ".pt" and extension != ".pth": + continue + + try: + modules.shared.sd_upscalers.append(UpscalerSwin(path, model_name)) + except Exception: + print(f"Error loading SwinIR model: {path}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + + +def upscale( + img, + model, + tile=opts.SWIN_tile, + tile_overlap=opts.SWIN_tile_overlap, + window_size=8, + scale=4, +): + img = np.array(img) + img = img[:, :, ::-1] + img = np.moveaxis(img, 2, 0) / 255 + img = torch.from_numpy(img).float() + img = img.unsqueeze(0).to(device) + with torch.no_grad(), precision_scope("cuda"): + _, _, h_old, w_old = img.size() + h_pad = (h_old // window_size + 1) * window_size - h_old + w_pad = (w_old // window_size + 1) * window_size - w_old + img = torch.cat([img, torch.flip(img, [2])], 2)[:, :, : h_old + h_pad, :] + img = torch.cat([img, torch.flip(img, [3])], 3)[:, :, :, : w_old + w_pad] + output = inference(img, model, tile, tile_overlap, window_size, scale) + output = output[..., : h_old * scale, : w_old * scale] + output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy() + if output.ndim == 3: + output = np.transpose( + output[[2, 1, 0], :, :], (1, 2, 0) + ) # CHW-RGB to HCW-BGR + output = (output * 255.0).round().astype(np.uint8) # float32 to uint8 + return Image.fromarray(output, "RGB") + + +def inference(img, model, tile, tile_overlap, window_size, scale): + # test the image tile by tile + b, c, h, w = img.size() + tile = min(tile, h, w) + assert tile % window_size == 0, "tile size should be a multiple of window_size" + sf = scale + + stride = tile - tile_overlap + h_idx_list = list(range(0, h - tile, stride)) + [h - tile] + w_idx_list = list(range(0, w - tile, stride)) + [w - tile] + E = torch.zeros(b, c, h * sf, w * sf, dtype=torch.half, device=device).type_as(img) + W = torch.zeros_like(E, dtype=torch.half, device=device) + + for h_idx in h_idx_list: + for w_idx in w_idx_list: + in_patch = img[..., h_idx : h_idx + tile, w_idx : w_idx + tile] + out_patch = model(in_patch) + out_patch_mask = torch.ones_like(out_patch) + + E[ + ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf + ].add_(out_patch) + W[ + ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf + ].add_(out_patch_mask) + output = E.div_(W) + + return output + + +class UpscalerSwin(modules.images.Upscaler): + def __init__(self, filename, title): + self.name = title + self.model = load_model(filename) + + def do_upscale(self, img): + model = self.model.to(device) + img = upscale(img, model) + return img From 740070ea9cdb254209f66417418f2a4af8b099d6 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 09:29:50 -0500 Subject: [PATCH 03/10] Re-implement universal model loading --- modules/codeformer_model.py | 35 +++++--- modules/esrgan_model.py | 56 +++++++++---- modules/extras.py | 2 + modules/gfpgan_model.py | 62 +++++++-------- modules/gfpgan_model_arch.py | 150 +++++++++++++++++++++++++++++++++++ modules/ldsr_model.py | 45 ++++++----- modules/modelloader.py | 65 +++++++++++++++ modules/paths.py | 3 +- modules/realesrgan_model.py | 23 +++++- modules/shared.py | 12 ++- modules/swinir_model.py | 83 +++++++++++++------ webui.py | 47 +++++------ 12 files changed, 449 insertions(+), 134 deletions(-) create mode 100644 modules/gfpgan_model_arch.py create mode 100644 modules/modelloader.py diff --git a/modules/codeformer_model.py b/modules/codeformer_model.py index 8fbdea24..dc0a5eee 100644 --- a/modules/codeformer_model.py +++ b/modules/codeformer_model.py @@ -5,22 +5,28 @@ import traceback import cv2 import torch -from modules import shared, devices -from modules.paths import script_path +from modules import shared, devices, modelloader +from modules.paths import script_path, models_path import modules.shared import modules.face_restoration from importlib import reload -# codeformer people made a choice to include modified basicsr librry to their projectwhich makes -# it utterly impossiblr to use it alongside with other libraries that also use basicsr, like GFPGAN. +# codeformer people made a choice to include modified basicsr library to their project, which makes +# it utterly impossible to use it alongside other libraries that also use basicsr, like GFPGAN. # I am making a choice to include some files from codeformer to work around this issue. - -pretrain_model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth' +model_dir = "Codeformer" +model_path = os.path.join(models_path, model_dir) +model_url = 'https://github.com/sczhou/CodeFormer/releases/download/v0.1.0/codeformer.pth' have_codeformer = False codeformer = None -def setup_codeformer(): + +def setup_model(dirname): + global model_path + if not os.path.exists(model_path): + os.makedirs(model_path) + path = modules.paths.paths.get("CodeFormer", None) if path is None: return @@ -44,16 +50,22 @@ def setup_codeformer(): def name(self): return "CodeFormer" - def __init__(self): + def __init__(self, dirname): self.net = None self.face_helper = None + self.cmd_dir = dirname def create_models(self): if self.net is not None and self.face_helper is not None: self.net.to(devices.device_codeformer) return self.net, self.face_helper - + model_paths = modelloader.load_models(model_path, model_url, self.cmd_dir) + if len(model_paths) != 0: + ckpt_path = model_paths[0] + else: + print("Unable to load codeformer model.") + return None, None net = net_class(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, connect_list=['32', '64', '128', '256']).to(devices.device_codeformer) ckpt_path = load_file_from_url(url=pretrain_model_url, model_dir=os.path.join(path, 'weights/CodeFormer'), progress=True) checkpoint = torch.load(ckpt_path)['params_ema'] @@ -74,6 +86,9 @@ def setup_codeformer(): original_resolution = np_image.shape[0:2] self.create_models() + if self.net is None or self.face_helper is None: + return np_image + self.face_helper.clean_all() self.face_helper.read_image(np_image) self.face_helper.get_face_landmarks_5(only_center_face=False, resize=640, eye_dist_threshold=5) @@ -114,7 +129,7 @@ def setup_codeformer(): have_codeformer = True global codeformer - codeformer = FaceRestorerCodeFormer() + codeformer = FaceRestorerCodeFormer(dirname) shared.face_restorers.append(codeformer) except Exception: diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py index 7f3baf31..dd0ee629 100644 --- a/modules/esrgan_model.py +++ b/modules/esrgan_model.py @@ -5,15 +5,35 @@ import traceback import numpy as np import torch from PIL import Image +from basicsr.utils.download_util import load_file_from_url import modules.esrgam_model_arch as arch -from modules import shared -from modules.shared import opts -from modules.devices import has_mps import modules.images +from modules import shared +from modules import shared, modelloader +from modules.devices import has_mps +from modules.paths import models_path +from modules.shared import opts + +model_dir = "ESRGAN" +model_path = os.path.join(models_path, model_dir) +model_url = "https://drive.google.com/u/0/uc?id=1TPrz5QKd8DHHt1k8SRtm6tMiPjz_Qene&export=download" +model_name = "ESRGAN_x4.pth" -def load_model(filename): +def load_model(path: str, name: str): + global model_path + global model_url + global model_dir + global model_name + if "http" in path: + filename = load_file_from_url(url=model_url, model_dir=model_path, file_name=model_name, progress=True) + else: + filename = path + if not os.path.exists(filename) or filename is None: + print("Unable to load %s from %s" % (model_dir, filename)) + return None + print("Loading %s from %s" % (model_dir, filename)) # this code is adapted from https://github.com/xinntao/ESRGAN pretrained_net = torch.load(filename, map_location='cpu' if has_mps else None) crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32) @@ -118,24 +138,30 @@ def esrgan_upscale(model, img): class UpscalerESRGAN(modules.images.Upscaler): def __init__(self, filename, title): self.name = title - self.model = load_model(filename) + self.filename = filename def do_upscale(self, img): - model = self.model.to(shared.device) + model = load_model(self.filename, self.name) + if model is None: + return img + model.to(shared.device) img = esrgan_upscale(model, img) return img -def load_models(dirname): - for file in os.listdir(dirname): - path = os.path.join(dirname, file) - model_name, extension = os.path.splitext(file) - - if extension != '.pt' and extension != '.pth': - continue +def setup_model(dirname): + global model_path + global model_name + if not os.path.exists(model_path): + os.makedirs(model_path) + model_paths = modelloader.load_models(model_path, command_path=dirname, ext_filter=[".pt", ".pth"]) + if len(model_paths) == 0: + modules.shared.sd_upscalers.append(UpscalerESRGAN(model_url, model_name)) + for file in model_paths: + name = modelloader.friendly_name(file) try: - modules.shared.sd_upscalers.append(UpscalerESRGAN(path, model_name)) + modules.shared.sd_upscalers.append(UpscalerESRGAN(file, name)) except Exception: - print(f"Error loading ESRGAN model: {path}", file=sys.stderr) + print(f"Error loading ESRGAN model: {file}", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) diff --git a/modules/extras.py b/modules/extras.py index 382ffa7d..4c95cf76 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -36,6 +36,8 @@ def run_extras(extras_mode, image, image_folder, gfpgan_visibility, codeformer_v outputs = [] for image, image_name in zip(imageArr, imageNameArr): + if image is None: + return outputs, "Please select an input image.", '' existing_pnginfo = image.info or {} image = image.convert("RGB") diff --git a/modules/gfpgan_model.py b/modules/gfpgan_model.py index 44c5dc6c..ffb6960d 100644 --- a/modules/gfpgan_model.py +++ b/modules/gfpgan_model.py @@ -7,33 +7,20 @@ from modules import shared, devices from modules.shared import cmd_opts from modules.paths import script_path import modules.face_restoration +from modules import shared, devices, modelloader +from modules.paths import models_path - -def gfpgan_model_path(): - from modules.shared import cmd_opts - - filemask = 'GFPGAN*.pth' - - if cmd_opts.gfpgan_model is not None: - return cmd_opts.gfpgan_model - - places = [script_path, '.', os.path.join(cmd_opts.gfpgan_dir, 'experiments/pretrained_models')] - - filename = None - for place in places: - filename = next(iter(glob(os.path.join(place, filemask))), None) - if filename is not None: - break - - return filename - +model_dir = "GFPGAN" +cmd_dir = None +model_path = os.path.join(models_path, model_dir) +model_url = "https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth" loaded_gfpgan_model = None def gfpgan(): global loaded_gfpgan_model - + global model_path if loaded_gfpgan_model is not None: loaded_gfpgan_model.gfpgan.to(shared.device) return loaded_gfpgan_model @@ -41,7 +28,15 @@ def gfpgan(): if gfpgan_constructor is None: return None - model = gfpgan_constructor(model_path=gfpgan_model_path() or 'https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth', upscale=1, arch='clean', channel_multiplier=2, bg_upsampler=None) + models = modelloader.load_models(model_path, model_url, cmd_dir) + if len(models) != 0: + latest_file = max(models, key=os.path.getctime) + model_file = latest_file + else: + print("Unable to load gfpgan model!") + return None + model = gfpgan_constructor(model_path=model_file, model_dir=model_path, upscale=1, arch='clean', channel_multiplier=2, + bg_upsampler=None) model.gfpgan.to(shared.device) loaded_gfpgan_model = model @@ -50,7 +45,8 @@ def gfpgan(): def gfpgan_fix_faces(np_image): model = gfpgan() - + if model is None: + return np_image np_image_bgr = np_image[:, :, ::-1] cropped_faces, restored_faces, gfpgan_output_bgr = model.enhance(np_image_bgr, has_aligned=False, only_center_face=False, paste_back=True) np_image = gfpgan_output_bgr[:, :, ::-1] @@ -64,19 +60,21 @@ def gfpgan_fix_faces(np_image): have_gfpgan = False gfpgan_constructor = None -def setup_gfpgan(): + +def setup_model(dirname): + global model_path + if not os.path.exists(model_path): + os.makedirs(model_path) + try: - gfpgan_model_path() - - if os.path.exists(cmd_opts.gfpgan_dir): - sys.path.append(os.path.abspath(cmd_opts.gfpgan_dir)) - from gfpgan import GFPGANer - + from modules.gfpgan_model_arch import GFPGANerr + global cmd_dir global have_gfpgan - have_gfpgan = True - global gfpgan_constructor - gfpgan_constructor = GFPGANer + + cmd_dir = dirname + have_gfpgan = True + gfpgan_constructor = GFPGANerr class FaceRestorerGFPGAN(modules.face_restoration.FaceRestoration): def name(self): diff --git a/modules/gfpgan_model_arch.py b/modules/gfpgan_model_arch.py new file mode 100644 index 00000000..d81cea96 --- /dev/null +++ b/modules/gfpgan_model_arch.py @@ -0,0 +1,150 @@ +# GFPGAN likes to download stuff "wherever", and we're trying to fix that, so this is a copy of the original... + +import cv2 +import os +import torch +from basicsr.utils import img2tensor, tensor2img +from basicsr.utils.download_util import load_file_from_url +from facexlib.utils.face_restoration_helper import FaceRestoreHelper +from torchvision.transforms.functional import normalize + +from gfpgan.archs.gfpgan_bilinear_arch import GFPGANBilinear +from gfpgan.archs.gfpganv1_arch import GFPGANv1 +from gfpgan.archs.gfpganv1_clean_arch import GFPGANv1Clean + +ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) + + +class GFPGANerr(): + """Helper for restoration with GFPGAN. + + It will detect and crop faces, and then resize the faces to 512x512. + GFPGAN is used to restored the resized faces. + The background is upsampled with the bg_upsampler. + Finally, the faces will be pasted back to the upsample background image. + + Args: + model_path (str): The path to the GFPGAN model. It can be urls (will first download it automatically). + upscale (float): The upscale of the final output. Default: 2. + arch (str): The GFPGAN architecture. Option: clean | original. Default: clean. + channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. + bg_upsampler (nn.Module): The upsampler for the background. Default: None. + """ + + def __init__(self, model_path, model_dir, upscale=2, arch='clean', channel_multiplier=2, bg_upsampler=None, device=None): + self.upscale = upscale + self.bg_upsampler = bg_upsampler + + # initialize model + self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device + # initialize the GFP-GAN + if arch == 'clean': + self.gfpgan = GFPGANv1Clean( + out_size=512, + num_style_feat=512, + channel_multiplier=channel_multiplier, + decoder_load_path=None, + fix_decoder=False, + num_mlp=8, + input_is_latent=True, + different_w=True, + narrow=1, + sft_half=True) + elif arch == 'bilinear': + self.gfpgan = GFPGANBilinear( + out_size=512, + num_style_feat=512, + channel_multiplier=channel_multiplier, + decoder_load_path=None, + fix_decoder=False, + num_mlp=8, + input_is_latent=True, + different_w=True, + narrow=1, + sft_half=True) + elif arch == 'original': + self.gfpgan = GFPGANv1( + out_size=512, + num_style_feat=512, + channel_multiplier=channel_multiplier, + decoder_load_path=None, + fix_decoder=True, + num_mlp=8, + input_is_latent=True, + different_w=True, + narrow=1, + sft_half=True) + elif arch == 'RestoreFormer': + from gfpgan.archs.restoreformer_arch import RestoreFormer + self.gfpgan = RestoreFormer() + # initialize face helper + self.face_helper = FaceRestoreHelper( + upscale, + face_size=512, + crop_ratio=(1, 1), + det_model='retinaface_resnet50', + save_ext='png', + use_parse=True, + device=self.device, + model_rootpath=model_dir) + + if model_path.startswith('https://'): + model_path = load_file_from_url( + url=model_path, model_dir=model_dir, progress=True, file_name=None) + loadnet = torch.load(model_path) + if 'params_ema' in loadnet: + keyname = 'params_ema' + else: + keyname = 'params' + self.gfpgan.load_state_dict(loadnet[keyname], strict=True) + self.gfpgan.eval() + self.gfpgan = self.gfpgan.to(self.device) + + @torch.no_grad() + def enhance(self, img, has_aligned=False, only_center_face=False, paste_back=True, weight=0.5): + self.face_helper.clean_all() + + if has_aligned: # the inputs are already aligned + img = cv2.resize(img, (512, 512)) + self.face_helper.cropped_faces = [img] + else: + self.face_helper.read_image(img) + # get face landmarks for each face + self.face_helper.get_face_landmarks_5(only_center_face=only_center_face, eye_dist_threshold=5) + # eye_dist_threshold=5: skip faces whose eye distance is smaller than 5 pixels + # TODO: even with eye_dist_threshold, it will still introduce wrong detections and restorations. + # align and warp each face + self.face_helper.align_warp_face() + + # face restoration + for cropped_face in self.face_helper.cropped_faces: + # prepare data + cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True) + normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True) + cropped_face_t = cropped_face_t.unsqueeze(0).to(self.device) + + try: + output = self.gfpgan(cropped_face_t, return_rgb=False, weight=weight)[0] + # convert to image + restored_face = tensor2img(output.squeeze(0), rgb2bgr=True, min_max=(-1, 1)) + except RuntimeError as error: + print(f'\tFailed inference for GFPGAN: {error}.') + restored_face = cropped_face + + restored_face = restored_face.astype('uint8') + self.face_helper.add_restored_face(restored_face) + + if not has_aligned and paste_back: + # upsample the background + if self.bg_upsampler is not None: + # Now only support RealESRGAN for upsampling background + bg_img = self.bg_upsampler.enhance(img, outscale=self.upscale)[0] + else: + bg_img = None + + self.face_helper.get_inverse_affine(None) + # paste each restored face to the input image + restored_img = self.face_helper.paste_faces_to_input_image(upsample_img=bg_img) + return self.face_helper.cropped_faces, self.face_helper.restored_faces, restored_img + else: + return self.face_helper.cropped_faces, self.face_helper.restored_faces, None diff --git a/modules/ldsr_model.py b/modules/ldsr_model.py index 95e84659..e6e7ff74 100644 --- a/modules/ldsr_model.py +++ b/modules/ldsr_model.py @@ -3,11 +3,14 @@ import sys import traceback from collections import namedtuple -from basicsr.utils.download_util import load_file_from_url +from modules import shared, images, modelloader, paths +from modules.paths import models_path -import modules.images -from modules import shared -from modules.paths import script_path +model_dir = "LDSR" +model_path = os.path.join(models_path, model_dir) +cmd_path = None +model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1" +yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1" LDSRModelInfo = namedtuple("LDSRModelInfo", ["name", "location", "model", "netscale"]) @@ -25,28 +28,32 @@ class UpscalerLDSR(modules.images.Upscaler): return upscale_with_ldsr(img) -def add_lsdr(): - modules.shared.sd_upscalers.append(UpscalerLDSR(100)) +def setup_model(dirname): + global cmd_path + global model_path + if not os.path.exists(model_path): + os.makedirs(model_path) + cmd_path = dirname + shared.sd_upscalers.append(UpscalerLDSR(100)) -def setup_ldsr(): - path = modules.paths.paths.get("LDSR", None) +def prepare_ldsr(): + path = paths.paths.get("LDSR", None) if path is None: return global have_ldsr global LDSR_obj try: from LDSR import LDSR - model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1" - yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1" - repo_path = 'latent-diffusion/experiments/pretrained_models/' - model_path = load_file_from_url(url=model_url, model_dir=os.path.join("repositories", repo_path), - progress=True, file_name="model.chkpt") - yaml_path = load_file_from_url(url=yaml_url, model_dir=os.path.join("repositories", repo_path), - progress=True, file_name="project.yaml") - have_ldsr = True - LDSR_obj = LDSR(model_path, yaml_path) - + model_files = modelloader.load_models(model_path, model_url, cmd_path, dl_name="model.ckpt", ext_filter=[".ckpt"]) + yaml_files = modelloader.load_models(model_path, yaml_url, cmd_path, dl_name="project.yaml", ext_filter=[".yaml"]) + if len(model_files) != 0 and len(yaml_files) != 0: + model_file = model_files[0] + yaml_file = yaml_files[0] + have_ldsr = True + LDSR_obj = LDSR(model_file, yaml_file) + else: + return except Exception: print("Error importing LDSR:", file=sys.stderr) @@ -55,7 +62,7 @@ def setup_ldsr(): def upscale_with_ldsr(image): - setup_ldsr() + prepare_ldsr() if not have_ldsr or LDSR_obj is None: return image diff --git a/modules/modelloader.py b/modules/modelloader.py new file mode 100644 index 00000000..d59fbe05 --- /dev/null +++ b/modules/modelloader.py @@ -0,0 +1,65 @@ +import os +from urllib.parse import urlparse + +from basicsr.utils.download_util import load_file_from_url + + +def load_models(model_path: str, model_url: str = None, command_path: str = None, dl_name: str = None, existing=None, + ext_filter=None) -> list: + """ + A one-and done loader to try finding the desired models in specified directories. + + @param dl_name: The file name to use for downloading a model. If not specified, it will be used from the URL. + @param model_url: If specified, attempt to download model from the given URL. + @param model_path: The location to store/find models in. + @param command_path: A command-line argument to search for models in first. + @param existing: An array of existing model paths. + @param ext_filter: An optional list of filename extensions to filter by + @return: A list of paths containing the desired model(s) + """ + if ext_filter is None: + ext_filter = [] + if existing is None: + existing = [] + try: + places = [] + if command_path is not None and command_path != model_path: + pretrained_path = os.path.join(command_path, 'experiments/pretrained_models') + if os.path.exists(pretrained_path): + places.append(pretrained_path) + elif os.path.exists(command_path): + places.append(command_path) + places.append(model_path) + for place in places: + if os.path.exists(place): + for file in os.listdir(place): + if os.path.isdir(file): + continue + if len(ext_filter) != 0: + model_name, extension = os.path.splitext(file) + if extension not in ext_filter: + continue + if file not in existing: + path = os.path.join(place, file) + existing.append(path) + if model_url is not None: + if dl_name is not None: + model_file = load_file_from_url(url=model_url, model_dir=model_path, file_name=dl_name, progress=True) + else: + model_file = load_file_from_url(url=model_url, model_dir=model_path, progress=True) + + if os.path.exists(model_file) and os.path.isfile(model_file) and model_file not in existing: + existing.append(model_file) + except: + pass + return existing + + +def friendly_name(file: str): + if "http" in file: + file = urlparse(file).path + + file = os.path.basename(file) + model_name, extension = os.path.splitext(file) + model_name = model_name.replace("_", " ").title() + return model_name diff --git a/modules/paths.py b/modules/paths.py index 3a19f9e5..015fa672 100644 --- a/modules/paths.py +++ b/modules/paths.py @@ -3,9 +3,10 @@ import os import sys script_path = os.path.dirname(os.path.dirname(os.path.realpath(__file__))) +models_path = os.path.join(script_path, "models") sys.path.insert(0, script_path) -# search for directory of stable diffsuion in following palces +# search for directory of stable diffusion in following places sd_path = None possible_sd_paths = [os.path.join(script_path, 'repositories/stable-diffusion'), '.', os.path.dirname(script_path)] for possible_sd_path in possible_sd_paths: diff --git a/modules/realesrgan_model.py b/modules/realesrgan_model.py index c32d6c4c..458bf678 100644 --- a/modules/realesrgan_model.py +++ b/modules/realesrgan_model.py @@ -1,14 +1,20 @@ +import os import sys import traceback from collections import namedtuple import numpy as np from PIL import Image +from basicsr.utils.download_util import load_file_from_url from realesrgan import RealESRGANer import modules.images +from modules.paths import models_path from modules.shared import cmd_opts, opts +model_dir = "RealESRGAN" +model_path = os.path.join(models_path, model_dir) +cmd_dir = None RealesrganModelInfo = namedtuple("RealesrganModelInfo", ["name", "location", "model", "netscale"]) realesrgan_models = [] have_realesrgan = False @@ -17,7 +23,6 @@ have_realesrgan = False def get_realesrgan_models(): try: from basicsr.archs.rrdbnet_arch import RRDBNet - from realesrgan import RealESRGANer from realesrgan.archs.srvgg_arch import SRVGGNetCompact models = [ RealesrganModelInfo( @@ -59,7 +64,7 @@ def get_realesrgan_models(): ] return models except Exception as e: - print("Error makeing Real-ESRGAN midels list:", file=sys.stderr) + print("Error making Real-ESRGAN models list:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) @@ -73,10 +78,15 @@ class UpscalerRealESRGAN(modules.images.Upscaler): return upscale_with_realesrgan(img, self.upscaling, self.model_index) -def setup_realesrgan(): +def setup_model(dirname): + global model_path + if not os.path.exists(model_path): + os.makedirs(model_path) + global realesrgan_models global have_realesrgan - + if model_path != dirname: + model_path = dirname try: from basicsr.archs.rrdbnet_arch import RRDBNet from realesrgan import RealESRGANer @@ -104,6 +114,11 @@ def upscale_with_realesrgan(image, RealESRGAN_upscaling, RealESRGAN_model_index) info = realesrgan_models[RealESRGAN_model_index] model = info.model() + model_file = load_file_from_url(url=info.location, model_dir=model_path, progress=True) + if not os.path.exists(model_file): + print("Unable to load RealESRGAN model: %s" % info.name) + return image + upsampler = RealESRGANer( scale=info.netscale, model_path=info.location, diff --git a/modules/shared.py b/modules/shared.py index c32da110..1444040d 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -16,11 +16,11 @@ import modules.sd_models sd_model_file = os.path.join(script_path, 'model.ckpt') default_sd_model_file = sd_model_file - +model_path = os.path.join(script_path, 'models') parser = argparse.ArgumentParser() parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",) parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; this checkpoint will be added to the list of checkpoints and loaded by default if you don't have a checkpoint selected in settings",) -parser.add_argument("--ckpt-dir", type=str, default=os.path.join(script_path, 'models'), help="path to directory with stable diffusion checkpoints",) +parser.add_argument("--ckpt-dir", type=str, default=model_path, help="path to directory with stable diffusion checkpoints",) parser.add_argument("--gfpgan-dir", type=str, help="GFPGAN directory", default=('./src/gfpgan' if os.path.exists('./src/gfpgan') else './GFPGAN')) parser.add_argument("--gfpgan-model", type=str, help="GFPGAN model file name", default=None) parser.add_argument("--no-half", action='store_true', help="do not switch the model to 16-bit floats") @@ -34,8 +34,12 @@ parser.add_argument("--always-batch-cond-uncond", action='store_true', help="dis parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.") parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast") parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site (doesn't work for me but you might have better luck)") -parser.add_argument("--esrgan-models-path", type=str, help="path to directory with ESRGAN models", default=os.path.join(script_path, 'ESRGAN')) -parser.add_argument("--swinir-models-path", type=str, help="path to directory with SwinIR models", default=os.path.join(script_path, 'SwinIR')) +parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model(s)", default=os.path.join(model_path, 'Codeformer')) +parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model(s)", default=os.path.join(model_path, 'GFPGAN')) +parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN models", default=os.path.join(model_path, 'ESRGAN')) +parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN models", default=os.path.join(model_path, 'RealESRGAN')) +parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR models", default=os.path.join(model_path, 'SwinIR')) +parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR models", default=os.path.join(model_path, 'LDSR')) parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.") parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization") parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find") diff --git a/modules/swinir_model.py b/modules/swinir_model.py index e86d0789..f515779e 100644 --- a/modules/swinir_model.py +++ b/modules/swinir_model.py @@ -1,21 +1,39 @@ +import contextlib +import os import sys import traceback -import cv2 -import os -import contextlib -import numpy as np -from PIL import Image -import torch -import modules.images -from modules.shared import cmd_opts, opts, device -from modules.swinir_arch import SwinIR as net +import numpy as np +import torch +from PIL import Image +from basicsr.utils.download_util import load_file_from_url + +import modules.images +from modules import modelloader +from modules.paths import models_path +from modules.shared import cmd_opts, opts, device +from modules.swinir_model_arch import SwinIR as net + +model_dir = "SwinIR" +model_url = "https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth" +model_name = "SwinIR x4" +model_path = os.path.join(models_path, model_dir) +cmd_path = "" precision_scope = ( torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext ) -def load_model(filename, scale=4): +def load_model(path, scale=4): + global model_path + global model_name + if "http" in path: + dl_name = "%s%s" % (model_name.replace(" ", "_"), ".pth") + filename = load_file_from_url(url=path, model_dir=model_path, file_name=dl_name, progress=True) + else: + filename = path + if filename is None or not os.path.exists(filename): + return None model = net( upscale=scale, in_chans=3, @@ -37,19 +55,29 @@ def load_model(filename, scale=4): return model -def load_models(dirname): - for file in os.listdir(dirname): - path = os.path.join(dirname, file) - model_name, extension = os.path.splitext(file) +def setup_model(dirname): + global model_path + global model_name + global cmd_path + if not os.path.exists(model_path): + os.makedirs(model_path) + cmd_path = dirname + model_file = "" + try: + models = modelloader.load_models(model_path, ext_filter=[".pt", ".pth"], command_path=cmd_path) - if extension != ".pt" and extension != ".pth": - continue + if len(models) != 0: + model_file = models[0] + name = modelloader.friendly_name(model_file) + else: + # Add the "default" model if none are found. + model_file = model_url + name = model_name - try: - modules.shared.sd_upscalers.append(UpscalerSwin(path, model_name)) - except Exception: - print(f"Error loading SwinIR model: {path}", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) + modules.shared.sd_upscalers.append(UpscalerSwin(model_file, name)) + except Exception: + print(f"Error loading SwinIR model: {model_file}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) def upscale( @@ -115,9 +143,16 @@ def inference(img, model, tile, tile_overlap, window_size, scale): class UpscalerSwin(modules.images.Upscaler): def __init__(self, filename, title): self.name = title - self.model = load_model(filename) + self.filename = filename def do_upscale(self, img): - model = self.model.to(device) + model = load_model(self.filename) + if model is None: + return img + model = model.to(device) img = upscale(img, model) - return img + try: + torch.cuda.empty_cache() + except: + pass + return img \ No newline at end of file diff --git a/webui.py b/webui.py index 9ea5f5a3..7e0b3296 100644 --- a/webui.py +++ b/webui.py @@ -1,37 +1,34 @@ import os +import signal import threading -from modules.paths import script_path - -import signal - -from modules.shared import opts, cmd_opts, state -import modules.shared as shared -import modules.ui +import modules.codeformer_model as codeformer +import modules.esrgan_model as esrgan +import modules.extras +import modules.face_restoration +import modules.gfpgan_model as gfpgan +import modules.img2img +import modules.ldsr_model as ldsr +import modules.lowvram +import modules.realesrgan_model as realesrgan import modules.scripts import modules.sd_hijack -import modules.codeformer_model -import modules.gfpgan_model -import modules.face_restoration -import modules.realesrgan_model as realesrgan -import modules.esrgan_model as esrgan -import modules.ldsr_model as ldsr -import modules.extras -import modules.lowvram -import modules.txt2img -import modules.img2img -import modules.swinir as swinir import modules.sd_models +import modules.shared as shared +import modules.swinir_model as swinir +import modules.txt2img +import modules.ui +from modules.paths import script_path +from modules.shared import cmd_opts - -modules.codeformer_model.setup_codeformer() -modules.gfpgan_model.setup_gfpgan() +codeformer.setup_model(cmd_opts.codeformer_models_path) +gfpgan.setup_model(cmd_opts.gfpgan_models_path) shared.face_restorers.append(modules.face_restoration.FaceRestoration()) -esrgan.load_models(cmd_opts.esrgan_models_path) -swinir.load_models(cmd_opts.swinir_models_path) -realesrgan.setup_realesrgan() -ldsr.add_lsdr() +esrgan.setup_model(cmd_opts.esrgan_models_path) +swinir.setup_model(cmd_opts.swinir_models_path) +realesrgan.setup_model(cmd_opts.realesrgan_models_path) +ldsr.setup_model(cmd_opts.ldsr_models_path) queue_lock = threading.Lock() From 7d5c29b674bacc5654f8613af134632b7cbdb158 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 10:27:18 -0500 Subject: [PATCH 04/10] Cleanup existing directories, fixes --- .gitignore | 1 + modules/codeformer_model.py | 15 ++++----------- modules/esrgan_model.py | 4 ++-- modules/ldsr_model.py | 2 +- modules/modelloader.py | 38 +++++++++++++++++++++++++++++++++++++ webui.py | 2 ++ 6 files changed, 48 insertions(+), 14 deletions(-) diff --git a/.gitignore b/.gitignore index 9d78853a..57fa2506 100644 --- a/.gitignore +++ b/.gitignore @@ -4,6 +4,7 @@ __pycache__ /venv /tmp /model.ckpt +/models/**/* /models/*.ckpt /GFPGANv1.3.pth /gfpgan/weights/*.pth diff --git a/modules/codeformer_model.py b/modules/codeformer_model.py index dc0a5eee..efd881eb 100644 --- a/modules/codeformer_model.py +++ b/modules/codeformer_model.py @@ -5,14 +5,13 @@ import traceback import cv2 import torch +import modules.face_restoration +import modules.shared from modules import shared, devices, modelloader from modules.paths import script_path, models_path -import modules.shared -import modules.face_restoration -from importlib import reload -# codeformer people made a choice to include modified basicsr library to their project, which makes -# it utterly impossible to use it alongside other libraries that also use basicsr, like GFPGAN. +# codeformer people made a choice to include modified basicsr library to their project which makes +# it utterly impossible to use it alongside with other libraries that also use basicsr, like GFPGAN. # I am making a choice to include some files from codeformer to work around this issue. model_dir = "Codeformer" model_path = os.path.join(models_path, model_dir) @@ -31,11 +30,6 @@ def setup_model(dirname): if path is None: return - - # both GFPGAN and CodeFormer use bascisr, one has it installed from pip the other uses its own - #stored_sys_path = sys.path - #sys.path = [path] + sys.path - try: from torchvision.transforms.functional import normalize from modules.codeformer.codeformer_arch import CodeFormer @@ -67,7 +61,6 @@ def setup_model(dirname): print("Unable to load codeformer model.") return None, None net = net_class(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, connect_list=['32', '64', '128', '256']).to(devices.device_codeformer) - ckpt_path = load_file_from_url(url=pretrain_model_url, model_dir=os.path.join(path, 'weights/CodeFormer'), progress=True) checkpoint = torch.load(ckpt_path)['params_ema'] net.load_state_dict(checkpoint) net.eval() diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py index dd0ee629..5e10c49c 100644 --- a/modules/esrgan_model.py +++ b/modules/esrgan_model.py @@ -18,7 +18,7 @@ from modules.shared import opts model_dir = "ESRGAN" model_path = os.path.join(models_path, model_dir) model_url = "https://drive.google.com/u/0/uc?id=1TPrz5QKd8DHHt1k8SRtm6tMiPjz_Qene&export=download" -model_name = "ESRGAN_x4.pth" +model_name = "ESRGAN_x4" def load_model(path: str, name: str): @@ -27,7 +27,7 @@ def load_model(path: str, name: str): global model_dir global model_name if "http" in path: - filename = load_file_from_url(url=model_url, model_dir=model_path, file_name=model_name, progress=True) + filename = load_file_from_url(url=model_url, model_dir=model_path, file_name="%s.pth" % model_name, progress=True) else: filename = path if not os.path.exists(filename) or filename is None: diff --git a/modules/ldsr_model.py b/modules/ldsr_model.py index e6e7ff74..4f9b1657 100644 --- a/modules/ldsr_model.py +++ b/modules/ldsr_model.py @@ -19,7 +19,7 @@ have_ldsr = False LDSR_obj = None -class UpscalerLDSR(modules.images.Upscaler): +class UpscalerLDSR(images.Upscaler): def __init__(self, steps): self.steps = steps self.name = "LDSR" diff --git a/modules/modelloader.py b/modules/modelloader.py index d59fbe05..9520a681 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -1,8 +1,11 @@ import os +import shutil from urllib.parse import urlparse from basicsr.utils.download_util import load_file_from_url +from modules.paths import script_path, models_path + def load_models(model_path: str, model_url: str = None, command_path: str = None, dl_name: str = None, existing=None, ext_filter=None) -> list: @@ -63,3 +66,38 @@ def friendly_name(file: str): model_name, extension = os.path.splitext(file) model_name = model_name.replace("_", " ").title() return model_name + + +def cleanup_models(): + root_path = script_path + src_path = os.path.join(root_path, "ESRGAN") + dest_path = os.path.join(models_path, "ESRGAN") + move_files(src_path, dest_path) + src_path = os.path.join(root_path, "gfpgan") + dest_path = os.path.join(models_path, "GFPGAN") + move_files(src_path, dest_path) + src_path = os.path.join(root_path, "SwinIR") + dest_path = os.path.join(models_path, "SwinIR") + move_files(src_path, dest_path) + src_path = os.path.join(root_path, "repositories/latent-diffusion/experiments/pretrained_models/") + dest_path = os.path.join(models_path, "LDSR") + move_files(src_path, dest_path) + + +def move_files(src_path: str, dest_path: str): + try: + if not os.path.exists(dest_path): + os.makedirs(dest_path) + if os.path.exists(src_path): + for file in os.listdir(src_path): + if os.path.isfile(file): + fullpath = os.path.join(src_path, file) + print("Moving file: %s to %s" % (fullpath, dest_path)) + try: + shutil.move(fullpath, dest_path) + except: + pass + print("Removing folder: %s" % src_path) + shutil.rmtree(src_path, True) + except: + pass \ No newline at end of file diff --git a/webui.py b/webui.py index 7e0b3296..e71a217c 100644 --- a/webui.py +++ b/webui.py @@ -18,9 +18,11 @@ import modules.shared as shared import modules.swinir_model as swinir import modules.txt2img import modules.ui +from modules import modelloader from modules.paths import script_path from modules.shared import cmd_opts +modelloader.cleanup_models() codeformer.setup_model(cmd_opts.codeformer_models_path) gfpgan.setup_model(cmd_opts.gfpgan_models_path) shared.face_restorers.append(modules.face_restoration.FaceRestoration()) From 8604e469fdec4c484780937cd1737b40c0ad3d4b Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 10:30:39 -0500 Subject: [PATCH 05/10] Update .gitignore --- .gitignore | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/.gitignore b/.gitignore index 57fa2506..653b9d05 100644 --- a/.gitignore +++ b/.gitignore @@ -1,13 +1,15 @@ __pycache__ -/ESRGAN +*.ckpt +*.pth +/ESRGAN/ +/SwinIR/ +/ESRGAN/* +/SwinIR/* /repositories /venv /tmp /model.ckpt /models/**/* -/models/*.ckpt -/GFPGANv1.3.pth -/gfpgan/weights/*.pth /ui-config.json /outputs /config.json From f860f94c64d98ff575132396b0636cc3e08f8332 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Mon, 26 Sep 2022 10:34:26 -0500 Subject: [PATCH 06/10] Cleanup gitignore, old files --- .gitignore | 4 +--- ESRGAN/Put ESRGAN models here.txt | 0 SwinIR/put_swinir_models_here.txt | 1 - 3 files changed, 1 insertion(+), 4 deletions(-) delete mode 100644 ESRGAN/Put ESRGAN models here.txt delete mode 100644 SwinIR/put_swinir_models_here.txt diff --git a/.gitignore b/.gitignore index 653b9d05..2e188eec 100644 --- a/.gitignore +++ b/.gitignore @@ -1,8 +1,6 @@ __pycache__ *.ckpt *.pth -/ESRGAN/ -/SwinIR/ /ESRGAN/* /SwinIR/* /repositories @@ -22,4 +20,4 @@ __pycache__ /webui-user.sh /interrogate /user.css -/.idea \ No newline at end of file +/.idea diff --git a/ESRGAN/Put ESRGAN models here.txt b/ESRGAN/Put ESRGAN models here.txt deleted file mode 100644 index e69de29b..00000000 diff --git a/SwinIR/put_swinir_models_here.txt b/SwinIR/put_swinir_models_here.txt deleted file mode 100644 index 8b137891..00000000 --- a/SwinIR/put_swinir_models_here.txt +++ /dev/null @@ -1 +0,0 @@ - From 11875f586323cea7c5b8398976449788a83dee76 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Tue, 27 Sep 2022 11:01:13 -0500 Subject: [PATCH 07/10] Use model loader with stable-diffusion too. Hook the model loader into the SD_models file. Add default url/download if checkpoint is not found. Add matching stablediffusion-models-path argument. Add message that --ckpt-dir will be removed in the future, but have it pipe to stablediffusion-models-path for now. Update help strings for models-path args so they're more or less uniform. Move sd_model "setup" call to webUI with the others. Ensure "cleanup_models" method moves existing models to the new locations, including SD, and that we aren't deleting folders that still have stuff in them. --- modules/modelloader.py | 25 +++++++++++++------ modules/sd_models.py | 56 +++++++++++++++++++++++++----------------- modules/shared.py | 28 ++++++++++++--------- webui.py | 1 + 4 files changed, 69 insertions(+), 41 deletions(-) diff --git a/modules/modelloader.py b/modules/modelloader.py index 9520a681..2ee364f0 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -45,7 +45,7 @@ def load_models(model_path: str, model_url: str = None, command_path: str = None if file not in existing: path = os.path.join(place, file) existing.append(path) - if model_url is not None: + if model_url is not None and len(existing) == 0: if dl_name is not None: model_file = load_file_from_url(url=model_url, model_dir=model_path, file_name=dl_name, progress=True) else: @@ -69,7 +69,13 @@ def friendly_name(file: str): def cleanup_models(): + # This code could probably be more efficient if we used a tuple list or something to store the src/destinations + # and then enumerate that, but this works for now. In the future, it'd be nice to just have every "model" scaler + # somehow auto-register and just do these things... root_path = script_path + src_path = models_path + dest_path = os.path.join(models_path, "Stable-diffusion") + move_files(src_path, dest_path, ".ckpt") src_path = os.path.join(root_path, "ESRGAN") dest_path = os.path.join(models_path, "ESRGAN") move_files(src_path, dest_path) @@ -84,20 +90,25 @@ def cleanup_models(): move_files(src_path, dest_path) -def move_files(src_path: str, dest_path: str): +def move_files(src_path: str, dest_path: str, ext_filter: str = None): try: if not os.path.exists(dest_path): os.makedirs(dest_path) if os.path.exists(src_path): for file in os.listdir(src_path): - if os.path.isfile(file): - fullpath = os.path.join(src_path, file) - print("Moving file: %s to %s" % (fullpath, dest_path)) + fullpath = os.path.join(src_path, file) + if os.path.isfile(fullpath): + print(f"Checking file {file} in {src_path}") + if ext_filter is not None: + if ext_filter not in file: + continue + print(f"Moving {file} from {src_path} to {dest_path}.") try: shutil.move(fullpath, dest_path) except: pass - print("Removing folder: %s" % src_path) - shutil.rmtree(src_path, True) + if len(os.listdir(src_path)) == 0: + print(f"Removing empty folder: {src_path}") + shutil.rmtree(src_path, True) except: pass \ No newline at end of file diff --git a/modules/sd_models.py b/modules/sd_models.py index dc81b0dc..89b7d276 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -8,7 +8,13 @@ from omegaconf import OmegaConf from ldm.util import instantiate_from_config -from modules import shared +from modules import shared, modelloader +from modules.paths import models_path + +model_dir = "Stable-diffusion" +model_path = os.path.join(models_path, model_dir) +model_name = "sd-v1-4.ckpt" +model_url = "https://drive.yerf.org/wl/?id=EBfTrmcCCUAGaQBXVIj5lJmEhjoP1tgl&mode=grid&download=1" CheckpointInfo = namedtuple("CheckpointInfo", ['filename', 'title', 'hash']) checkpoints_list = {} @@ -23,23 +29,28 @@ except Exception: pass -def list_models(): +def modeltitle(path, h): + abspath = os.path.abspath(path) + + if abspath.startswith(model_dir): + name = abspath.replace(model_dir, '') + else: + name = os.path.basename(path) + + if name.startswith("\\") or name.startswith("/"): + name = name[1:] + + return f'{name} [{h}]' + + +def setup_model(dirname): + global model_path + global model_name + global model_url + if not os.path.exists(model_path): + os.makedirs(model_path) checkpoints_list.clear() - - model_dir = os.path.abspath(shared.cmd_opts.ckpt_dir) - - def modeltitle(path, h): - abspath = os.path.abspath(path) - - if abspath.startswith(model_dir): - name = abspath.replace(model_dir, '') - else: - name = os.path.basename(path) - - if name.startswith("\\") or name.startswith("/"): - name = name[1:] - - return f'{name} [{h}]' + model_list = modelloader.load_models(model_path, model_url, dirname, model_name, ext_filter=".ckpt") cmd_ckpt = shared.cmd_opts.ckpt if os.path.exists(cmd_ckpt): @@ -47,13 +58,12 @@ def list_models(): title = modeltitle(cmd_ckpt, h) checkpoints_list[title] = CheckpointInfo(cmd_ckpt, title, h) elif cmd_ckpt is not None and cmd_ckpt != shared.default_sd_model_file: - print(f"Checkpoint in --ckpt argument not found: {cmd_ckpt}", file=sys.stderr) + print(f"Checkpoint in --ckpt argument not found (Possible it was moved to {model_path}: {cmd_ckpt}", file=sys.stderr) - if os.path.exists(model_dir): - for filename in glob.glob(model_dir + '/**/*.ckpt', recursive=True): - h = model_hash(filename) - title = modeltitle(filename, h) - checkpoints_list[title] = CheckpointInfo(filename, title, h) + for filename in model_list: + h = model_hash(filename) + title = modeltitle(filename, h) + checkpoints_list[title] = CheckpointInfo(filename, title, h) def model_hash(filename): diff --git a/modules/shared.py b/modules/shared.py index 1444040d..e617bce4 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -20,7 +20,8 @@ model_path = os.path.join(script_path, 'models') parser = argparse.ArgumentParser() parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",) parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; this checkpoint will be added to the list of checkpoints and loaded by default if you don't have a checkpoint selected in settings",) -parser.add_argument("--ckpt-dir", type=str, default=model_path, help="path to directory with stable diffusion checkpoints",) +# This should be deprecated, but we'll leave it for a few iterations +parser.add_argument("--ckpt-dir", type=str, default=None, help="Path to directory with stable diffusion checkpoints (Deprecated, use '--stablediffusion-models-path'", ) parser.add_argument("--gfpgan-dir", type=str, help="GFPGAN directory", default=('./src/gfpgan' if os.path.exists('./src/gfpgan') else './GFPGAN')) parser.add_argument("--gfpgan-model", type=str, help="GFPGAN model file name", default=None) parser.add_argument("--no-half", action='store_true', help="do not switch the model to 16-bit floats") @@ -34,12 +35,13 @@ parser.add_argument("--always-batch-cond-uncond", action='store_true', help="dis parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.") parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast") parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site (doesn't work for me but you might have better luck)") -parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model(s)", default=os.path.join(model_path, 'Codeformer')) -parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model(s)", default=os.path.join(model_path, 'GFPGAN')) -parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN models", default=os.path.join(model_path, 'ESRGAN')) -parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN models", default=os.path.join(model_path, 'RealESRGAN')) -parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR models", default=os.path.join(model_path, 'SwinIR')) -parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR models", default=os.path.join(model_path, 'LDSR')) +parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model file(s).", default=os.path.join(model_path, 'Codeformer')) +parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model file(s).", default=os.path.join(model_path, 'GFPGAN')) +parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN model file(s).", default=os.path.join(model_path, 'ESRGAN')) +parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN model file(s).", default=os.path.join(model_path, 'RealESRGAN')) +parser.add_argument("--stablediffusion-models-path", type=str, help="Path to directory with Stable-diffusion checkpoints.", default=os.path.join(model_path, 'SwinIR')) +parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(model_path, 'SwinIR')) +parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(model_path, 'LDSR')) parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.") parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization") parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find") @@ -57,7 +59,10 @@ parser.add_argument("--autolaunch", action='store_true', help="open the webui UR parser.add_argument("--use-textbox-seed", action='store_true', help="use textbox for seeds in UI (no up/down, but possible to input long seeds)", default=False) cmd_opts = parser.parse_args() - +if cmd_opts.ckpt_dir is not None: + print("The 'ckpt-dir' arg is deprecated in favor of the 'stablediffusion-models-path' argument and will be " + "removed in a future release. Please use the new option if you wish to use a custom checkpoint directory.") + cmd_opts.__setattr__("stablediffusion-models-path", cmd_opts.ckpt_dir) device = get_optimal_device() batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram) @@ -65,6 +70,7 @@ parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram config_filename = cmd_opts.ui_settings_file + class State: interrupted = False job = "" @@ -98,8 +104,8 @@ prompt_styles = modules.styles.StyleDatabase(styles_filename) interrogator = modules.interrogate.InterrogateModels("interrogate") face_restorers = [] - -modules.sd_models.list_models() +# This was moved to webui.py with the other model "setup" calls. +# modules.sd_models.list_models() def realesrgan_models_names(): @@ -195,7 +201,7 @@ options_templates.update(options_section(('system', "System"), { options_templates.update(options_section(('sd', "Stable Diffusion"), { "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Radio, lambda: {"choices": [x.title for x in modules.sd_models.checkpoints_list.values()]}), "img2img_color_correction": OptionInfo(False, "Apply color correction to img2img results to match original colors."), - "save_images_before_color_correction": OptionInfo(False, "Save a copy of image before applying color correction to img2img results"), + "save_images_before_color_correction": OptionInfo(False, "Save a copy of image before applying color correction to img2img results"), "img2img_fix_steps": OptionInfo(False, "With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising)."), "enable_quantization": OptionInfo(False, "Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds. Requires restart to apply."), "enable_emphasis": OptionInfo(True, "Use (text) to make model pay more attention to text and [text] to make it pay less attention"), diff --git a/webui.py b/webui.py index e71a217c..15e49bf6 100644 --- a/webui.py +++ b/webui.py @@ -23,6 +23,7 @@ from modules.paths import script_path from modules.shared import cmd_opts modelloader.cleanup_models() +modules.sd_models.setup_model(cmd_opts.stablediffusion_models_path) codeformer.setup_model(cmd_opts.codeformer_models_path) gfpgan.setup_model(cmd_opts.gfpgan_models_path) shared.face_restorers.append(modules.face_restoration.FaceRestoration()) From 31ad536c331df14dd785bfd2a1f93f91a8f7839e Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Tue, 27 Sep 2022 11:05:25 -0500 Subject: [PATCH 08/10] Remove debugging message --- modules/modelloader.py | 1 - 1 file changed, 1 deletion(-) diff --git a/modules/modelloader.py b/modules/modelloader.py index 2ee364f0..3bd1de4d 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -98,7 +98,6 @@ def move_files(src_path: str, dest_path: str, ext_filter: str = None): for file in os.listdir(src_path): fullpath = os.path.join(src_path, file) if os.path.isfile(fullpath): - print(f"Checking file {file} in {src_path}") if ext_filter is not None: if ext_filter not in file: continue From 0dce0df1ee63b2f158805c1a1f1a3743cc4a104b Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Thu, 29 Sep 2022 17:46:23 -0500 Subject: [PATCH 09/10] Holy $hit. Yep. Fix gfpgan_model_arch requirement(s). Add Upscaler base class, move from images. Add a lot of methods to Upscaler. Re-work all the child upscalers to be proper classes. Add BSRGAN scaler. Add ldsr_model_arch class, removing the dependency for another repo that just uses regular latent-diffusion stuff. Add one universal method that will always find and load new upscaler models without having to add new "setup_model" calls. Still need to add command line params, but that could probably be automated. Add a "self.scale" property to all Upscalers so the scalers themselves can do "things" in response to the requested upscaling size. Ensure LDSR doesn't get stuck in a longer loop of "upscale/downscale/upscale" as we try to reach the target upscale size. Add typehints for IDE sanity. PEP-8 improvements. Moar. --- launch.py | 11 +- modules/bsrgan_model.py | 79 +++++++++++++ modules/bsrgan_model_arch.py | 103 ++++++++++++++++ modules/esrgan_model.py | 223 +++++++++++++++++------------------ modules/extras.py | 35 +++--- modules/gfpgan_model.py | 58 ++++++--- modules/gfpgan_model_arch.py | 150 ----------------------- modules/images.py | 84 ++++++------- modules/ldsr_model.py | 95 ++++++--------- modules/ldsr_model_arch.py | 223 +++++++++++++++++++++++++++++++++++ modules/modelloader.py | 74 ++++++++---- modules/realesrgan_model.py | 209 ++++++++++++++++---------------- modules/sd_models.py | 3 +- modules/sd_samplers.py | 4 +- modules/shared.py | 18 +-- modules/swinir_model.py | 151 +++++++++++------------- modules/upscaler.py | 121 +++++++++++++++++++ webui.py | 9 +- 18 files changed, 1009 insertions(+), 641 deletions(-) create mode 100644 modules/bsrgan_model.py create mode 100644 modules/bsrgan_model_arch.py delete mode 100644 modules/gfpgan_model_arch.py create mode 100644 modules/ldsr_model_arch.py create mode 100644 modules/upscaler.py diff --git a/launch.py b/launch.py index 4462631c..77edea26 100644 --- a/launch.py +++ b/launch.py @@ -1,5 +1,5 @@ # this scripts installs necessary requirements and launches main program in webui.py - +import shutil import subprocess import os import sys @@ -22,7 +22,6 @@ stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "6 taming_transformers_commit_hash = os.environ.get('TAMING_TRANSFORMERS_COMMIT_HASH', "24268930bf1dce879235a7fddd0b2355b84d7ea6") codeformer_commit_hash = os.environ.get('CODEFORMER_COMMIT_HASH', "c5b4593074ba6214284d6acd5f1719b6c5d739af") blip_commit_hash = os.environ.get('BLIP_COMMIT_HASH', "48211a1594f1321b00f14c9f7a5b4813144b2fb9") -ldsr_commit_hash = os.environ.get('LDSR_COMMIT_HASH',"abf33e7002d59d9085081bce93ec798dcabd49af") args = shlex.split(commandline_args) @@ -122,9 +121,11 @@ git_clone("https://github.com/CompVis/stable-diffusion.git", repo_dir('stable-di git_clone("https://github.com/CompVis/taming-transformers.git", repo_dir('taming-transformers'), "Taming Transformers", taming_transformers_commit_hash) git_clone("https://github.com/sczhou/CodeFormer.git", repo_dir('CodeFormer'), "CodeFormer", codeformer_commit_hash) git_clone("https://github.com/salesforce/BLIP.git", repo_dir('BLIP'), "BLIP", blip_commit_hash) -# Using my repo until my changes are merged, as this makes interfacing with our version of SD-web a lot easier -git_clone("https://github.com/Hafiidz/latent-diffusion", repo_dir('latent-diffusion'), "LDSR", ldsr_commit_hash) - +if os.path.isdir(repo_dir('latent-diffusion')): + try: + shutil.rmtree(repo_dir('latent-diffusion')) + except: + pass if not is_installed("lpips"): run_pip(f"install -r {os.path.join(repo_dir('CodeFormer'), 'requirements.txt')}", "requirements for CodeFormer") diff --git a/modules/bsrgan_model.py b/modules/bsrgan_model.py new file mode 100644 index 00000000..77141545 --- /dev/null +++ b/modules/bsrgan_model.py @@ -0,0 +1,79 @@ +import os.path +import sys +import traceback + +import PIL.Image +import numpy as np +import torch +from basicsr.utils.download_util import load_file_from_url + +import modules.upscaler +from modules import shared, modelloader +from modules.bsrgan_model_arch import RRDBNet +from modules.paths import models_path + + +class UpscalerBSRGAN(modules.upscaler.Upscaler): + def __init__(self, dirname): + self.name = "BSRGAN" + self.model_path = os.path.join(models_path, self.name) + self.model_name = "BSRGAN 4x" + self.model_url = "https://github.com/cszn/KAIR/releases/download/v1.0/BSRGAN.pth" + self.user_path = dirname + super().__init__() + model_paths = self.find_models(ext_filter=[".pt", ".pth"]) + scalers = [] + if len(model_paths) == 0: + scaler_data = modules.upscaler.UpscalerData(self.model_name, self.model_url, self, 4) + scalers.append(scaler_data) + for file in model_paths: + if "http" in file: + name = self.model_name + else: + name = modelloader.friendly_name(file) + try: + scaler_data = modules.upscaler.UpscalerData(name, file, self, 4) + scalers.append(scaler_data) + except Exception: + print(f"Error loading BSRGAN model: {file}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + self.scalers = scalers + + def do_upscale(self, img: PIL.Image, selected_file): + torch.cuda.empty_cache() + model = self.load_model(selected_file) + if model is None: + return img + model.to(shared.device) + torch.cuda.empty_cache() + img = np.array(img) + img = img[:, :, ::-1] + img = np.moveaxis(img, 2, 0) / 255 + img = torch.from_numpy(img).float() + img = img.unsqueeze(0).to(shared.device) + with torch.no_grad(): + output = model(img) + output = output.squeeze().float().cpu().clamp_(0, 1).numpy() + output = 255. * np.moveaxis(output, 0, 2) + output = output.astype(np.uint8) + output = output[:, :, ::-1] + torch.cuda.empty_cache() + return PIL.Image.fromarray(output, 'RGB') + + def load_model(self, path: str): + if "http" in path: + filename = load_file_from_url(url=self.model_url, model_dir=self.model_path, file_name="%s.pth" % self.name, + progress=True) + else: + filename = path + if not os.path.exists(filename) or filename is None: + print("Unable to load %s from %s" % (self.model_dir, filename)) + return None + print("Loading %s from %s" % (self.model_dir, filename)) + model = RRDBNet(in_nc=3, out_nc=3, nf=64, nb=23, gc=32, sf=2) # define network + model.load_state_dict(torch.load(filename), strict=True) + model.eval() + for k, v in model.named_parameters(): + v.requires_grad = False + return model + diff --git a/modules/bsrgan_model_arch.py b/modules/bsrgan_model_arch.py new file mode 100644 index 00000000..d72647db --- /dev/null +++ b/modules/bsrgan_model_arch.py @@ -0,0 +1,103 @@ +import functools +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.nn.init as init + + +def initialize_weights(net_l, scale=1): + if not isinstance(net_l, list): + net_l = [net_l] + for net in net_l: + for m in net.modules(): + if isinstance(m, nn.Conv2d): + init.kaiming_normal_(m.weight, a=0, mode='fan_in') + m.weight.data *= scale # for residual block + if m.bias is not None: + m.bias.data.zero_() + elif isinstance(m, nn.Linear): + init.kaiming_normal_(m.weight, a=0, mode='fan_in') + m.weight.data *= scale + if m.bias is not None: + m.bias.data.zero_() + elif isinstance(m, nn.BatchNorm2d): + init.constant_(m.weight, 1) + init.constant_(m.bias.data, 0.0) + + +def make_layer(block, n_layers): + layers = [] + for _ in range(n_layers): + layers.append(block()) + return nn.Sequential(*layers) + + +class ResidualDenseBlock_5C(nn.Module): + def __init__(self, nf=64, gc=32, bias=True): + super(ResidualDenseBlock_5C, self).__init__() + # gc: growth channel, i.e. intermediate channels + self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias) + self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias) + self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias) + self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias) + self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias) + self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) + + # initialization + initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1) + + def forward(self, x): + x1 = self.lrelu(self.conv1(x)) + x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1))) + x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1))) + x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1))) + x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1)) + return x5 * 0.2 + x + + +class RRDB(nn.Module): + '''Residual in Residual Dense Block''' + + def __init__(self, nf, gc=32): + super(RRDB, self).__init__() + self.RDB1 = ResidualDenseBlock_5C(nf, gc) + self.RDB2 = ResidualDenseBlock_5C(nf, gc) + self.RDB3 = ResidualDenseBlock_5C(nf, gc) + + def forward(self, x): + out = self.RDB1(x) + out = self.RDB2(out) + out = self.RDB3(out) + return out * 0.2 + x + + +class RRDBNet(nn.Module): + def __init__(self, in_nc=3, out_nc=3, nf=64, nb=23, gc=32, sf=4): + super(RRDBNet, self).__init__() + RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc) + self.sf = sf + print([in_nc, out_nc, nf, nb, gc, sf]) + + self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1, bias=True) + self.RRDB_trunk = make_layer(RRDB_block_f, nb) + self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) + #### upsampling + self.upconv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) + if self.sf==4: + self.upconv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) + self.HRconv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True) + self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True) + + self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True) + + def forward(self, x): + fea = self.conv_first(x) + trunk = self.trunk_conv(self.RRDB_trunk(fea)) + fea = fea + trunk + + fea = self.lrelu(self.upconv1(F.interpolate(fea, scale_factor=2, mode='nearest'))) + if self.sf==4: + fea = self.lrelu(self.upconv2(F.interpolate(fea, scale_factor=2, mode='nearest'))) + out = self.conv_last(self.lrelu(self.HRconv(fea))) + + return out \ No newline at end of file diff --git a/modules/esrgan_model.py b/modules/esrgan_model.py index 5e10c49c..ce841aa4 100644 --- a/modules/esrgan_model.py +++ b/modules/esrgan_model.py @@ -1,6 +1,4 @@ import os -import sys -import traceback import numpy as np import torch @@ -8,93 +6,119 @@ from PIL import Image from basicsr.utils.download_util import load_file_from_url import modules.esrgam_model_arch as arch -import modules.images -from modules import shared -from modules import shared, modelloader +from modules import shared, modelloader, images from modules.devices import has_mps from modules.paths import models_path +from modules.upscaler import Upscaler, UpscalerData from modules.shared import opts -model_dir = "ESRGAN" -model_path = os.path.join(models_path, model_dir) -model_url = "https://drive.google.com/u/0/uc?id=1TPrz5QKd8DHHt1k8SRtm6tMiPjz_Qene&export=download" -model_name = "ESRGAN_x4" +class UpscalerESRGAN(Upscaler): + def __init__(self, dirname): + self.name = "ESRGAN" + self.model_url = "https://drive.google.com/u/0/uc?id=1TPrz5QKd8DHHt1k8SRtm6tMiPjz_Qene&export=download" + self.model_name = "ESRGAN 4x" + self.scalers = [] + self.user_path = dirname + self.model_path = os.path.join(models_path, self.name) + super().__init__() + model_paths = self.find_models(ext_filter=[".pt", ".pth"]) + scalers = [] + if len(model_paths) == 0: + scaler_data = UpscalerData(self.model_name, self.model_url, self, 4) + scalers.append(scaler_data) + for file in model_paths: + print(f"File: {file}") + if "http" in file: + name = self.model_name + else: + name = modelloader.friendly_name(file) -def load_model(path: str, name: str): - global model_path - global model_url - global model_dir - global model_name - if "http" in path: - filename = load_file_from_url(url=model_url, model_dir=model_path, file_name="%s.pth" % model_name, progress=True) - else: - filename = path - if not os.path.exists(filename) or filename is None: - print("Unable to load %s from %s" % (model_dir, filename)) - return None - print("Loading %s from %s" % (model_dir, filename)) - # this code is adapted from https://github.com/xinntao/ESRGAN - pretrained_net = torch.load(filename, map_location='cpu' if has_mps else None) - crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32) + scaler_data = UpscalerData(name, file, self, 4) + print(f"ESRGAN: Adding scaler {name}") + self.scalers.append(scaler_data) - if 'conv_first.weight' in pretrained_net: - crt_model.load_state_dict(pretrained_net) + def do_upscale(self, img, selected_model): + model = self.load_model(selected_model) + if model is None: + return img + model.to(shared.device) + img = esrgan_upscale(model, img) + return img + + def load_model(self, path: str): + if "http" in path: + filename = load_file_from_url(url=self.model_url, model_dir=self.model_path, + file_name="%s.pth" % self.model_name, + progress=True) + else: + filename = path + if not os.path.exists(filename) or filename is None: + print("Unable to load %s from %s" % (self.model_path, filename)) + return None + # this code is adapted from https://github.com/xinntao/ESRGAN + pretrained_net = torch.load(filename, map_location='cpu' if has_mps else None) + crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32) + + if 'conv_first.weight' in pretrained_net: + crt_model.load_state_dict(pretrained_net) + return crt_model + + if 'model.0.weight' not in pretrained_net: + is_realesrgan = "params_ema" in pretrained_net and 'body.0.rdb1.conv1.weight' in pretrained_net[ + "params_ema"] + if is_realesrgan: + raise Exception("The file is a RealESRGAN model, it can't be used as a ESRGAN model.") + else: + raise Exception("The file is not a ESRGAN model.") + + crt_net = crt_model.state_dict() + load_net_clean = {} + for k, v in pretrained_net.items(): + if k.startswith('module.'): + load_net_clean[k[7:]] = v + else: + load_net_clean[k] = v + pretrained_net = load_net_clean + + tbd = [] + for k, v in crt_net.items(): + tbd.append(k) + + # directly copy + for k, v in crt_net.items(): + if k in pretrained_net and pretrained_net[k].size() == v.size(): + crt_net[k] = pretrained_net[k] + tbd.remove(k) + + crt_net['conv_first.weight'] = pretrained_net['model.0.weight'] + crt_net['conv_first.bias'] = pretrained_net['model.0.bias'] + + for k in tbd.copy(): + if 'RDB' in k: + ori_k = k.replace('RRDB_trunk.', 'model.1.sub.') + if '.weight' in k: + ori_k = ori_k.replace('.weight', '.0.weight') + elif '.bias' in k: + ori_k = ori_k.replace('.bias', '.0.bias') + crt_net[k] = pretrained_net[ori_k] + tbd.remove(k) + + crt_net['trunk_conv.weight'] = pretrained_net['model.1.sub.23.weight'] + crt_net['trunk_conv.bias'] = pretrained_net['model.1.sub.23.bias'] + crt_net['upconv1.weight'] = pretrained_net['model.3.weight'] + crt_net['upconv1.bias'] = pretrained_net['model.3.bias'] + crt_net['upconv2.weight'] = pretrained_net['model.6.weight'] + crt_net['upconv2.bias'] = pretrained_net['model.6.bias'] + crt_net['HRconv.weight'] = pretrained_net['model.8.weight'] + crt_net['HRconv.bias'] = pretrained_net['model.8.bias'] + crt_net['conv_last.weight'] = pretrained_net['model.10.weight'] + crt_net['conv_last.bias'] = pretrained_net['model.10.bias'] + + crt_model.load_state_dict(crt_net) + crt_model.eval() return crt_model - if 'model.0.weight' not in pretrained_net: - is_realesrgan = "params_ema" in pretrained_net and 'body.0.rdb1.conv1.weight' in pretrained_net["params_ema"] - if is_realesrgan: - raise Exception("The file is a RealESRGAN model, it can't be used as a ESRGAN model.") - else: - raise Exception("The file is not a ESRGAN model.") - - crt_net = crt_model.state_dict() - load_net_clean = {} - for k, v in pretrained_net.items(): - if k.startswith('module.'): - load_net_clean[k[7:]] = v - else: - load_net_clean[k] = v - pretrained_net = load_net_clean - - tbd = [] - for k, v in crt_net.items(): - tbd.append(k) - - # directly copy - for k, v in crt_net.items(): - if k in pretrained_net and pretrained_net[k].size() == v.size(): - crt_net[k] = pretrained_net[k] - tbd.remove(k) - - crt_net['conv_first.weight'] = pretrained_net['model.0.weight'] - crt_net['conv_first.bias'] = pretrained_net['model.0.bias'] - - for k in tbd.copy(): - if 'RDB' in k: - ori_k = k.replace('RRDB_trunk.', 'model.1.sub.') - if '.weight' in k: - ori_k = ori_k.replace('.weight', '.0.weight') - elif '.bias' in k: - ori_k = ori_k.replace('.bias', '.0.bias') - crt_net[k] = pretrained_net[ori_k] - tbd.remove(k) - - crt_net['trunk_conv.weight'] = pretrained_net['model.1.sub.23.weight'] - crt_net['trunk_conv.bias'] = pretrained_net['model.1.sub.23.bias'] - crt_net['upconv1.weight'] = pretrained_net['model.3.weight'] - crt_net['upconv1.bias'] = pretrained_net['model.3.bias'] - crt_net['upconv2.weight'] = pretrained_net['model.6.weight'] - crt_net['upconv2.bias'] = pretrained_net['model.6.bias'] - crt_net['HRconv.weight'] = pretrained_net['model.8.weight'] - crt_net['HRconv.bias'] = pretrained_net['model.8.bias'] - crt_net['conv_last.weight'] = pretrained_net['model.10.weight'] - crt_net['conv_last.bias'] = pretrained_net['model.10.bias'] - - crt_model.load_state_dict(crt_net) - crt_model.eval() - return crt_model def upscale_without_tiling(model, img): img = np.array(img) @@ -115,7 +139,7 @@ def esrgan_upscale(model, img): if opts.ESRGAN_tile == 0: return upscale_without_tiling(model, img) - grid = modules.images.split_grid(img, opts.ESRGAN_tile, opts.ESRGAN_tile, opts.ESRGAN_tile_overlap) + grid = images.split_grid(img, opts.ESRGAN_tile, opts.ESRGAN_tile, opts.ESRGAN_tile_overlap) newtiles = [] scale_factor = 1 @@ -130,38 +154,7 @@ def esrgan_upscale(model, img): newrow.append([x * scale_factor, w * scale_factor, output]) newtiles.append([y * scale_factor, h * scale_factor, newrow]) - newgrid = modules.images.Grid(newtiles, grid.tile_w * scale_factor, grid.tile_h * scale_factor, grid.image_w * scale_factor, grid.image_h * scale_factor, grid.overlap * scale_factor) - output = modules.images.combine_grid(newgrid) + newgrid = images.Grid(newtiles, grid.tile_w * scale_factor, grid.tile_h * scale_factor, + grid.image_w * scale_factor, grid.image_h * scale_factor, grid.overlap * scale_factor) + output = images.combine_grid(newgrid) return output - - -class UpscalerESRGAN(modules.images.Upscaler): - def __init__(self, filename, title): - self.name = title - self.filename = filename - - def do_upscale(self, img): - model = load_model(self.filename, self.name) - if model is None: - return img - model.to(shared.device) - img = esrgan_upscale(model, img) - return img - - -def setup_model(dirname): - global model_path - global model_name - if not os.path.exists(model_path): - os.makedirs(model_path) - - model_paths = modelloader.load_models(model_path, command_path=dirname, ext_filter=[".pt", ".pth"]) - if len(model_paths) == 0: - modules.shared.sd_upscalers.append(UpscalerESRGAN(model_url, model_name)) - for file in model_paths: - name = modelloader.friendly_name(file) - try: - modules.shared.sd_upscalers.append(UpscalerESRGAN(file, name)) - except Exception: - print(f"Error loading ESRGAN model: {file}", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) diff --git a/modules/extras.py b/modules/extras.py index af6e631f..d7d0fa54 100644 --- a/modules/extras.py +++ b/modules/extras.py @@ -66,29 +66,28 @@ def run_extras(extras_mode, image, image_folder, gfpgan_visibility, codeformer_v info += f"CodeFormer w: {round(codeformer_weight, 2)}, CodeFormer visibility:{round(codeformer_visibility, 2)}\n" image = res - if upscaling_resize != 1.0: - def upscale(image, scaler_index, resize): - small = image.crop((image.width // 2, image.height // 2, image.width // 2 + 10, image.height // 2 + 10)) - pixels = tuple(np.array(small).flatten().tolist()) - key = (resize, scaler_index, image.width, image.height, gfpgan_visibility, codeformer_visibility, codeformer_weight) + pixels + def upscale(image, scaler_index, resize): + small = image.crop((image.width // 2, image.height // 2, image.width // 2 + 10, image.height // 2 + 10)) + pixels = tuple(np.array(small).flatten().tolist()) + key = (resize, scaler_index, image.width, image.height, gfpgan_visibility, codeformer_visibility, codeformer_weight) + pixels - c = cached_images.get(key) - if c is None: - upscaler = shared.sd_upscalers[scaler_index] - c = upscaler.upscale(image, image.width * resize, image.height * resize) - cached_images[key] = c + c = cached_images.get(key) + if c is None: + upscaler = shared.sd_upscalers[scaler_index] + c = upscaler.scaler.upscale(image, resize, upscaler.data_path) + cached_images[key] = c - return c + return c - info += f"Upscale: {round(upscaling_resize, 3)}, model:{shared.sd_upscalers[extras_upscaler_1].name}\n" - res = upscale(image, extras_upscaler_1, upscaling_resize) + info += f"Upscale: {round(upscaling_resize, 3)}, model:{shared.sd_upscalers[extras_upscaler_1].name}\n" + res = upscale(image, extras_upscaler_1, upscaling_resize) - if extras_upscaler_2 != 0 and extras_upscaler_2_visibility > 0: - res2 = upscale(image, extras_upscaler_2, upscaling_resize) - info += f"Upscale: {round(upscaling_resize, 3)}, visibility: {round(extras_upscaler_2_visibility, 3)}, model:{shared.sd_upscalers[extras_upscaler_2].name}\n" - res = Image.blend(res, res2, extras_upscaler_2_visibility) + if extras_upscaler_2 != 0 and extras_upscaler_2_visibility > 0: + res2 = upscale(image, extras_upscaler_2, upscaling_resize) + info += f"Upscale: {round(upscaling_resize, 3)}, visibility: {round(extras_upscaler_2_visibility, 3)}, model:{shared.sd_upscalers[extras_upscaler_2].name}\n" + res = Image.blend(res, res2, extras_upscaler_2_visibility) - image = res + image = res while len(cached_images) > 2: del cached_images[next(iter(cached_images.keys()))] diff --git a/modules/gfpgan_model.py b/modules/gfpgan_model.py index ffb6960d..2bf8a1ee 100644 --- a/modules/gfpgan_model.py +++ b/modules/gfpgan_model.py @@ -1,24 +1,23 @@ import os import sys import traceback -from glob import glob -from modules import shared, devices -from modules.shared import cmd_opts -from modules.paths import script_path +import facexlib +import gfpgan + import modules.face_restoration from modules import shared, devices, modelloader from modules.paths import models_path model_dir = "GFPGAN" -cmd_dir = None +user_path = None model_path = os.path.join(models_path, model_dir) model_url = "https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth" - +have_gfpgan = False loaded_gfpgan_model = None -def gfpgan(): +def gfpgann(): global loaded_gfpgan_model global model_path if loaded_gfpgan_model is not None: @@ -28,14 +27,16 @@ def gfpgan(): if gfpgan_constructor is None: return None - models = modelloader.load_models(model_path, model_url, cmd_dir) - if len(models) != 0: + models = modelloader.load_models(model_path, model_url, user_path, ext_filter="GFPGAN") + if len(models) == 1 and "http" in models[0]: + model_file = models[0] + elif len(models) != 0: latest_file = max(models, key=os.path.getctime) model_file = latest_file else: print("Unable to load gfpgan model!") return None - model = gfpgan_constructor(model_path=model_file, model_dir=model_path, upscale=1, arch='clean', channel_multiplier=2, + model = gfpgan_constructor(model_path=model_file, upscale=1, arch='clean', channel_multiplier=2, bg_upsampler=None) model.gfpgan.to(shared.device) loaded_gfpgan_model = model @@ -44,11 +45,12 @@ def gfpgan(): def gfpgan_fix_faces(np_image): - model = gfpgan() + model = gfpgann() if model is None: return np_image np_image_bgr = np_image[:, :, ::-1] - cropped_faces, restored_faces, gfpgan_output_bgr = model.enhance(np_image_bgr, has_aligned=False, only_center_face=False, paste_back=True) + cropped_faces, restored_faces, gfpgan_output_bgr = model.enhance(np_image_bgr, has_aligned=False, + only_center_face=False, paste_back=True) np_image = gfpgan_output_bgr[:, :, ::-1] if shared.opts.face_restoration_unload: @@ -57,7 +59,6 @@ def gfpgan_fix_faces(np_image): return np_image -have_gfpgan = False gfpgan_constructor = None @@ -67,14 +68,33 @@ def setup_model(dirname): os.makedirs(model_path) try: - from modules.gfpgan_model_arch import GFPGANerr - global cmd_dir + from gfpgan import GFPGANer + from facexlib import detection, parsing + global user_path global have_gfpgan global gfpgan_constructor - cmd_dir = dirname + load_file_from_url_orig = gfpgan.utils.load_file_from_url + facex_load_file_from_url_orig = facexlib.detection.load_file_from_url + facex_load_file_from_url_orig2 = facexlib.parsing.load_file_from_url + + def my_load_file_from_url(**kwargs): + print("Setting model_dir to " + model_path) + return load_file_from_url_orig(**dict(kwargs, model_dir=model_path)) + + def facex_load_file_from_url(**kwargs): + return facex_load_file_from_url_orig(**dict(kwargs, save_dir=model_path, model_dir=None)) + + def facex_load_file_from_url2(**kwargs): + return facex_load_file_from_url_orig2(**dict(kwargs, save_dir=model_path, model_dir=None)) + + gfpgan.utils.load_file_from_url = my_load_file_from_url + facexlib.detection.load_file_from_url = facex_load_file_from_url + facexlib.parsing.load_file_from_url = facex_load_file_from_url2 + user_path = dirname + print("Have gfpgan should be true?") have_gfpgan = True - gfpgan_constructor = GFPGANerr + gfpgan_constructor = GFPGANer class FaceRestorerGFPGAN(modules.face_restoration.FaceRestoration): def name(self): @@ -82,7 +102,9 @@ def setup_model(dirname): def restore(self, np_image): np_image_bgr = np_image[:, :, ::-1] - cropped_faces, restored_faces, gfpgan_output_bgr = gfpgan().enhance(np_image_bgr, has_aligned=False, only_center_face=False, paste_back=True) + cropped_faces, restored_faces, gfpgan_output_bgr = gfpgann().enhance(np_image_bgr, has_aligned=False, + only_center_face=False, + paste_back=True) np_image = gfpgan_output_bgr[:, :, ::-1] return np_image diff --git a/modules/gfpgan_model_arch.py b/modules/gfpgan_model_arch.py deleted file mode 100644 index d81cea96..00000000 --- a/modules/gfpgan_model_arch.py +++ /dev/null @@ -1,150 +0,0 @@ -# GFPGAN likes to download stuff "wherever", and we're trying to fix that, so this is a copy of the original... - -import cv2 -import os -import torch -from basicsr.utils import img2tensor, tensor2img -from basicsr.utils.download_util import load_file_from_url -from facexlib.utils.face_restoration_helper import FaceRestoreHelper -from torchvision.transforms.functional import normalize - -from gfpgan.archs.gfpgan_bilinear_arch import GFPGANBilinear -from gfpgan.archs.gfpganv1_arch import GFPGANv1 -from gfpgan.archs.gfpganv1_clean_arch import GFPGANv1Clean - -ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) - - -class GFPGANerr(): - """Helper for restoration with GFPGAN. - - It will detect and crop faces, and then resize the faces to 512x512. - GFPGAN is used to restored the resized faces. - The background is upsampled with the bg_upsampler. - Finally, the faces will be pasted back to the upsample background image. - - Args: - model_path (str): The path to the GFPGAN model. It can be urls (will first download it automatically). - upscale (float): The upscale of the final output. Default: 2. - arch (str): The GFPGAN architecture. Option: clean | original. Default: clean. - channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2. - bg_upsampler (nn.Module): The upsampler for the background. Default: None. - """ - - def __init__(self, model_path, model_dir, upscale=2, arch='clean', channel_multiplier=2, bg_upsampler=None, device=None): - self.upscale = upscale - self.bg_upsampler = bg_upsampler - - # initialize model - self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device - # initialize the GFP-GAN - if arch == 'clean': - self.gfpgan = GFPGANv1Clean( - out_size=512, - num_style_feat=512, - channel_multiplier=channel_multiplier, - decoder_load_path=None, - fix_decoder=False, - num_mlp=8, - input_is_latent=True, - different_w=True, - narrow=1, - sft_half=True) - elif arch == 'bilinear': - self.gfpgan = GFPGANBilinear( - out_size=512, - num_style_feat=512, - channel_multiplier=channel_multiplier, - decoder_load_path=None, - fix_decoder=False, - num_mlp=8, - input_is_latent=True, - different_w=True, - narrow=1, - sft_half=True) - elif arch == 'original': - self.gfpgan = GFPGANv1( - out_size=512, - num_style_feat=512, - channel_multiplier=channel_multiplier, - decoder_load_path=None, - fix_decoder=True, - num_mlp=8, - input_is_latent=True, - different_w=True, - narrow=1, - sft_half=True) - elif arch == 'RestoreFormer': - from gfpgan.archs.restoreformer_arch import RestoreFormer - self.gfpgan = RestoreFormer() - # initialize face helper - self.face_helper = FaceRestoreHelper( - upscale, - face_size=512, - crop_ratio=(1, 1), - det_model='retinaface_resnet50', - save_ext='png', - use_parse=True, - device=self.device, - model_rootpath=model_dir) - - if model_path.startswith('https://'): - model_path = load_file_from_url( - url=model_path, model_dir=model_dir, progress=True, file_name=None) - loadnet = torch.load(model_path) - if 'params_ema' in loadnet: - keyname = 'params_ema' - else: - keyname = 'params' - self.gfpgan.load_state_dict(loadnet[keyname], strict=True) - self.gfpgan.eval() - self.gfpgan = self.gfpgan.to(self.device) - - @torch.no_grad() - def enhance(self, img, has_aligned=False, only_center_face=False, paste_back=True, weight=0.5): - self.face_helper.clean_all() - - if has_aligned: # the inputs are already aligned - img = cv2.resize(img, (512, 512)) - self.face_helper.cropped_faces = [img] - else: - self.face_helper.read_image(img) - # get face landmarks for each face - self.face_helper.get_face_landmarks_5(only_center_face=only_center_face, eye_dist_threshold=5) - # eye_dist_threshold=5: skip faces whose eye distance is smaller than 5 pixels - # TODO: even with eye_dist_threshold, it will still introduce wrong detections and restorations. - # align and warp each face - self.face_helper.align_warp_face() - - # face restoration - for cropped_face in self.face_helper.cropped_faces: - # prepare data - cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True) - normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True) - cropped_face_t = cropped_face_t.unsqueeze(0).to(self.device) - - try: - output = self.gfpgan(cropped_face_t, return_rgb=False, weight=weight)[0] - # convert to image - restored_face = tensor2img(output.squeeze(0), rgb2bgr=True, min_max=(-1, 1)) - except RuntimeError as error: - print(f'\tFailed inference for GFPGAN: {error}.') - restored_face = cropped_face - - restored_face = restored_face.astype('uint8') - self.face_helper.add_restored_face(restored_face) - - if not has_aligned and paste_back: - # upsample the background - if self.bg_upsampler is not None: - # Now only support RealESRGAN for upsampling background - bg_img = self.bg_upsampler.enhance(img, outscale=self.upscale)[0] - else: - bg_img = None - - self.face_helper.get_inverse_affine(None) - # paste each restored face to the input image - restored_img = self.face_helper.paste_faces_to_input_image(upsample_img=bg_img) - return self.face_helper.cropped_faces, self.face_helper.restored_faces, restored_img - else: - return self.face_helper.cropped_faces, self.face_helper.restored_faces, None diff --git a/modules/images.py b/modules/images.py index 9458bf8d..a6538dbe 100644 --- a/modules/images.py +++ b/modules/images.py @@ -11,7 +11,6 @@ from PIL import Image, ImageFont, ImageDraw, PngImagePlugin from fonts.ttf import Roboto import string -import modules.shared from modules import sd_samplers, shared from modules.shared import opts, cmd_opts @@ -52,8 +51,8 @@ def split_grid(image, tile_w=512, tile_h=512, overlap=64): cols = math.ceil((w - overlap) / non_overlap_width) rows = math.ceil((h - overlap) / non_overlap_height) - dx = (w - tile_w) / (cols-1) if cols > 1 else 0 - dy = (h - tile_h) / (rows-1) if rows > 1 else 0 + dx = (w - tile_w) / (cols - 1) if cols > 1 else 0 + dy = (h - tile_h) / (rows - 1) if rows > 1 else 0 grid = Grid([], tile_w, tile_h, w, h, overlap) for row in range(rows): @@ -67,7 +66,7 @@ def split_grid(image, tile_w=512, tile_h=512, overlap=64): for col in range(cols): x = int(col * dx) - if x+tile_w >= w: + if x + tile_w >= w: x = w - tile_w tile = image.crop((x, y, x + tile_w, y + tile_h)) @@ -85,8 +84,10 @@ def combine_grid(grid): r = r.astype(np.uint8) return Image.fromarray(r, 'L') - mask_w = make_mask_image(np.arange(grid.overlap, dtype=np.float32).reshape((1, grid.overlap)).repeat(grid.tile_h, axis=0)) - mask_h = make_mask_image(np.arange(grid.overlap, dtype=np.float32).reshape((grid.overlap, 1)).repeat(grid.image_w, axis=1)) + mask_w = make_mask_image( + np.arange(grid.overlap, dtype=np.float32).reshape((1, grid.overlap)).repeat(grid.tile_h, axis=0)) + mask_h = make_mask_image( + np.arange(grid.overlap, dtype=np.float32).reshape((grid.overlap, 1)).repeat(grid.image_w, axis=1)) combined_image = Image.new("RGB", (grid.image_w, grid.image_h)) for y, h, row in grid.tiles: @@ -129,10 +130,12 @@ def draw_grid_annotations(im, width, height, hor_texts, ver_texts): def draw_texts(drawing, draw_x, draw_y, lines): for i, line in enumerate(lines): - drawing.multiline_text((draw_x, draw_y + line.size[1] / 2), line.text, font=fnt, fill=color_active if line.is_active else color_inactive, anchor="mm", align="center") + drawing.multiline_text((draw_x, draw_y + line.size[1] / 2), line.text, font=fnt, + fill=color_active if line.is_active else color_inactive, anchor="mm", align="center") if not line.is_active: - drawing.line((draw_x - line.size[0]//2, draw_y + line.size[1]//2, draw_x + line.size[0]//2, draw_y + line.size[1]//2), fill=color_inactive, width=4) + drawing.line((draw_x - line.size[0] // 2, draw_y + line.size[1] // 2, draw_x + line.size[0] // 2, + draw_y + line.size[1] // 2), fill=color_inactive, width=4) draw_y += line.size[1] + line_spacing @@ -171,7 +174,8 @@ def draw_grid_annotations(im, width, height, hor_texts, ver_texts): line.size = (bbox[2] - bbox[0], bbox[3] - bbox[1]) hor_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing for lines in hor_texts] - ver_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing * len(lines) for lines in ver_texts] + ver_text_heights = [sum([line.size[1] + line_spacing for line in lines]) - line_spacing * len(lines) for lines in + ver_texts] pad_top = max(hor_text_heights) + line_spacing * 2 @@ -202,8 +206,10 @@ def draw_prompt_matrix(im, width, height, all_prompts): prompts_horiz = prompts[:boundary] prompts_vert = prompts[boundary:] - hor_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_horiz)] for pos in range(1 << len(prompts_horiz))] - ver_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_vert)] for pos in range(1 << len(prompts_vert))] + hor_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_horiz)] for pos in + range(1 << len(prompts_horiz))] + ver_texts = [[GridAnnotation(x, is_active=pos & (1 << i) != 0) for i, x in enumerate(prompts_vert)] for pos in + range(1 << len(prompts_vert))] return draw_grid_annotations(im, width, height, hor_texts, ver_texts) @@ -214,7 +220,8 @@ def resize_image(resize_mode, im, width, height): return im.resize((w, h), resample=LANCZOS) upscaler = [x for x in shared.sd_upscalers if x.name == opts.upscaler_for_img2img][0] - return upscaler.upscale(im, w, h) + scale = w / im.width + return upscaler.scaler.upscale(im, scale) if resize_mode == 0: res = resize(im, width, height) @@ -244,11 +251,13 @@ def resize_image(resize_mode, im, width, height): if ratio < src_ratio: fill_height = height // 2 - src_h // 2 res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0)) - res.paste(resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), box=(0, fill_height + src_h)) + res.paste(resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), + box=(0, fill_height + src_h)) elif ratio > src_ratio: fill_width = width // 2 - src_w // 2 res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0)) - res.paste(resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), box=(fill_width + src_w, 0)) + res.paste(resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), + box=(fill_width + src_w, 0)) return res @@ -256,7 +265,7 @@ def resize_image(resize_mode, im, width, height): invalid_filename_chars = '<>:"/\\|?*\n' invalid_filename_prefix = ' ' invalid_filename_postfix = ' .' -re_nonletters = re.compile(r'[\s'+string.punctuation+']+') +re_nonletters = re.compile(r'[\s' + string.punctuation + ']+') max_filename_part_length = 128 @@ -283,7 +292,8 @@ def apply_filename_pattern(x, p, seed, prompt): words = [x for x in re_nonletters.split(prompt or "") if len(x) > 0] if len(words) == 0: words = ["empty"] - x = x.replace("[prompt_words]", sanitize_filename_part(" ".join(words[0:max_prompt_words]), replace_spaces=False)) + x = x.replace("[prompt_words]", + sanitize_filename_part(" ".join(words[0:max_prompt_words]), replace_spaces=False)) if p is not None: x = x.replace("[steps]", str(p.steps)) @@ -291,7 +301,8 @@ def apply_filename_pattern(x, p, seed, prompt): x = x.replace("[width]", str(p.width)) x = x.replace("[height]", str(p.height)) x = x.replace("[styles]", sanitize_filename_part(", ".join(p.styles), replace_spaces=False)) - x = x.replace("[sampler]", sanitize_filename_part(sd_samplers.samplers[p.sampler_index].name, replace_spaces=False)) + x = x.replace("[sampler]", + sanitize_filename_part(sd_samplers.samplers[p.sampler_index].name, replace_spaces=False)) x = x.replace("[model_hash]", shared.sd_model.sd_model_hash) x = x.replace("[date]", datetime.date.today().isoformat()) @@ -303,6 +314,7 @@ def apply_filename_pattern(x, p, seed, prompt): return x + def get_next_sequence_number(path, basename): """ Determines and returns the next sequence number to use when saving an image in the specified directory. @@ -316,7 +328,8 @@ def get_next_sequence_number(path, basename): prefix_length = len(basename) for p in os.listdir(path): if p.startswith(basename): - l = os.path.splitext(p[prefix_length:])[0].split('-') #splits the filename (removing the basename first if one is defined, so the sequence number is always the first element) + l = os.path.splitext(p[prefix_length:])[0].split( + '-') # splits the filename (removing the basename first if one is defined, so the sequence number is always the first element) try: result = max(int(l[0]), result) except ValueError: @@ -324,7 +337,10 @@ def get_next_sequence_number(path, basename): return result + 1 -def save_image(image, path, basename, seed=None, prompt=None, extension='png', info=None, short_filename=False, no_prompt=False, grid=False, pnginfo_section_name='parameters', p=None, existing_info=None, forced_filename=None, suffix=""): + +def save_image(image, path, basename, seed=None, prompt=None, extension='png', info=None, short_filename=False, + no_prompt=False, grid=False, pnginfo_section_name='parameters', p=None, existing_info=None, + forced_filename=None, suffix=""): if short_filename or prompt is None or seed is None: file_decoration = "" elif opts.save_to_dirs: @@ -361,7 +377,7 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i fullfn = "a.png" fullfn_without_extension = "a" for i in range(500): - fn = f"{basecount+i:05}" if basename == '' else f"{basename}-{basecount+i:04}" + fn = f"{basecount + i:05}" if basename == '' else f"{basename}-{basecount + i:04}" fullfn = os.path.join(path, f"{fn}{file_decoration}.{extension}") fullfn_without_extension = os.path.join(path, f"{fn}{file_decoration}") if not os.path.exists(fullfn): @@ -403,31 +419,3 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i file.write(info + "\n") -class Upscaler: - name = "Lanczos" - - def do_upscale(self, img): - return img - - def upscale(self, img, w, h): - for i in range(3): - if img.width >= w and img.height >= h: - break - - img = self.do_upscale(img) - - if img.width != w or img.height != h: - img = img.resize((int(w), int(h)), resample=LANCZOS) - - return img - - -class UpscalerNone(Upscaler): - name = "None" - - def upscale(self, img, w, h): - return img - - -modules.shared.sd_upscalers.append(UpscalerNone()) -modules.shared.sd_upscalers.append(Upscaler()) diff --git a/modules/ldsr_model.py b/modules/ldsr_model.py index 4f9b1657..969d1a0d 100644 --- a/modules/ldsr_model.py +++ b/modules/ldsr_model.py @@ -1,74 +1,45 @@ import os import sys import traceback -from collections import namedtuple -from modules import shared, images, modelloader, paths +from basicsr.utils.download_util import load_file_from_url + +from modules.upscaler import Upscaler, UpscalerData +from modules.ldsr_model_arch import LDSR +from modules import shared from modules.paths import models_path -model_dir = "LDSR" -model_path = os.path.join(models_path, model_dir) -cmd_path = None -model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1" -yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1" -LDSRModelInfo = namedtuple("LDSRModelInfo", ["name", "location", "model", "netscale"]) - -ldsr_models = [] -have_ldsr = False -LDSR_obj = None - - -class UpscalerLDSR(images.Upscaler): - def __init__(self, steps): - self.steps = steps +class UpscalerLDSR(Upscaler): + def __init__(self, user_path): self.name = "LDSR" + self.model_path = os.path.join(models_path, self.name) + self.user_path = user_path + self.model_url = "https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1" + self.yaml_url = "https://heibox.uni-heidelberg.de/f/31a76b13ea27482981b4/?dl=1" + super().__init__() + scaler_data = UpscalerData("LDSR", None, self) + self.scalers = [scaler_data] - def do_upscale(self, img): - return upscale_with_ldsr(img) + def load_model(self, path: str): + model = load_file_from_url(url=self.model_url, model_dir=self.model_path, + file_name="model.pth", progress=True) + yaml = load_file_from_url(url=self.model_url, model_dir=self.model_path, + file_name="project.yaml", progress=True) + try: + return LDSR(model, yaml) -def setup_model(dirname): - global cmd_path - global model_path - if not os.path.exists(model_path): - os.makedirs(model_path) - cmd_path = dirname - shared.sd_upscalers.append(UpscalerLDSR(100)) + except Exception: + print("Error importing LDSR:", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + return None - -def prepare_ldsr(): - path = paths.paths.get("LDSR", None) - if path is None: - return - global have_ldsr - global LDSR_obj - try: - from LDSR import LDSR - model_files = modelloader.load_models(model_path, model_url, cmd_path, dl_name="model.ckpt", ext_filter=[".ckpt"]) - yaml_files = modelloader.load_models(model_path, yaml_url, cmd_path, dl_name="project.yaml", ext_filter=[".yaml"]) - if len(model_files) != 0 and len(yaml_files) != 0: - model_file = model_files[0] - yaml_file = yaml_files[0] - have_ldsr = True - LDSR_obj = LDSR(model_file, yaml_file) - else: - return - - except Exception: - print("Error importing LDSR:", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) - have_ldsr = False - - -def upscale_with_ldsr(image): - prepare_ldsr() - if not have_ldsr or LDSR_obj is None: - return image - - ddim_steps = shared.opts.ldsr_steps - pre_scale = shared.opts.ldsr_pre_down - post_scale = shared.opts.ldsr_post_down - - image = LDSR_obj.super_resolution(image, ddim_steps, pre_scale, post_scale) - return image + def do_upscale(self, img, path): + ldsr = self.load_model(path) + if ldsr is None: + print("NO LDSR!") + return img + ddim_steps = shared.opts.ldsr_steps + pre_scale = shared.opts.ldsr_pre_down + return ldsr.super_resolution(img, ddim_steps, self.scale) diff --git a/modules/ldsr_model_arch.py b/modules/ldsr_model_arch.py new file mode 100644 index 00000000..8fe87c6a --- /dev/null +++ b/modules/ldsr_model_arch.py @@ -0,0 +1,223 @@ +import gc +import time +import warnings + +import numpy as np +import torch +import torchvision +from PIL import Image +from einops import rearrange, repeat +from omegaconf import OmegaConf + +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.util import instantiate_from_config, ismap + +warnings.filterwarnings("ignore", category=UserWarning) + + +# Create LDSR Class +class LDSR: + def load_model_from_config(self, half_attention): + print(f"Loading model from {self.modelPath}") + pl_sd = torch.load(self.modelPath, map_location="cpu") + sd = pl_sd["state_dict"] + config = OmegaConf.load(self.yamlPath) + model = instantiate_from_config(config.model) + model.load_state_dict(sd, strict=False) + model.cuda() + if half_attention: + model = model.half() + + model.eval() + return {"model": model} + + def __init__(self, model_path, yaml_path): + self.modelPath = model_path + self.yamlPath = yaml_path + + @staticmethod + def run(model, selected_path, custom_steps, eta): + example = get_cond(selected_path) + + n_runs = 1 + guider = None + ckwargs = None + ddim_use_x0_pred = False + temperature = 1. + eta = eta + custom_shape = None + + height, width = example["image"].shape[1:3] + split_input = height >= 128 and width >= 128 + + if split_input: + ks = 128 + stride = 64 + vqf = 4 # + model.split_input_params = {"ks": (ks, ks), "stride": (stride, stride), + "vqf": vqf, + "patch_distributed_vq": True, + "tie_braker": False, + "clip_max_weight": 0.5, + "clip_min_weight": 0.01, + "clip_max_tie_weight": 0.5, + "clip_min_tie_weight": 0.01} + else: + if hasattr(model, "split_input_params"): + delattr(model, "split_input_params") + + x_t = None + logs = None + for n in range(n_runs): + if custom_shape is not None: + x_t = torch.randn(1, custom_shape[1], custom_shape[2], custom_shape[3]).to(model.device) + x_t = repeat(x_t, '1 c h w -> b c h w', b=custom_shape[0]) + + logs = make_convolutional_sample(example, model, + custom_steps=custom_steps, + eta=eta, quantize_x0=False, + custom_shape=custom_shape, + temperature=temperature, noise_dropout=0., + corrector=guider, corrector_kwargs=ckwargs, x_T=x_t, + ddim_use_x0_pred=ddim_use_x0_pred + ) + return logs + + def super_resolution(self, image, steps=100, target_scale=2, half_attention=False): + model = self.load_model_from_config(half_attention) + + # Run settings + diffusion_steps = int(steps) + eta = 1.0 + + down_sample_method = 'Lanczos' + + gc.collect() + torch.cuda.empty_cache() + + im_og = image + width_og, height_og = im_og.size + # If we can adjust the max upscale size, then the 4 below should be our variable + print("Foo") + down_sample_rate = target_scale / 4 + print(f"Downsample rate is {down_sample_rate}") + width_downsampled_pre = width_og * down_sample_rate + height_downsampled_pre = height_og * down_sample_method + + if down_sample_rate != 1: + print( + f'Downsampling from [{width_og}, {height_og}] to [{width_downsampled_pre}, {height_downsampled_pre}]') + im_og = im_og.resize((width_downsampled_pre, height_downsampled_pre), Image.LANCZOS) + else: + print(f"Down sample rate is 1 from {target_scale} / 4") + logs = self.run(model["model"], im_og, diffusion_steps, eta) + + sample = logs["sample"] + sample = sample.detach().cpu() + sample = torch.clamp(sample, -1., 1.) + sample = (sample + 1.) / 2. * 255 + sample = sample.numpy().astype(np.uint8) + sample = np.transpose(sample, (0, 2, 3, 1)) + a = Image.fromarray(sample[0]) + + del model + gc.collect() + torch.cuda.empty_cache() + print(f'Processing finished!') + return a + + +def get_cond(selected_path): + example = dict() + up_f = 4 + c = selected_path.convert('RGB') + c = torch.unsqueeze(torchvision.transforms.ToTensor()(c), 0) + c_up = torchvision.transforms.functional.resize(c, size=[up_f * c.shape[2], up_f * c.shape[3]], + antialias=True) + c_up = rearrange(c_up, '1 c h w -> 1 h w c') + c = rearrange(c, '1 c h w -> 1 h w c') + c = 2. * c - 1. + + c = c.to(torch.device("cuda")) + example["LR_image"] = c + example["image"] = c_up + + return example + + +@torch.no_grad() +def convsample_ddim(model, cond, steps, shape, eta=1.0, callback=None, normals_sequence=None, + mask=None, x0=None, quantize_x0=False, temperature=1., score_corrector=None, + corrector_kwargs=None, x_t=None + ): + ddim = DDIMSampler(model) + bs = shape[0] + shape = shape[1:] + print(f"Sampling with eta = {eta}; steps: {steps}") + samples, intermediates = ddim.sample(steps, batch_size=bs, shape=shape, conditioning=cond, callback=callback, + normals_sequence=normals_sequence, quantize_x0=quantize_x0, eta=eta, + mask=mask, x0=x0, temperature=temperature, verbose=False, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, x_t=x_t) + + return samples, intermediates + + +@torch.no_grad() +def make_convolutional_sample(batch, model, custom_steps=None, eta=1.0, quantize_x0=False, custom_shape=None, temperature=1., noise_dropout=0., corrector=None, + corrector_kwargs=None, x_T=None, ddim_use_x0_pred=False): + log = dict() + + z, c, x, xrec, xc = model.get_input(batch, model.first_stage_key, + return_first_stage_outputs=True, + force_c_encode=not (hasattr(model, 'split_input_params') + and model.cond_stage_key == 'coordinates_bbox'), + return_original_cond=True) + + if custom_shape is not None: + z = torch.randn(custom_shape) + print(f"Generating {custom_shape[0]} samples of shape {custom_shape[1:]}") + + z0 = None + + log["input"] = x + log["reconstruction"] = xrec + + if ismap(xc): + log["original_conditioning"] = model.to_rgb(xc) + if hasattr(model, 'cond_stage_key'): + log[model.cond_stage_key] = model.to_rgb(xc) + + else: + log["original_conditioning"] = xc if xc is not None else torch.zeros_like(x) + if model.cond_stage_model: + log[model.cond_stage_key] = xc if xc is not None else torch.zeros_like(x) + if model.cond_stage_key == 'class_label': + log[model.cond_stage_key] = xc[model.cond_stage_key] + + with model.ema_scope("Plotting"): + t0 = time.time() + + sample, intermediates = convsample_ddim(model, c, steps=custom_steps, shape=z.shape, + eta=eta, + quantize_x0=quantize_x0, mask=None, x0=z0, + temperature=temperature, score_corrector=corrector, corrector_kwargs=corrector_kwargs, + x_t=x_T) + t1 = time.time() + + if ddim_use_x0_pred: + sample = intermediates['pred_x0'][-1] + + x_sample = model.decode_first_stage(sample) + + try: + x_sample_noquant = model.decode_first_stage(sample, force_not_quantize=True) + log["sample_noquant"] = x_sample_noquant + log["sample_diff"] = torch.abs(x_sample_noquant - x_sample) + except: + pass + + log["sample"] = x_sample + log["time"] = t1 - t0 + + return log diff --git a/modules/modelloader.py b/modules/modelloader.py index 3bd1de4d..6de65c69 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -1,34 +1,36 @@ import os import shutil +import importlib from urllib.parse import urlparse from basicsr.utils.download_util import load_file_from_url +from modules import shared +from modules.upscaler import Upscaler from modules.paths import script_path, models_path -def load_models(model_path: str, model_url: str = None, command_path: str = None, dl_name: str = None, existing=None, - ext_filter=None) -> list: +def load_models(model_path: str, model_url: str = None, command_path: str = None, ext_filter=None, download_name=None) -> list: """ A one-and done loader to try finding the desired models in specified directories. - @param dl_name: The file name to use for downloading a model. If not specified, it will be used from the URL. - @param model_url: If specified, attempt to download model from the given URL. + @param download_name: Specify to download from model_url immediately. + @param model_url: If no other models are found, this will be downloaded on upscale. @param model_path: The location to store/find models in. @param command_path: A command-line argument to search for models in first. - @param existing: An array of existing model paths. @param ext_filter: An optional list of filename extensions to filter by @return: A list of paths containing the desired model(s) """ + output = [] + if ext_filter is None: ext_filter = [] - if existing is None: - existing = [] try: places = [] if command_path is not None and command_path != model_path: pretrained_path = os.path.join(command_path, 'experiments/pretrained_models') if os.path.exists(pretrained_path): + print(f"Appending path: {pretrained_path}") places.append(pretrained_path) elif os.path.exists(command_path): places.append(command_path) @@ -36,26 +38,24 @@ def load_models(model_path: str, model_url: str = None, command_path: str = None for place in places: if os.path.exists(place): for file in os.listdir(place): - if os.path.isdir(file): + full_path = os.path.join(place, file) + if os.path.isdir(full_path): continue if len(ext_filter) != 0: model_name, extension = os.path.splitext(file) if extension not in ext_filter: continue - if file not in existing: - path = os.path.join(place, file) - existing.append(path) - if model_url is not None and len(existing) == 0: - if dl_name is not None: - model_file = load_file_from_url(url=model_url, model_dir=model_path, file_name=dl_name, progress=True) + if file not in output: + output.append(full_path) + if model_url is not None and len(output) == 0: + if download_name is not None: + dl = load_file_from_url(model_url, model_path, True, download_name) + output.append(dl) else: - model_file = load_file_from_url(url=model_url, model_dir=model_path, progress=True) - - if os.path.exists(model_file) and os.path.isfile(model_file) and model_file not in existing: - existing.append(model_file) + output.append(model_url) except: pass - return existing + return output def friendly_name(file: str): @@ -110,4 +110,38 @@ def move_files(src_path: str, dest_path: str, ext_filter: str = None): print(f"Removing empty folder: {src_path}") shutil.rmtree(src_path, True) except: - pass \ No newline at end of file + pass + + +def load_upscalers(): + datas = [] + for cls in Upscaler.__subclasses__(): + name = cls.__name__ + module_name = cls.__module__ + print(f"Class: {name} and {module_name}") + module = importlib.import_module(module_name) + class_ = getattr(module, name) + cmd_name = f"{name.lower().replace('upscaler', '')}-models-path" + print(f"CMD Name: {cmd_name}") + opt_string = None + try: + opt_string = shared.opts.__getattr__(cmd_name) + except: + pass + scaler = class_(opt_string) + for child in scaler.scalers: + print(f"Appending {child.name}") + datas.append(child) + + shared.sd_upscalers = datas + + # for scaler in subclasses: + # print(f"Found scaler: {type(scaler).__name__}") + # try: + # scaler = scaler() + # for child in scaler.scalers: + # print(f"Appending {child.name}") + # datas.append[child] + # except: + # pass + # shared.sd_upscalers = datas diff --git a/modules/realesrgan_model.py b/modules/realesrgan_model.py index 458bf678..0a2eb896 100644 --- a/modules/realesrgan_model.py +++ b/modules/realesrgan_model.py @@ -1,64 +1,135 @@ import os import sys import traceback -from collections import namedtuple import numpy as np from PIL import Image from basicsr.utils.download_util import load_file_from_url from realesrgan import RealESRGANer -import modules.images +from modules.upscaler import Upscaler, UpscalerData from modules.paths import models_path from modules.shared import cmd_opts, opts -model_dir = "RealESRGAN" -model_path = os.path.join(models_path, model_dir) -cmd_dir = None -RealesrganModelInfo = namedtuple("RealesrganModelInfo", ["name", "location", "model", "netscale"]) -realesrgan_models = [] -have_realesrgan = False + +class UpscalerRealESRGAN(Upscaler): + def __init__(self, path): + self.name = "RealESRGAN" + self.model_path = os.path.join(models_path, self.name) + self.user_path = path + super().__init__() + try: + from basicsr.archs.rrdbnet_arch import RRDBNet + from realesrgan import RealESRGANer + from realesrgan.archs.srvgg_arch import SRVGGNetCompact + self.enable = True + self.scalers = [] + scalers = self.load_models(path) + for scaler in scalers: + if scaler.name in opts.realesrgan_enabled_models: + self.scalers.append(scaler) + + except Exception: + print("Error importing Real-ESRGAN:", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + self.enable = False + self.scalers = [] + + def do_upscale(self, img, path): + if not self.enable: + return img + + info = self.load_model(path) + if not os.path.exists(info.data_path): + print("Unable to load RealESRGAN model: %s" % info.name) + return img + + upsampler = RealESRGANer( + scale=info.scale, + model_path=info.data_path, + model=info.model(), + half=not cmd_opts.no_half, + tile=opts.ESRGAN_tile, + tile_pad=opts.ESRGAN_tile_overlap, + ) + + upsampled = upsampler.enhance(np.array(img), outscale=info.scale)[0] + + image = Image.fromarray(upsampled) + return image + + def load_model(self, path): + try: + info = None + for scaler in self.scalers: + if scaler.data_path == path: + info = scaler + + if info is None: + print(f"Unable to find model info: {path}") + return None + + model_file = load_file_from_url(url=info.data_path, model_dir=self.model_path, progress=True) + info.data_path = model_file + return info + except Exception as e: + print(f"Error making Real-ESRGAN models list: {e}", file=sys.stderr) + print(traceback.format_exc(), file=sys.stderr) + return None + + def load_models(self, _): + return get_realesrgan_models(self) -def get_realesrgan_models(): +def get_realesrgan_models(scaler): try: from basicsr.archs.rrdbnet_arch import RRDBNet from realesrgan.archs.srvgg_arch import SRVGGNetCompact models = [ - RealesrganModelInfo( - name="Real-ESRGAN General x4x3", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3.pth", - netscale=4, - model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu') + UpscalerData( + name="R-ESRGAN General 4xV3", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-x4v3" + ".pth", + scale=4, + upscaler=scaler, + model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, + act_type='prelu') ), - RealesrganModelInfo( - name="Real-ESRGAN General WDN x4x3", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-wdn-x4v3.pth", - netscale=4, - model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, act_type='prelu') + UpscalerData( + name="R-ESRGAN General WDN 4xV3", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-general-wdn-x4v3.pth", + scale=4, + upscaler=scaler, + model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=32, upscale=4, + act_type='prelu') ), - RealesrganModelInfo( - name="Real-ESRGAN AnimeVideo", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth", - netscale=4, - model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu') + UpscalerData( + name="R-ESRGAN AnimeVideo", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.5.0/realesr-animevideov3.pth", + scale=4, + upscaler=scaler, + model=lambda: SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, + act_type='prelu') ), - RealesrganModelInfo( - name="Real-ESRGAN 4x plus", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth", - netscale=4, + UpscalerData( + name="R-ESRGAN 4x+", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth", + scale=4, + upscaler=scaler, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) ), - RealesrganModelInfo( - name="Real-ESRGAN 4x plus anime 6B", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth", - netscale=4, + UpscalerData( + name="R-ESRGAN 4x+ Anime6B", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth", + scale=4, + upscaler=scaler, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=6, num_grow_ch=32, scale=4) ), - RealesrganModelInfo( - name="Real-ESRGAN 2x plus", - location="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth", - netscale=2, + UpscalerData( + name="R-ESRGAN 2x+", + path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth", + scale=2, + upscaler=scaler, model=lambda: RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2) ), ] @@ -66,69 +137,3 @@ def get_realesrgan_models(): except Exception as e: print("Error making Real-ESRGAN models list:", file=sys.stderr) print(traceback.format_exc(), file=sys.stderr) - - -class UpscalerRealESRGAN(modules.images.Upscaler): - def __init__(self, upscaling, model_index): - self.upscaling = upscaling - self.model_index = model_index - self.name = realesrgan_models[model_index].name - - def do_upscale(self, img): - return upscale_with_realesrgan(img, self.upscaling, self.model_index) - - -def setup_model(dirname): - global model_path - if not os.path.exists(model_path): - os.makedirs(model_path) - - global realesrgan_models - global have_realesrgan - if model_path != dirname: - model_path = dirname - try: - from basicsr.archs.rrdbnet_arch import RRDBNet - from realesrgan import RealESRGANer - from realesrgan.archs.srvgg_arch import SRVGGNetCompact - - realesrgan_models = get_realesrgan_models() - have_realesrgan = True - - for i, model in enumerate(realesrgan_models): - if model.name in opts.realesrgan_enabled_models: - modules.shared.sd_upscalers.append(UpscalerRealESRGAN(model.netscale, i)) - - except Exception: - print("Error importing Real-ESRGAN:", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) - - realesrgan_models = [RealesrganModelInfo('None', '', 0, None)] - have_realesrgan = False - - -def upscale_with_realesrgan(image, RealESRGAN_upscaling, RealESRGAN_model_index): - if not have_realesrgan: - return image - - info = realesrgan_models[RealESRGAN_model_index] - - model = info.model() - model_file = load_file_from_url(url=info.location, model_dir=model_path, progress=True) - if not os.path.exists(model_file): - print("Unable to load RealESRGAN model: %s" % info.name) - return image - - upsampler = RealESRGANer( - scale=info.netscale, - model_path=info.location, - model=model, - half=not cmd_opts.no_half, - tile=opts.ESRGAN_tile, - tile_pad=opts.ESRGAN_tile_overlap, - ) - - upsampled = upsampler.enhance(np.array(image), outscale=RealESRGAN_upscaling)[0] - - image = Image.fromarray(upsampled) - return image diff --git a/modules/sd_models.py b/modules/sd_models.py index 89b7d276..23826727 100644 --- a/modules/sd_models.py +++ b/modules/sd_models.py @@ -50,7 +50,7 @@ def setup_model(dirname): if not os.path.exists(model_path): os.makedirs(model_path) checkpoints_list.clear() - model_list = modelloader.load_models(model_path, model_url, dirname, model_name, ext_filter=".ckpt") + model_list = modelloader.load_models(model_path=model_path, model_url=model_url, command_path=dirname, download_name=model_name, ext_filter=".ckpt") cmd_ckpt = shared.cmd_opts.ckpt if os.path.exists(cmd_ckpt): @@ -68,6 +68,7 @@ def setup_model(dirname): def model_hash(filename): try: + print(f"Opening: {filename}") with open(filename, "rb") as file: import hashlib m = hashlib.sha256() diff --git a/modules/sd_samplers.py b/modules/sd_samplers.py index 666ee1ee..cfc3ee40 100644 --- a/modules/sd_samplers.py +++ b/modules/sd_samplers.py @@ -154,9 +154,9 @@ class VanillaStableDiffusionSampler: # existing code fails with cetin step counts, like 9 try: - samples_ddim, _ = self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=p.ddim_eta) + samples_ddim, _ = self.sampler.sample(S=steps, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_t=x, eta=p.ddim_eta) except Exception: - samples_ddim, _ = self.sampler.sample(S=steps+1, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_T=x, eta=p.ddim_eta) + samples_ddim, _ = self.sampler.sample(S=steps+1, conditioning=conditioning, batch_size=int(x.shape[0]), shape=x[0].shape, verbose=False, unconditional_guidance_scale=p.cfg_scale, unconditional_conditioning=unconditional_conditioning, x_t=x, eta=p.ddim_eta) return samples_ddim diff --git a/modules/shared.py b/modules/shared.py index c27079eb..4c31039d 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -1,18 +1,19 @@ -import sys import argparse +import datetime import json import os +import sys + import gradio as gr import tqdm -import datetime import modules.artists -from modules.paths import script_path, sd_path -from modules.devices import get_optimal_device -import modules.styles import modules.interrogate import modules.memmon import modules.sd_models +import modules.styles +from modules.devices import get_optimal_device +from modules.paths import script_path, sd_path sd_model_file = os.path.join(script_path, 'model.ckpt') default_sd_model_file = sd_model_file @@ -38,6 +39,7 @@ parser.add_argument("--share", action='store_true', help="use share=True for gra parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model file(s).", default=os.path.join(model_path, 'Codeformer')) parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model file(s).", default=os.path.join(model_path, 'GFPGAN')) parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN model file(s).", default=os.path.join(model_path, 'ESRGAN')) +parser.add_argument("--bsrgan-models-path", type=str, help="Path to directory with BSRGAN model file(s).", default=os.path.join(model_path, 'BSRGAN')) parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN model file(s).", default=os.path.join(model_path, 'RealESRGAN')) parser.add_argument("--stablediffusion-models-path", type=str, help="Path to directory with Stable-diffusion checkpoints.", default=os.path.join(model_path, 'SwinIR')) parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(model_path, 'SwinIR')) @@ -111,7 +113,7 @@ face_restorers = [] def realesrgan_models_names(): import modules.realesrgan_model - return [x.name for x in modules.realesrgan_model.get_realesrgan_models()] + return [x.name for x in modules.realesrgan_model.get_realesrgan_models(None)] class OptionInfo: @@ -176,13 +178,11 @@ options_templates.update(options_section(('saving-to-dirs', "Saving to a directo options_templates.update(options_section(('upscaling', "Upscaling"), { "ESRGAN_tile": OptionInfo(192, "Tile size for ESRGAN upscalers. 0 = no tiling.", gr.Slider, {"minimum": 0, "maximum": 512, "step": 16}), "ESRGAN_tile_overlap": OptionInfo(8, "Tile overlap, in pixels for ESRGAN upscalers. Low values = visible seam.", gr.Slider, {"minimum": 0, "maximum": 48, "step": 1}), - "realesrgan_enabled_models": OptionInfo(["Real-ESRGAN 4x plus", "Real-ESRGAN 4x plus anime 6B"], "Select which RealESRGAN models to show in the web UI. (Requires restart)", gr.CheckboxGroup, lambda: {"choices": realesrgan_models_names()}), + "realesrgan_enabled_models": OptionInfo(["R-ESRGAN x4+", "R-ESRGAN x4+ Anime6B"], "Select which Real-ESRGAN models to show in the web UI. (Requires restart)", gr.CheckboxGroup, lambda: {"choices": realesrgan_models_names()}), "SWIN_tile": OptionInfo(192, "Tile size for all SwinIR.", gr.Slider, {"minimum": 16, "maximum": 512, "step": 16}), "SWIN_tile_overlap": OptionInfo(8, "Tile overlap, in pixels for SwinIR. Low values = visible seam.", gr.Slider, {"minimum": 0, "maximum": 48, "step": 1}), "ldsr_steps": OptionInfo(100, "LDSR processing steps. Lower = faster", gr.Slider, {"minimum": 1, "maximum": 200, "step": 1}), "ldsr_pre_down": OptionInfo(1, "LDSR Pre-process downssample scale. 1 = no down-sampling, 4 = 1/4 scale.", gr.Slider, {"minimum": 1, "maximum": 4, "step": 1}), - "ldsr_post_down": OptionInfo(1, "LDSR Post-process down-sample scale. 1 = no down-sampling, 4 = 1/4 scale.", gr.Slider, {"minimum": 1, "maximum": 4, "step": 1}), - "upscaler_for_img2img": OptionInfo(None, "Upscaler for img2img", gr.Radio, lambda: {"choices": [x.name for x in sd_upscalers]}), })) diff --git a/modules/swinir_model.py b/modules/swinir_model.py index f515779e..ea7b6301 100644 --- a/modules/swinir_model.py +++ b/modules/swinir_model.py @@ -1,92 +1,91 @@ import contextlib import os -import sys -import traceback import numpy as np import torch from PIL import Image from basicsr.utils.download_util import load_file_from_url -import modules.images from modules import modelloader from modules.paths import models_path from modules.shared import cmd_opts, opts, device from modules.swinir_model_arch import SwinIR as net +from modules.upscaler import Upscaler, UpscalerData -model_dir = "SwinIR" -model_url = "https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth" -model_name = "SwinIR x4" -model_path = os.path.join(models_path, model_dir) -cmd_path = "" precision_scope = ( torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext ) -def load_model(path, scale=4): - global model_path - global model_name - if "http" in path: - dl_name = "%s%s" % (model_name.replace(" ", "_"), ".pth") - filename = load_file_from_url(url=path, model_dir=model_path, file_name=dl_name, progress=True) - else: - filename = path - if filename is None or not os.path.exists(filename): - return None - model = net( - upscale=scale, - in_chans=3, - img_size=64, - window_size=8, - img_range=1.0, - depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], - embed_dim=240, - num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8], - mlp_ratio=2, - upsampler="nearest+conv", - resi_connection="3conv", - ) +class UpscalerSwinIR(Upscaler): + def __init__(self, dirname): + self.name = "SwinIR" + self.model_url = "https://github.com/JingyunLiang/SwinIR/releases/download/v0.0" \ + "/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR" \ + "-L_x4_GAN.pth " + self.model_name = "SwinIR 4x" + self.model_path = os.path.join(models_path, self.name) + self.user_path = dirname + super().__init__() + scalers = [] + model_files = self.find_models(ext_filter=[".pt", ".pth"]) + for model in model_files: + if "http" in model: + name = self.model_name + else: + name = modelloader.friendly_name(model) + model_data = UpscalerData(name, model, self) + scalers.append(model_data) + self.scalers = scalers - pretrained_model = torch.load(filename) - model.load_state_dict(pretrained_model["params_ema"], strict=True) - if not cmd_opts.no_half: - model = model.half() - return model + def do_upscale(self, img, model_file): + model = self.load_model(model_file) + if model is None: + return img + model = model.to(device) + img = upscale(img, model) + try: + torch.cuda.empty_cache() + except: + pass + return img - -def setup_model(dirname): - global model_path - global model_name - global cmd_path - if not os.path.exists(model_path): - os.makedirs(model_path) - cmd_path = dirname - model_file = "" - try: - models = modelloader.load_models(model_path, ext_filter=[".pt", ".pth"], command_path=cmd_path) - - if len(models) != 0: - model_file = models[0] - name = modelloader.friendly_name(model_file) + def load_model(self, path, scale=4): + if "http" in path: + dl_name = "%s%s" % (self.name.replace(" ", "_"), ".pth") + filename = load_file_from_url(url=path, model_dir=self.model_path, file_name=dl_name, progress=True) else: - # Add the "default" model if none are found. - model_file = model_url - name = model_name + filename = path + if filename is None or not os.path.exists(filename): + return None + model = net( + upscale=scale, + in_chans=3, + img_size=64, + window_size=8, + img_range=1.0, + depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], + embed_dim=240, + num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8], + mlp_ratio=2, + upsampler="nearest+conv", + resi_connection="3conv", + ) - modules.shared.sd_upscalers.append(UpscalerSwin(model_file, name)) - except Exception: - print(f"Error loading SwinIR model: {model_file}", file=sys.stderr) - print(traceback.format_exc(), file=sys.stderr) + pretrained_model = torch.load(filename) + model.load_state_dict(pretrained_model["params_ema"], strict=True) + if not cmd_opts.no_half: + model = model.half() + return model def upscale( - img, - model, - tile=opts.SWIN_tile, - tile_overlap=opts.SWIN_tile_overlap, - window_size=8, - scale=4, + img, + model, + tile=opts.SWIN_tile, + tile_overlap=opts.SWIN_tile_overlap, + window_size=8, + scale=4, ): img = np.array(img) img = img[:, :, ::-1] @@ -125,34 +124,16 @@ def inference(img, model, tile, tile_overlap, window_size, scale): for h_idx in h_idx_list: for w_idx in w_idx_list: - in_patch = img[..., h_idx : h_idx + tile, w_idx : w_idx + tile] + in_patch = img[..., h_idx: h_idx + tile, w_idx: w_idx + tile] out_patch = model(in_patch) out_patch_mask = torch.ones_like(out_patch) E[ - ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf + ..., h_idx * sf: (h_idx + tile) * sf, w_idx * sf: (w_idx + tile) * sf ].add_(out_patch) W[ - ..., h_idx * sf : (h_idx + tile) * sf, w_idx * sf : (w_idx + tile) * sf + ..., h_idx * sf: (h_idx + tile) * sf, w_idx * sf: (w_idx + tile) * sf ].add_(out_patch_mask) output = E.div_(W) return output - - -class UpscalerSwin(modules.images.Upscaler): - def __init__(self, filename, title): - self.name = title - self.filename = filename - - def do_upscale(self, img): - model = load_model(self.filename) - if model is None: - return img - model = model.to(device) - img = upscale(img, model) - try: - torch.cuda.empty_cache() - except: - pass - return img \ No newline at end of file diff --git a/modules/upscaler.py b/modules/upscaler.py new file mode 100644 index 00000000..d698282f --- /dev/null +++ b/modules/upscaler.py @@ -0,0 +1,121 @@ +import os +from abc import abstractmethod + +import PIL +import numpy as np +import torch +from PIL import Image + +import modules.shared +from modules import modelloader, shared + +LANCZOS = (Image.Resampling.LANCZOS if hasattr(Image, 'Resampling') else Image.LANCZOS) +from modules.paths import models_path + + +class Upscaler: + name = None + model_path = None + model_name = None + model_url = None + enable = True + filter = None + model = None + user_path = None + scalers: [] + tile = True + + def __init__(self, create_dirs=False): + self.mod_pad_h = None + self.tile_size = modules.shared.opts.ESRGAN_tile + self.tile_pad = modules.shared.opts.ESRGAN_tile_overlap + self.device = modules.shared.device + self.img = None + self.output = None + self.scale = 1 + self.half = not modules.shared.cmd_opts.no_half + self.pre_pad = 0 + self.mod_scale = None + if self.name is not None and create_dirs: + self.model_path = os.path.join(models_path, self.name) + if not os.path.exists(self.model_path): + os.makedirs(self.model_path) + + try: + import cv2 + self.can_tile = True + except: + pass + + @abstractmethod + def do_upscale(self, img: PIL.Image, selected_model: str): + return img + + def upscale(self, img: PIL.Image, scale: int, selected_model: str = None): + self.scale = scale + dest_w = img.width * scale + dest_h = img.height * scale + for i in range(3): + if img.width >= dest_w and img.height >= dest_h: + break + img = self.do_upscale(img, selected_model) + if img.width != dest_w or img.height != dest_h: + img = img.resize(dest_w, dest_h, resample=LANCZOS) + + return img + + @abstractmethod + def load_model(self, path: str): + pass + + def find_models(self, ext_filter=None) -> list: + return modelloader.load_models(model_path=self.model_path, model_url=self.model_url, command_path=self.user_path) + + def update_status(self, prompt): + print(f"\nextras: {prompt}", file=shared.progress_print_out) + + +class UpscalerData: + name = None + data_path = None + scale: int = 4 + scaler: Upscaler = None + model: None + + def __init__(self, name: str, path: str, upscaler: Upscaler = None, scale: int = 4, model=None): + self.name = name + self.data_path = path + self.scaler = upscaler + self.scale = scale + self.model = model + + +class UpscalerNone(Upscaler): + name = "None" + scalers = [] + + def load_model(self, path): + pass + + def do_upscale(self, img, selected_model=None): + return img + + def __init__(self, dirname=None): + super().__init__(False) + self.scalers = [UpscalerData("None", None, self)] + + +class UpscalerLanczos(Upscaler): + scalers = [] + + def do_upscale(self, img, selected_model=None): + return img.resize((int(img.width * self.scale), int(img.height * self.scale)), resample=LANCZOS) + + def load_model(self, _): + pass + + def __init__(self, dirname=None): + super().__init__(False) + self.name = "Lanczos" + self.scalers = [UpscalerData("Lanczos", None, self)] + diff --git a/webui.py b/webui.py index 76d392a2..be1bc769 100644 --- a/webui.py +++ b/webui.py @@ -1,9 +1,10 @@ import os import signal import threading - +import modules.paths import modules.codeformer_model as codeformer import modules.esrgan_model as esrgan +import modules.bsrgan_model as bsrgan import modules.extras import modules.face_restoration import modules.gfpgan_model as gfpgan @@ -27,11 +28,7 @@ modules.sd_models.setup_model(cmd_opts.stablediffusion_models_path) codeformer.setup_model(cmd_opts.codeformer_models_path) gfpgan.setup_model(cmd_opts.gfpgan_models_path) shared.face_restorers.append(modules.face_restoration.FaceRestoration()) - -esrgan.setup_model(cmd_opts.esrgan_models_path) -swinir.setup_model(cmd_opts.swinir_models_path) -realesrgan.setup_model(cmd_opts.realesrgan_models_path) -ldsr.setup_model(cmd_opts.ldsr_models_path) +modelloader.load_upscalers() queue_lock = threading.Lock() From 435fd2112aee9a0e61408ac56663e41beea1e446 Mon Sep 17 00:00:00 2001 From: d8ahazard Date: Thu, 29 Sep 2022 19:59:53 -0500 Subject: [PATCH 10/10] Fixes, cleanup. --- modules/ldsr_model_arch.py | 6 ++++-- modules/modelloader.py | 14 -------------- modules/swinir_model.py | 2 +- 3 files changed, 5 insertions(+), 17 deletions(-) diff --git a/modules/ldsr_model_arch.py b/modules/ldsr_model_arch.py index 8fe87c6a..f8f3c3d3 100644 --- a/modules/ldsr_model_arch.py +++ b/modules/ldsr_model_arch.py @@ -101,8 +101,10 @@ class LDSR: print("Foo") down_sample_rate = target_scale / 4 print(f"Downsample rate is {down_sample_rate}") - width_downsampled_pre = width_og * down_sample_rate - height_downsampled_pre = height_og * down_sample_method + wd = width_og * down_sample_rate + hd = height_og * down_sample_rate + width_downsampled_pre = int(wd) + height_downsampled_pre = int(hd) if down_sample_rate != 1: print( diff --git a/modules/modelloader.py b/modules/modelloader.py index 6de65c69..51b3ecd5 100644 --- a/modules/modelloader.py +++ b/modules/modelloader.py @@ -118,11 +118,9 @@ def load_upscalers(): for cls in Upscaler.__subclasses__(): name = cls.__name__ module_name = cls.__module__ - print(f"Class: {name} and {module_name}") module = importlib.import_module(module_name) class_ = getattr(module, name) cmd_name = f"{name.lower().replace('upscaler', '')}-models-path" - print(f"CMD Name: {cmd_name}") opt_string = None try: opt_string = shared.opts.__getattr__(cmd_name) @@ -130,18 +128,6 @@ def load_upscalers(): pass scaler = class_(opt_string) for child in scaler.scalers: - print(f"Appending {child.name}") datas.append(child) shared.sd_upscalers = datas - - # for scaler in subclasses: - # print(f"Found scaler: {type(scaler).__name__}") - # try: - # scaler = scaler() - # for child in scaler.scalers: - # print(f"Appending {child.name}") - # datas.append[child] - # except: - # pass - # shared.sd_upscalers = datas diff --git a/modules/swinir_model.py b/modules/swinir_model.py index ea7b6301..41fda5a7 100644 --- a/modules/swinir_model.py +++ b/modules/swinir_model.py @@ -52,7 +52,7 @@ class UpscalerSwinIR(Upscaler): def load_model(self, path, scale=4): if "http" in path: - dl_name = "%s%s" % (self.name.replace(" ", "_"), ".pth") + dl_name = "%s%s" % (self.model_name.replace(" ", "_"), ".pth") filename = load_file_from_url(url=path, model_dir=self.model_path, file_name=dl_name, progress=True) else: filename = path