From 4b8a192f680101de247dca79e48974b53bf961fe Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Sat, 29 Oct 2022 16:36:43 +0900 Subject: [PATCH 01/24] add optimizer save option to shared.opts --- modules/shared.py | 1 + 1 file changed, 1 insertion(+) diff --git a/modules/shared.py b/modules/shared.py index e4f163c1..065b893d 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -286,6 +286,7 @@ options_templates.update(options_section(('system', "System"), { options_templates.update(options_section(('training', "Training"), { "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training hypernetwork. Saves VRAM."), + "save_optimizer_state": OptionInfo(False, "Saves Optimizer state with checkpoints. This will cause file size to increase VERY much."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), From 20194fd9752a280306fb66b57b258609b0918c46 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Sat, 29 Oct 2022 16:56:42 +0900 Subject: [PATCH 02/24] We have duplicate linear now --- modules/hypernetworks/ui.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/ui.py b/modules/hypernetworks/ui.py index aad09ffc..c2d4b51c 100644 --- a/modules/hypernetworks/ui.py +++ b/modules/hypernetworks/ui.py @@ -9,7 +9,7 @@ from modules import devices, sd_hijack, shared from modules.hypernetworks import hypernetwork not_available = ["hardswish", "multiheadattention"] -keys = ["linear"] + list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available) +keys = list(x for x in hypernetwork.HypernetworkModule.activation_dict.keys() if x not in not_available) def create_hypernetwork(name, enable_sizes, overwrite_old, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False): # Remove illegal characters from name. From 9d96d7d0a0aa0a966a9aefd24342345eb65952ed Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Sun, 30 Oct 2022 20:39:04 +0900 Subject: [PATCH 03/24] resolve conflicts --- modules/hypernetworks/hypernetwork.py | 44 +++++++++++++++++++++++---- 1 file changed, 38 insertions(+), 6 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index a11e01d6..8f74cdea 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -21,6 +21,7 @@ from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_norm from collections import defaultdict, deque from statistics import stdev, mean +optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"} class HypernetworkModule(torch.nn.Module): multiplier = 1.0 @@ -139,6 +140,8 @@ class Hypernetwork: self.weight_init = weight_init self.add_layer_norm = add_layer_norm self.use_dropout = use_dropout + self.optimizer_name = None + self.optimizer_state_dict = None for size in enable_sizes or []: self.layers[size] = ( @@ -171,6 +174,10 @@ class Hypernetwork: state_dict['use_dropout'] = self.use_dropout state_dict['sd_checkpoint'] = self.sd_checkpoint state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name + if self.optimizer_name is not None: + state_dict['optimizer_name'] = self.optimizer_name + if self.optimizer_state_dict: + state_dict['optimizer_state_dict'] = self.optimizer_state_dict torch.save(state_dict, filename) @@ -190,7 +197,14 @@ class Hypernetwork: self.add_layer_norm = state_dict.get('is_layer_norm', False) print(f"Layer norm is set to {self.add_layer_norm}") self.use_dropout = state_dict.get('use_dropout', False) - print(f"Dropout usage is set to {self.use_dropout}" ) + print(f"Dropout usage is set to {self.use_dropout}") + self.optimizer_name = state_dict.get('optimizer_name', 'AdamW') + print(f"Optimizer name is {self.optimizer_name}") + self.optimizer_state_dict = state_dict.get('optimizer_state_dict', None) + if self.optimizer_state_dict: + print("Loaded existing optimizer from checkpoint") + else: + print("No saved optimizer exists in checkpoint") for size, sd in state_dict.items(): if type(size) == int: @@ -392,8 +406,19 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log weights = hypernetwork.weights() for weight in weights: weight.requires_grad = True - # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... - optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) + # Here we use optimizer from saved HN, or we can specify as UI option. + if (optimizer_name := hypernetwork.optimizer_name) in optimizer_dict: + optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate) + else: + print(f"Optimizer type {optimizer_name} is not defined!") + optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate) + optimizer_name = 'AdamW' + if hypernetwork.optimizer_state_dict: # This line must be changed if Optimizer type can be different from saved optimizer. + try: + optimizer.load_state_dict(hypernetwork.optimizer_state_dict) + except RuntimeError as e: + print("Cannot resume from saved optimizer!") + print(e) steps_without_grad = 0 @@ -455,8 +480,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log # Before saving, change name to match current checkpoint. hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}' last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt') + hypernetwork.optimizer_name = optimizer_name + if shared.opts.save_optimizer_state: + hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file) - + hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), { "loss": f"{previous_mean_loss:.7f}", "learn_rate": scheduler.learn_rate @@ -514,14 +542,18 @@ Last saved hypernetwork: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" - report_statistics(loss_dict) filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') + hypernetwork.optimizer_name = optimizer_name + if shared.opts.save_optimizer_state: + hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename) - + del optimizer + hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. return hypernetwork, filename + def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename): old_hypernetwork_name = hypernetwork.name old_sd_checkpoint = hypernetwork.sd_checkpoint if hasattr(hypernetwork, "sd_checkpoint") else None From 3178c35224467893cf8dcedb1028c59c6c23db58 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Wed, 2 Nov 2022 22:16:32 +0900 Subject: [PATCH 04/24] resolve conflicts --- modules/shared.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/shared.py b/modules/shared.py index 065b893d..959937d7 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -285,7 +285,7 @@ options_templates.update(options_section(('system', "System"), { })) options_templates.update(options_section(('training', "Training"), { - "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training hypernetwork. Saves VRAM."), + "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."), "save_optimizer_state": OptionInfo(False, "Saves Optimizer state with checkpoints. This will cause file size to increase VERY much."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), From 9b5f85ac83f864310fe19c9deab6670bad695b0d Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Wed, 2 Nov 2022 22:18:04 +0900 Subject: [PATCH 05/24] first revert --- modules/shared.py | 1 - 1 file changed, 1 deletion(-) diff --git a/modules/shared.py b/modules/shared.py index 959937d7..7e8c552b 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -286,7 +286,6 @@ options_templates.update(options_section(('system', "System"), { options_templates.update(options_section(('training', "Training"), { "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."), - "save_optimizer_state": OptionInfo(False, "Saves Optimizer state with checkpoints. This will cause file size to increase VERY much."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), From 7ea5956ad5fa925f92116e8a3bf78d7f6517b654 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Wed, 2 Nov 2022 22:18:55 +0900 Subject: [PATCH 06/24] now add --- modules/shared.py | 1 + 1 file changed, 1 insertion(+) diff --git a/modules/shared.py b/modules/shared.py index d8e99f85..7ecb40d8 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -309,6 +309,7 @@ options_templates.update(options_section(('system', "System"), { options_templates.update(options_section(('training', "Training"), { "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."), + "save_optimizer_state": OptionInfo(False, "Saves Optimizer state with checkpoints. This will cause file size to increase VERY much."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), From 0b143c1163a96b193a4e8512be9c5831c661a50d Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Thu, 3 Nov 2022 14:30:53 +0900 Subject: [PATCH 07/24] Separate .optim file from model --- modules/hypernetworks/hypernetwork.py | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 8f74cdea..63c25de8 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -161,6 +161,7 @@ class Hypernetwork: def save(self, filename): state_dict = {} + optimizer_saved_dict = {} for k, v in self.layers.items(): state_dict[k] = (v[0].state_dict(), v[1].state_dict()) @@ -175,9 +176,10 @@ class Hypernetwork: state_dict['sd_checkpoint'] = self.sd_checkpoint state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name if self.optimizer_name is not None: - state_dict['optimizer_name'] = self.optimizer_name + optimizer_saved_dict['optimizer_name'] = self.optimizer_name if self.optimizer_state_dict: - state_dict['optimizer_state_dict'] = self.optimizer_state_dict + optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict + torch.save(optimizer_saved_dict, filename + '.optim') torch.save(state_dict, filename) @@ -198,9 +200,11 @@ class Hypernetwork: print(f"Layer norm is set to {self.add_layer_norm}") self.use_dropout = state_dict.get('use_dropout', False) print(f"Dropout usage is set to {self.use_dropout}") - self.optimizer_name = state_dict.get('optimizer_name', 'AdamW') + + optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {} + self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW') print(f"Optimizer name is {self.optimizer_name}") - self.optimizer_state_dict = state_dict.get('optimizer_state_dict', None) + self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None) if self.optimizer_state_dict: print("Loaded existing optimizer from checkpoint") else: From 1764ac3c8bc482bd575987850e96630d9115e51a Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Thu, 3 Nov 2022 14:49:26 +0900 Subject: [PATCH 08/24] use hash to check valid optim --- modules/hypernetworks/hypernetwork.py | 15 ++++++++++----- 1 file changed, 10 insertions(+), 5 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 63c25de8..4230b8cf 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -177,11 +177,12 @@ class Hypernetwork: state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name if self.optimizer_name is not None: optimizer_saved_dict['optimizer_name'] = self.optimizer_name - if self.optimizer_state_dict: - optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict - torch.save(optimizer_saved_dict, filename + '.optim') torch.save(state_dict, filename) + if self.optimizer_state_dict: + optimizer_saved_dict['hash'] = sd_models.model_hash(filename) + optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict + torch.save(optimizer_saved_dict, filename + '.optim') def load(self, filename): self.filename = filename @@ -204,7 +205,10 @@ class Hypernetwork: optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {} self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW') print(f"Optimizer name is {self.optimizer_name}") - self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None) + if sd_models.model_hash(filename) == optimizer_saved_dict.get('hash', None): + self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None) + else: + self.optimizer_state_dict = None if self.optimizer_state_dict: print("Loaded existing optimizer from checkpoint") else: @@ -229,7 +233,7 @@ def list_hypernetworks(path): name = os.path.splitext(os.path.basename(filename))[0] # Prevent a hypothetical "None.pt" from being listed. if name != "None": - res[name] = filename + res[name + f"({sd_models.model_hash(filename)})"] = filename return res @@ -375,6 +379,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log else: hypernetwork_dir = None + hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0] if create_image_every > 0: images_dir = os.path.join(log_directory, "images") os.makedirs(images_dir, exist_ok=True) From 0abb39f461baa343ae7c23abffb261e57c3168d4 Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 15:47:19 +0900 Subject: [PATCH 09/24] resolve conflict - first revert --- modules/hypernetworks/hypernetwork.py | 123 +++++++++++--------------- 1 file changed, 52 insertions(+), 71 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 4230b8cf..674fcedd 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -21,7 +21,6 @@ from torch.nn.init import normal_, xavier_normal_, xavier_uniform_, kaiming_norm from collections import defaultdict, deque from statistics import stdev, mean -optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"} class HypernetworkModule(torch.nn.Module): multiplier = 1.0 @@ -34,9 +33,12 @@ class HypernetworkModule(torch.nn.Module): "tanh": torch.nn.Tanh, "sigmoid": torch.nn.Sigmoid, } - activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) + activation_dict.update( + {cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if + inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) - def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal', add_layer_norm=False, use_dropout=False): + def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal', + add_layer_norm=False, use_dropout=False): super().__init__() assert layer_structure is not None, "layer_structure must not be None" @@ -47,7 +49,7 @@ class HypernetworkModule(torch.nn.Module): for i in range(len(layer_structure) - 1): # Add a fully-connected layer - linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) + linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i + 1]))) # Add an activation func if activation_func == "linear" or activation_func is None: @@ -59,7 +61,7 @@ class HypernetworkModule(torch.nn.Module): # Add layer normalization if add_layer_norm: - linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) + linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i + 1]))) # Add dropout expect last layer if use_dropout and i < len(layer_structure) - 3: @@ -128,7 +130,8 @@ class Hypernetwork: filename = None name = None - def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False): + def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, + add_layer_norm=False, use_dropout=False): self.filename = None self.name = name self.layers = {} @@ -140,13 +143,13 @@ class Hypernetwork: self.weight_init = weight_init self.add_layer_norm = add_layer_norm self.use_dropout = use_dropout - self.optimizer_name = None - self.optimizer_state_dict = None for size in enable_sizes or []: self.layers[size] = ( - HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), - HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, + self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, + self.add_layer_norm, self.use_dropout), ) def weights(self): @@ -161,7 +164,6 @@ class Hypernetwork: def save(self, filename): state_dict = {} - optimizer_saved_dict = {} for k, v in self.layers.items(): state_dict[k] = (v[0].state_dict(), v[1].state_dict()) @@ -175,14 +177,8 @@ class Hypernetwork: state_dict['use_dropout'] = self.use_dropout state_dict['sd_checkpoint'] = self.sd_checkpoint state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name - if self.optimizer_name is not None: - optimizer_saved_dict['optimizer_name'] = self.optimizer_name torch.save(state_dict, filename) - if self.optimizer_state_dict: - optimizer_saved_dict['hash'] = sd_models.model_hash(filename) - optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict - torch.save(optimizer_saved_dict, filename + '.optim') def load(self, filename): self.filename = filename @@ -202,23 +198,13 @@ class Hypernetwork: self.use_dropout = state_dict.get('use_dropout', False) print(f"Dropout usage is set to {self.use_dropout}") - optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {} - self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW') - print(f"Optimizer name is {self.optimizer_name}") - if sd_models.model_hash(filename) == optimizer_saved_dict.get('hash', None): - self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None) - else: - self.optimizer_state_dict = None - if self.optimizer_state_dict: - print("Loaded existing optimizer from checkpoint") - else: - print("No saved optimizer exists in checkpoint") - for size, sd in state_dict.items(): if type(size) == int: self.layers[size] = ( - HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), - HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init, + self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init, + self.add_layer_norm, self.use_dropout), ) self.name = state_dict.get('name', self.name) @@ -233,7 +219,7 @@ def list_hypernetworks(path): name = os.path.splitext(os.path.basename(filename))[0] # Prevent a hypothetical "None.pt" from being listed. if name != "None": - res[name + f"({sd_models.model_hash(filename)})"] = filename + res[name] = filename return res @@ -330,7 +316,7 @@ def statistics(data): std = 0 else: std = stdev(data) - total_information = f"loss:{mean(data):.3f}" + u"\u00B1" + f"({std/ (len(data) ** 0.5):.3f})" + total_information = f"loss:{mean(data):.3f}" + u"\u00B1" + f"({std / (len(data) ** 0.5):.3f})" recent_data = data[-32:] if len(recent_data) < 2: std = 0 @@ -340,7 +326,7 @@ def statistics(data): return total_information, recent_information -def report_statistics(loss_info:dict): +def report_statistics(loss_info: dict): keys = sorted(loss_info.keys(), key=lambda x: sum(loss_info[x]) / len(loss_info[x])) for key in keys: try: @@ -352,14 +338,18 @@ def report_statistics(loss_info:dict): print(e) - -def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): +def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, + training_height, steps, create_image_every, save_hypernetwork_every, template_file, + preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, + preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): # images allows training previews to have infotext. Importing it at the top causes a circular import problem. from modules import images save_hypernetwork_every = save_hypernetwork_every or 0 create_image_every = create_image_every or 0 - textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork") + textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, + save_hypernetwork_every, create_image_every, log_directory, + name="hypernetwork") path = shared.hypernetworks.get(hypernetwork_name, None) shared.loaded_hypernetwork = Hypernetwork() @@ -379,7 +369,6 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log else: hypernetwork_dir = None - hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0] if create_image_every > 0: images_dir = os.path.join(log_directory, "images") os.makedirs(images_dir, exist_ok=True) @@ -395,39 +384,34 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - + # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, + height=training_height, + repeats=shared.opts.training_image_repeats_per_epoch, + placeholder_token=hypernetwork_name, + model=shared.sd_model, device=devices.device, + template_file=template_file, include_cond=True, + batch_size=batch_size) if unload: shared.sd_model.cond_stage_model.to(devices.cpu) shared.sd_model.first_stage_model.to(devices.cpu) size = len(ds.indexes) - loss_dict = defaultdict(lambda : deque(maxlen = 1024)) + loss_dict = defaultdict(lambda: deque(maxlen=1024)) losses = torch.zeros((size,)) previous_mean_losses = [0] previous_mean_loss = 0 print("Mean loss of {} elements".format(size)) - + weights = hypernetwork.weights() for weight in weights: weight.requires_grad = True - # Here we use optimizer from saved HN, or we can specify as UI option. - if (optimizer_name := hypernetwork.optimizer_name) in optimizer_dict: - optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate) - else: - print(f"Optimizer type {optimizer_name} is not defined!") - optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate) - optimizer_name = 'AdamW' - if hypernetwork.optimizer_state_dict: # This line must be changed if Optimizer type can be different from saved optimizer. - try: - optimizer.load_state_dict(hypernetwork.optimizer_state_dict) - except RuntimeError as e: - print("Cannot resume from saved optimizer!") - print(e) + # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... + optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) steps_without_grad = 0 @@ -441,7 +425,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if len(loss_dict) > 0: previous_mean_losses = [i[-1] for i in loss_dict.values()] previous_mean_loss = mean(previous_mean_losses) - + scheduler.apply(optimizer, hypernetwork.step) if scheduler.finished: break @@ -460,7 +444,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log losses[hypernetwork.step % losses.shape[0]] = loss.item() for entry in entries: loss_dict[entry.filename].append(loss.item()) - + optimizer.zero_grad() weights[0].grad = None loss.backward() @@ -475,9 +459,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log steps_done = hypernetwork.step + 1 - if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): + if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): raise RuntimeError("Loss diverged.") - + if len(previous_mean_losses) > 1: std = stdev(previous_mean_losses) else: @@ -489,11 +473,8 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log # Before saving, change name to match current checkpoint. hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}' last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt') - hypernetwork.optimizer_name = optimizer_name - if shared.opts.save_optimizer_state: - hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file) - hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. + textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), { "loss": f"{previous_mean_loss:.7f}", "learn_rate": scheduler.learn_rate @@ -529,7 +510,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log preview_text = p.prompt processed = processing.process_images(p) - image = processed.images[0] if len(processed.images)>0 else None + image = processed.images[0] if len(processed.images) > 0 else None if unload: shared.sd_model.cond_stage_model.to(devices.cpu) @@ -537,7 +518,10 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if image is not None: shared.state.current_image = image - last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False) + last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, + shared.opts.samples_format, processed.infotexts[0], + p=p, forced_filename=forced_filename, + save_to_dirs=False) last_saved_image += f", prompt: {preview_text}" shared.state.job_no = hypernetwork.step @@ -551,15 +535,12 @@ Last saved hypernetwork: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" + report_statistics(loss_dict) filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') - hypernetwork.optimizer_name = optimizer_name - if shared.opts.save_optimizer_state: - hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename) - del optimizer - hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. + return hypernetwork, filename @@ -576,4 +557,4 @@ def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename): hypernetwork.sd_checkpoint = old_sd_checkpoint hypernetwork.sd_checkpoint_name = old_sd_checkpoint_name hypernetwork.name = old_hypernetwork_name - raise + raise \ No newline at end of file From 0d07cbfa15d34294a4fa22d74359cdd6fe2f799c Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 15:50:54 +0900 Subject: [PATCH 10/24] I blame code autocomplete --- modules/hypernetworks/hypernetwork.py | 76 ++++++++++----------------- 1 file changed, 27 insertions(+), 49 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 674fcedd..a11e01d6 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -33,12 +33,9 @@ class HypernetworkModule(torch.nn.Module): "tanh": torch.nn.Tanh, "sigmoid": torch.nn.Sigmoid, } - activation_dict.update( - {cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if - inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) + activation_dict.update({cls_name.lower(): cls_obj for cls_name, cls_obj in inspect.getmembers(torch.nn.modules.activation) if inspect.isclass(cls_obj) and cls_obj.__module__ == 'torch.nn.modules.activation'}) - def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal', - add_layer_norm=False, use_dropout=False): + def __init__(self, dim, state_dict=None, layer_structure=None, activation_func=None, weight_init='Normal', add_layer_norm=False, use_dropout=False): super().__init__() assert layer_structure is not None, "layer_structure must not be None" @@ -49,7 +46,7 @@ class HypernetworkModule(torch.nn.Module): for i in range(len(layer_structure) - 1): # Add a fully-connected layer - linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i + 1]))) + linears.append(torch.nn.Linear(int(dim * layer_structure[i]), int(dim * layer_structure[i+1]))) # Add an activation func if activation_func == "linear" or activation_func is None: @@ -61,7 +58,7 @@ class HypernetworkModule(torch.nn.Module): # Add layer normalization if add_layer_norm: - linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i + 1]))) + linears.append(torch.nn.LayerNorm(int(dim * layer_structure[i+1]))) # Add dropout expect last layer if use_dropout and i < len(layer_structure) - 3: @@ -130,8 +127,7 @@ class Hypernetwork: filename = None name = None - def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, - add_layer_norm=False, use_dropout=False): + def __init__(self, name=None, enable_sizes=None, layer_structure=None, activation_func=None, weight_init=None, add_layer_norm=False, use_dropout=False): self.filename = None self.name = name self.layers = {} @@ -146,10 +142,8 @@ class Hypernetwork: for size in enable_sizes or []: self.layers[size] = ( - HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, - self.add_layer_norm, self.use_dropout), - HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, - self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, None, self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), ) def weights(self): @@ -196,15 +190,13 @@ class Hypernetwork: self.add_layer_norm = state_dict.get('is_layer_norm', False) print(f"Layer norm is set to {self.add_layer_norm}") self.use_dropout = state_dict.get('use_dropout', False) - print(f"Dropout usage is set to {self.use_dropout}") + print(f"Dropout usage is set to {self.use_dropout}" ) for size, sd in state_dict.items(): if type(size) == int: self.layers[size] = ( - HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init, - self.add_layer_norm, self.use_dropout), - HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init, - self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, sd[0], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), + HypernetworkModule(size, sd[1], self.layer_structure, self.activation_func, self.weight_init, self.add_layer_norm, self.use_dropout), ) self.name = state_dict.get('name', self.name) @@ -316,7 +308,7 @@ def statistics(data): std = 0 else: std = stdev(data) - total_information = f"loss:{mean(data):.3f}" + u"\u00B1" + f"({std / (len(data) ** 0.5):.3f})" + total_information = f"loss:{mean(data):.3f}" + u"\u00B1" + f"({std/ (len(data) ** 0.5):.3f})" recent_data = data[-32:] if len(recent_data) < 2: std = 0 @@ -326,7 +318,7 @@ def statistics(data): return total_information, recent_information -def report_statistics(loss_info: dict): +def report_statistics(loss_info:dict): keys = sorted(loss_info.keys(), key=lambda x: sum(loss_info[x]) / len(loss_info[x])) for key in keys: try: @@ -338,18 +330,14 @@ def report_statistics(loss_info: dict): print(e) -def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, - training_height, steps, create_image_every, save_hypernetwork_every, template_file, - preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, - preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): + +def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log_directory, training_width, training_height, steps, create_image_every, save_hypernetwork_every, template_file, preview_from_txt2img, preview_prompt, preview_negative_prompt, preview_steps, preview_sampler_index, preview_cfg_scale, preview_seed, preview_width, preview_height): # images allows training previews to have infotext. Importing it at the top causes a circular import problem. from modules import images save_hypernetwork_every = save_hypernetwork_every or 0 create_image_every = create_image_every or 0 - textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, - save_hypernetwork_every, create_image_every, log_directory, - name="hypernetwork") + textual_inversion.validate_train_inputs(hypernetwork_name, learn_rate, batch_size, data_root, template_file, steps, save_hypernetwork_every, create_image_every, log_directory, name="hypernetwork") path = shared.hypernetworks.get(hypernetwork_name, None) shared.loaded_hypernetwork = Hypernetwork() @@ -384,29 +372,23 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log return hypernetwork, filename scheduler = LearnRateScheduler(learn_rate, steps, ititial_step) - + # dataset loading may take a while, so input validations and early returns should be done before this shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..." with torch.autocast("cuda"): - ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, - height=training_height, - repeats=shared.opts.training_image_repeats_per_epoch, - placeholder_token=hypernetwork_name, - model=shared.sd_model, device=devices.device, - template_file=template_file, include_cond=True, - batch_size=batch_size) + ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file, include_cond=True, batch_size=batch_size) if unload: shared.sd_model.cond_stage_model.to(devices.cpu) shared.sd_model.first_stage_model.to(devices.cpu) size = len(ds.indexes) - loss_dict = defaultdict(lambda: deque(maxlen=1024)) + loss_dict = defaultdict(lambda : deque(maxlen = 1024)) losses = torch.zeros((size,)) previous_mean_losses = [0] previous_mean_loss = 0 print("Mean loss of {} elements".format(size)) - + weights = hypernetwork.weights() for weight in weights: weight.requires_grad = True @@ -425,7 +407,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if len(loss_dict) > 0: previous_mean_losses = [i[-1] for i in loss_dict.values()] previous_mean_loss = mean(previous_mean_losses) - + scheduler.apply(optimizer, hypernetwork.step) if scheduler.finished: break @@ -444,7 +426,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log losses[hypernetwork.step % losses.shape[0]] = loss.item() for entry in entries: loss_dict[entry.filename].append(loss.item()) - + optimizer.zero_grad() weights[0].grad = None loss.backward() @@ -459,9 +441,9 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log steps_done = hypernetwork.step + 1 - if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): + if torch.isnan(losses[hypernetwork.step % losses.shape[0]]): raise RuntimeError("Loss diverged.") - + if len(previous_mean_losses) > 1: std = stdev(previous_mean_losses) else: @@ -510,7 +492,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log preview_text = p.prompt processed = processing.process_images(p) - image = processed.images[0] if len(processed.images) > 0 else None + image = processed.images[0] if len(processed.images)>0 else None if unload: shared.sd_model.cond_stage_model.to(devices.cpu) @@ -518,10 +500,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log if image is not None: shared.state.current_image = image - last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, - shared.opts.samples_format, processed.infotexts[0], - p=p, forced_filename=forced_filename, - save_to_dirs=False) + last_saved_image, last_text_info = images.save_image(image, images_dir, "", p.seed, p.prompt, shared.opts.samples_format, processed.infotexts[0], p=p, forced_filename=forced_filename, save_to_dirs=False) last_saved_image += f", prompt: {preview_text}" shared.state.job_no = hypernetwork.step @@ -535,7 +514,7 @@ Last saved hypernetwork: {html.escape(last_saved_file)}
Last saved image: {html.escape(last_saved_image)}

""" - + report_statistics(loss_dict) filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') @@ -543,7 +522,6 @@ Last saved image: {html.escape(last_saved_image)}
return hypernetwork, filename - def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename): old_hypernetwork_name = hypernetwork.name old_sd_checkpoint = hypernetwork.sd_checkpoint if hasattr(hypernetwork, "sd_checkpoint") else None @@ -557,4 +535,4 @@ def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename): hypernetwork.sd_checkpoint = old_sd_checkpoint hypernetwork.sd_checkpoint_name = old_sd_checkpoint_name hypernetwork.name = old_hypernetwork_name - raise \ No newline at end of file + raise From 283249d2390f0f3a1c8a55d5d9aa551e3e9b2f9c Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 15:57:17 +0900 Subject: [PATCH 11/24] apply --- modules/hypernetworks/hypernetwork.py | 54 ++++++++++++++++++++++++--- 1 file changed, 49 insertions(+), 5 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 6e1a10cf..de8688a9 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -22,6 +22,8 @@ from collections import defaultdict, deque from statistics import stdev, mean +optimizer_dict = {optim_name : cls_obj for optim_name, cls_obj in inspect.getmembers(torch.optim, inspect.isclass) if optim_name != "Optimizer"} + class HypernetworkModule(torch.nn.Module): multiplier = 1.0 activation_dict = { @@ -142,6 +144,8 @@ class Hypernetwork: self.use_dropout = use_dropout self.activate_output = activate_output self.last_layer_dropout = kwargs['last_layer_dropout'] if 'last_layer_dropout' in kwargs else True + self.optimizer_name = None + self.optimizer_state_dict = None for size in enable_sizes or []: self.layers[size] = ( @@ -163,6 +167,7 @@ class Hypernetwork: def save(self, filename): state_dict = {} + optimizer_saved_dict = {} for k, v in self.layers.items(): state_dict[k] = (v[0].state_dict(), v[1].state_dict()) @@ -178,8 +183,15 @@ class Hypernetwork: state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name state_dict['activate_output'] = self.activate_output state_dict['last_layer_dropout'] = self.last_layer_dropout - + + if self.optimizer_name is not None: + optimizer_saved_dict['optimizer_name'] = self.optimizer_name + torch.save(state_dict, filename) + if self.optimizer_state_dict: + optimizer_saved_dict['hash'] = sd_models.model_hash(filename) + optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict + torch.save(optimizer_saved_dict, filename + '.optim') def load(self, filename): self.filename = filename @@ -202,6 +214,18 @@ class Hypernetwork: print(f"Activate last layer is set to {self.activate_output}") self.last_layer_dropout = state_dict.get('last_layer_dropout', False) + optimizer_saved_dict = torch.load(self.filename + '.optim', map_location = 'cpu') if os.path.exists(self.filename + '.optim') else {} + self.optimizer_name = optimizer_saved_dict.get('optimizer_name', 'AdamW') + print(f"Optimizer name is {self.optimizer_name}") + if sd_models.model_hash(filename) == optimizer_saved_dict.get('hash', None): + self.optimizer_state_dict = optimizer_saved_dict.get('optimizer_state_dict', None) + else: + self.optimizer_state_dict = None + if self.optimizer_state_dict: + print("Loaded existing optimizer from checkpoint") + else: + print("No saved optimizer exists in checkpoint") + for size, sd in state_dict.items(): if type(size) == int: self.layers[size] = ( @@ -223,7 +247,7 @@ def list_hypernetworks(path): name = os.path.splitext(os.path.basename(filename))[0] # Prevent a hypothetical "None.pt" from being listed. if name != "None": - res[name] = filename + res[name + f"({sd_models.model_hash(filename)})"] = filename return res @@ -369,6 +393,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log else: hypernetwork_dir = None + hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0] if create_image_every > 0: images_dir = os.path.join(log_directory, "images") os.makedirs(images_dir, exist_ok=True) @@ -404,8 +429,19 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log weights = hypernetwork.weights() for weight in weights: weight.requires_grad = True - # if optimizer == "AdamW": or else Adam / AdamW / SGD, etc... - optimizer = torch.optim.AdamW(weights, lr=scheduler.learn_rate) + # Here we use optimizer from saved HN, or we can specify as UI option. + if (optimizer_name := hypernetwork.optimizer_name) in optimizer_dict: + optimizer = optimizer_dict[hypernetwork.optimizer_name](params=weights, lr=scheduler.learn_rate) + else: + print(f"Optimizer type {optimizer_name} is not defined!") + optimizer = torch.optim.AdamW(params=weights, lr=scheduler.learn_rate) + optimizer_name = 'AdamW' + if hypernetwork.optimizer_state_dict: # This line must be changed if Optimizer type can be different from saved optimizer. + try: + optimizer.load_state_dict(hypernetwork.optimizer_state_dict) + except RuntimeError as e: + print("Cannot resume from saved optimizer!") + print(e) steps_without_grad = 0 @@ -467,7 +503,11 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log # Before saving, change name to match current checkpoint. hypernetwork_name_every = f'{hypernetwork_name}-{steps_done}' last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name_every}.pt') + hypernetwork.optimizer_name = optimizer_name + if shared.opts.save_optimizer_state: + hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, last_saved_file) + hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. textual_inversion.write_loss(log_directory, "hypernetwork_loss.csv", hypernetwork.step, len(ds), { "loss": f"{previous_mean_loss:.7f}", @@ -530,8 +570,12 @@ Last saved image: {html.escape(last_saved_image)}
report_statistics(loss_dict) filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') + hypernetwork.optimizer_name = optimizer_name + if shared.opts.save_optimizer_state: + hypernetwork.optimizer_state_dict = optimizer.state_dict() save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename) - + del optimizer + hypernetwork.optimizer_state_dict = None # dereference it after saving, to save memory. return hypernetwork, filename def save_hypernetwork(hypernetwork, checkpoint, hypernetwork_name, filename): From f5d394214d6ee74a682d0a1016bcbebc4b43c13a Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 16:04:03 +0900 Subject: [PATCH 12/24] split before declaring file name --- modules/hypernetworks/hypernetwork.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index de8688a9..9b6a3e62 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -382,6 +382,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log shared.state.textinfo = "Initializing hypernetwork training..." shared.state.job_count = steps + hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0] filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt') log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name) @@ -393,7 +394,6 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, data_root, log else: hypernetwork_dir = None - hypernetwork_name = hypernetwork_name.rsplit('(', 1)[0] if create_image_every > 0: images_dir = os.path.join(log_directory, "images") os.makedirs(images_dir, exist_ok=True) From 1ca0bcd3a7003dd2c1324de7d97fd2a6fc5ddc53 Mon Sep 17 00:00:00 2001 From: aria1th <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 16:09:19 +0900 Subject: [PATCH 13/24] only save if option is enabled --- modules/hypernetworks/hypernetwork.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 9b6a3e62..b1f308e2 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -188,7 +188,7 @@ class Hypernetwork: optimizer_saved_dict['optimizer_name'] = self.optimizer_name torch.save(state_dict, filename) - if self.optimizer_state_dict: + if shared.opts.save_optimizer_state and self.optimizer_state_dict: optimizer_saved_dict['hash'] = sd_models.model_hash(filename) optimizer_saved_dict['optimizer_state_dict'] = self.optimizer_state_dict torch.save(optimizer_saved_dict, filename + '.optim') From 7278897982bfb640ee95f144c97ed25fb3f77ea3 Mon Sep 17 00:00:00 2001 From: AngelBottomless <35677394+aria1th@users.noreply.github.com> Date: Fri, 4 Nov 2022 17:12:28 +0900 Subject: [PATCH 14/24] Update shared.py --- modules/shared.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/shared.py b/modules/shared.py index 4d6e1c8b..6e7a02e0 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -309,7 +309,7 @@ options_templates.update(options_section(('system', "System"), { options_templates.update(options_section(('training', "Training"), { "unload_models_when_training": OptionInfo(False, "Move VAE and CLIP to RAM when training if possible. Saves VRAM."), - "save_optimizer_state": OptionInfo(False, "Saves Optimizer state with checkpoints. This will cause file size to increase VERY much."), + "save_optimizer_state": OptionInfo(False, "Saves Optimizer state as separate *.optim file. Training can be resumed with HN itself and matching optim file."), "dataset_filename_word_regex": OptionInfo("", "Filename word regex"), "dataset_filename_join_string": OptionInfo(" ", "Filename join string"), "training_image_repeats_per_epoch": OptionInfo(1, "Number of repeats for a single input image per epoch; used only for displaying epoch number", gr.Number, {"precision": 0}), From fd62727893f9face287b0a9620251afaa38a627d Mon Sep 17 00:00:00 2001 From: Isaac Poulton Date: Fri, 4 Nov 2022 18:34:35 +0700 Subject: [PATCH 15/24] Sort hypernetworks --- modules/hypernetworks/hypernetwork.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index 6e1a10cf..f1f04a70 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -224,7 +224,7 @@ def list_hypernetworks(path): # Prevent a hypothetical "None.pt" from being listed. if name != "None": res[name] = filename - return res + return dict(sorted(res.items())) def load_hypernetwork(filename): From f316280ad3634a2343b086a6de0bfcd473e18599 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 4 Nov 2022 16:48:40 +0300 Subject: [PATCH 16/24] fix the error that prevents from setting some options --- modules/shared.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/modules/shared.py b/modules/shared.py index a9e28b9c..962115f6 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -406,7 +406,8 @@ class Options: if key in self.data or key in self.data_labels: assert not cmd_opts.freeze_settings, "changing settings is disabled" - comp_args = opts.data_labels[key].component_args + info = opts.data_labels.get(key, None) + comp_args = info.component_args if info else None if isinstance(comp_args, dict) and comp_args.get('visible', True) is False: raise RuntimeError(f"not possible to set {key} because it is restricted") From 116bcf730ade8d3ac5d76d04c5887b6bba000970 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 4 Nov 2022 16:48:46 +0300 Subject: [PATCH 17/24] disable setting options via API until it is fixed by the author --- modules/api/api.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/modules/api/api.py b/modules/api/api.py index a49f3755..8a7ab2f5 100644 --- a/modules/api/api.py +++ b/modules/api/api.py @@ -218,6 +218,10 @@ class Api: return options def set_config(self, req: OptionsModel): + # currently req has all options fields even if you send a dict like { "send_seed": false }, which means it will + # overwrite all options with default values. + raise RuntimeError('Setting options via API is not supported') + reqDict = vars(req) for o in reqDict: setattr(shared.opts, o, reqDict[o]) From 08feb4c364e8b2aed929fd7d22dfa21a93d78b2c Mon Sep 17 00:00:00 2001 From: Isaac Poulton Date: Fri, 4 Nov 2022 20:53:11 +0700 Subject: [PATCH 18/24] Sort straight out of the glob --- modules/hypernetworks/hypernetwork.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/hypernetworks/hypernetwork.py b/modules/hypernetworks/hypernetwork.py index f1f04a70..a441ab10 100644 --- a/modules/hypernetworks/hypernetwork.py +++ b/modules/hypernetworks/hypernetwork.py @@ -219,12 +219,12 @@ class Hypernetwork: def list_hypernetworks(path): res = {} - for filename in glob.iglob(os.path.join(path, '**/*.pt'), recursive=True): + for filename in sorted(glob.iglob(os.path.join(path, '**/*.pt'), recursive=True)): name = os.path.splitext(os.path.basename(filename))[0] # Prevent a hypothetical "None.pt" from being listed. if name != "None": res[name] = filename - return dict(sorted(res.items())) + return res def load_hypernetwork(filename): From 0d7e01d9950e013784c4b77c05aa7583ea69edc8 Mon Sep 17 00:00:00 2001 From: innovaciones Date: Fri, 4 Nov 2022 12:14:32 -0600 Subject: [PATCH 19/24] Open extensions links in new tab Fixed for "Available" tab --- modules/ui_extensions.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/modules/ui_extensions.py b/modules/ui_extensions.py index a81de9a7..8e0d41d5 100644 --- a/modules/ui_extensions.py +++ b/modules/ui_extensions.py @@ -188,7 +188,7 @@ def refresh_available_extensions_from_data(): code += f""" - {html.escape(name)} + {html.escape(name)} {html.escape(description)} {install_code} From 30b1bcc64e67ad50c5d3af3a6fe1bd1e9553f34e Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Fri, 4 Nov 2022 22:56:18 +0300 Subject: [PATCH 20/24] fix upscale loop erroneously applied multiple times --- modules/upscaler.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/modules/upscaler.py b/modules/upscaler.py index 83fde7ca..c4e6e6bd 100644 --- a/modules/upscaler.py +++ b/modules/upscaler.py @@ -57,10 +57,18 @@ class Upscaler: 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 + shape = (img.width, img.height) + img = self.do_upscale(img, selected_model) + + if shape == (img.width, img.height): + break + + if img.width >= dest_w and img.height >= dest_h: + break + if img.width != dest_w or img.height != dest_h: img = img.resize((int(dest_w), int(dest_h)), resample=LANCZOS) From ebce0c57c78a3f22178e3a38938d19ec0dfb703d Mon Sep 17 00:00:00 2001 From: Billy Cao Date: Sat, 5 Nov 2022 11:38:24 +0800 Subject: [PATCH 21/24] Use typing.Optional instead of | to add support for Python 3.9 and below. --- modules/api/models.py | 26 +++++++++++++------------- 1 file changed, 13 insertions(+), 13 deletions(-) diff --git a/modules/api/models.py b/modules/api/models.py index 2ae75f43..a44c5ddd 100644 --- a/modules/api/models.py +++ b/modules/api/models.py @@ -1,6 +1,6 @@ import inspect from pydantic import BaseModel, Field, create_model -from typing import Any, Optional, Union +from typing import Any, Optional from typing_extensions import Literal from inflection import underscore from modules.processing import StableDiffusionProcessingTxt2Img, StableDiffusionProcessingImg2Img @@ -185,22 +185,22 @@ _options = vars(parser)['_option_string_actions'] for key in _options: if(_options[key].dest != 'help'): flag = _options[key] - _type = str - if(_options[key].default != None): _type = type(_options[key].default) + _type = str + if _options[key].default is not None: _type = type(_options[key].default) flags.update({flag.dest: (_type,Field(default=flag.default, description=flag.help))}) FlagsModel = create_model("Flags", **flags) class SamplerItem(BaseModel): name: str = Field(title="Name") - aliases: list[str] = Field(title="Aliases") + aliases: list[str] = Field(title="Aliases") options: dict[str, str] = Field(title="Options") class UpscalerItem(BaseModel): name: str = Field(title="Name") - model_name: str | None = Field(title="Model Name") - model_path: str | None = Field(title="Path") - model_url: str | None = Field(title="URL") + model_name: Optional[str] = Field(title="Model Name") + model_path: Optional[str] = Field(title="Path") + model_url: Optional[str] = Field(title="URL") class SDModelItem(BaseModel): title: str = Field(title="Title") @@ -211,21 +211,21 @@ class SDModelItem(BaseModel): class HypernetworkItem(BaseModel): name: str = Field(title="Name") - path: str | None = Field(title="Path") + path: Optional[str] = Field(title="Path") class FaceRestorerItem(BaseModel): name: str = Field(title="Name") - cmd_dir: str | None = Field(title="Path") + cmd_dir: Optional[str] = Field(title="Path") class RealesrganItem(BaseModel): name: str = Field(title="Name") - path: str | None = Field(title="Path") - scale: int | None = Field(title="Scale") + path: Optional[str] = Field(title="Path") + scale: Optional[int] = Field(title="Scale") class PromptStyleItem(BaseModel): name: str = Field(title="Name") - prompt: str | None = Field(title="Prompt") - negative_prompt: str | None = Field(title="Negative Prompt") + prompt: Optional[str] = Field(title="Prompt") + negative_prompt: Optional[str] = Field(title="Negative Prompt") class ArtistItem(BaseModel): name: str = Field(title="Name") From e9a5562b9b27a1a4f9c282637b111cefd9727a41 Mon Sep 17 00:00:00 2001 From: papuSpartan Date: Sat, 5 Nov 2022 04:06:51 -0500 Subject: [PATCH 22/24] add support for tls (gradio tls options) --- modules/shared.py | 3 +++ webui.py | 22 ++++++++++++++++++++-- 2 files changed, 23 insertions(+), 2 deletions(-) diff --git a/modules/shared.py b/modules/shared.py index 962115f6..7a20c3af 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -86,6 +86,9 @@ parser.add_argument("--nowebui", action='store_true', help="use api=True to laun parser.add_argument("--ui-debug-mode", action='store_true', help="Don't load model to quickly launch UI") parser.add_argument("--device-id", type=str, help="Select the default CUDA device to use (export CUDA_VISIBLE_DEVICES=0,1,etc might be needed before)", default=None) parser.add_argument("--administrator", action='store_true', help="Administrator rights", default=False) +parser.add_argument("--tls-keyfile", type=str, help="Partially enables TLS, requires --tls-certfile to fully function", default=None) +parser.add_argument("--tls-certfile", type=str, help="Partially enables TLS, requires --tls-keyfile to fully function", default=None) +parser.add_argument("--server-name", type=str, help="Sets hostname of server", default=None) cmd_opts = parser.parse_args() restricted_opts = { diff --git a/webui.py b/webui.py index 81df09dd..d366f4ca 100644 --- a/webui.py +++ b/webui.py @@ -34,7 +34,7 @@ from modules.shared import cmd_opts import modules.hypernetworks.hypernetwork queue_lock = threading.Lock() - +server_name = "0.0.0.0" if cmd_opts.listen else cmd_opts.server_name def wrap_queued_call(func): def f(*args, **kwargs): @@ -85,6 +85,22 @@ def initialize(): shared.opts.onchange("sd_hypernetwork", wrap_queued_call(lambda: modules.hypernetworks.hypernetwork.load_hypernetwork(shared.opts.sd_hypernetwork))) shared.opts.onchange("sd_hypernetwork_strength", modules.hypernetworks.hypernetwork.apply_strength) + if cmd_opts.tls_keyfile is not None and cmd_opts.tls_keyfile is not None: + + try: + if not os.path.exists(cmd_opts.tls_keyfile): + print("Invalid path to TLS keyfile given") + if not os.path.exists(cmd_opts.tls_certfile): + print(f"Invalid path to TLS certfile: '{cmd_opts.tls_certfile}'") + except TypeError: + cmd_opts.tls_keyfile = cmd_opts.tls_certfile = None + print(f"path: '{cmd_opts.tls_keyfile}' {type(cmd_opts.tls_keyfile)}") + print(f"path: '{cmd_opts.tls_certfile}' {type(cmd_opts.tls_certfile)}") + print("TLS setup invalid, running webui without TLS") + else: + print("Running with TLS") + + # make the program just exit at ctrl+c without waiting for anything def sigint_handler(sig, frame): print(f'Interrupted with signal {sig} in {frame}') @@ -131,8 +147,10 @@ def webui(): app, local_url, share_url = demo.launch( share=cmd_opts.share, - server_name="0.0.0.0" if cmd_opts.listen else None, + server_name=server_name, server_port=cmd_opts.port, + ssl_keyfile=cmd_opts.tls_keyfile, + ssl_certfile=cmd_opts.tls_certfile, debug=cmd_opts.gradio_debug, auth=[tuple(cred.split(':')) for cred in cmd_opts.gradio_auth.strip('"').split(',')] if cmd_opts.gradio_auth else None, inbrowser=cmd_opts.autolaunch, From a02bad570ef7718436369bb4e4aa5b8e0f1f5689 Mon Sep 17 00:00:00 2001 From: papuSpartan Date: Sat, 5 Nov 2022 04:14:21 -0500 Subject: [PATCH 23/24] rm dbg --- webui.py | 2 -- 1 file changed, 2 deletions(-) diff --git a/webui.py b/webui.py index d366f4ca..222dbeee 100644 --- a/webui.py +++ b/webui.py @@ -94,8 +94,6 @@ def initialize(): print(f"Invalid path to TLS certfile: '{cmd_opts.tls_certfile}'") except TypeError: cmd_opts.tls_keyfile = cmd_opts.tls_certfile = None - print(f"path: '{cmd_opts.tls_keyfile}' {type(cmd_opts.tls_keyfile)}") - print(f"path: '{cmd_opts.tls_certfile}' {type(cmd_opts.tls_certfile)}") print("TLS setup invalid, running webui without TLS") else: print("Running with TLS") From 03b08c4a6b0609f24ec789d40100529b92ef0612 Mon Sep 17 00:00:00 2001 From: AUTOMATIC <16777216c@gmail.com> Date: Sat, 5 Nov 2022 15:04:48 +0300 Subject: [PATCH 24/24] do not die when an extension's repo has no remote --- modules/extensions.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/modules/extensions.py b/modules/extensions.py index 897af96e..8e0977fd 100644 --- a/modules/extensions.py +++ b/modules/extensions.py @@ -34,8 +34,11 @@ class Extension: if repo is None or repo.bare: self.remote = None else: - self.remote = next(repo.remote().urls, None) - self.status = 'unknown' + try: + self.remote = next(repo.remote().urls, None) + self.status = 'unknown' + except Exception: + self.remote = None def list_files(self, subdir, extension): from modules import scripts