WebUI/modules/codeformer_model.py

141 lines
5.8 KiB
Python
Raw Normal View History

2022-09-07 09:32:28 +00:00
import os
import sys
import traceback
import cv2
2022-09-07 09:32:28 +00:00
import torch
2022-09-26 15:27:18 +00:00
import modules.face_restoration
import modules.shared
2022-09-26 14:29:50 +00:00
from modules import shared, devices, modelloader
from modules.paths import script_path, models_path
2022-09-07 09:32:28 +00:00
2022-09-26 15:27:18 +00:00
# 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.
2022-09-07 09:32:28 +00:00
# I am making a choice to include some files from codeformer to work around this issue.
2022-09-26 14:29:50 +00:00
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'
2022-09-07 09:32:28 +00:00
have_codeformer = False
2022-09-07 10:35:02 +00:00
codeformer = None
2022-09-07 09:32:28 +00:00
2022-09-26 14:29:50 +00:00
def setup_model(dirname):
global model_path
if not os.path.exists(model_path):
os.makedirs(model_path)
2022-09-07 09:32:28 +00:00
path = modules.paths.paths.get("CodeFormer", None)
if path is None:
return
try:
from torchvision.transforms.functional import normalize
from modules.codeformer.codeformer_arch import CodeFormer
from basicsr.utils.download_util import load_file_from_url
from basicsr.utils import imwrite, img2tensor, tensor2img
from facelib.utils.face_restoration_helper import FaceRestoreHelper
from modules.shared import cmd_opts
net_class = CodeFormer
class FaceRestorerCodeFormer(modules.face_restoration.FaceRestoration):
def name(self):
return "CodeFormer"
2022-09-26 14:29:50 +00:00
def __init__(self, dirname):
2022-09-07 09:32:28 +00:00
self.net = None
self.face_helper = None
2022-09-26 14:29:50 +00:00
self.cmd_dir = dirname
2022-09-07 09:32:28 +00:00
def create_models(self):
if self.net is not None and self.face_helper is not None:
self.net.to(devices.device_codeformer)
2022-09-07 09:32:28 +00:00
return self.net, self.face_helper
2022-09-30 08:03:00 +00:00
model_paths = modelloader.load_models(model_path, model_url, self.cmd_dir, download_name='codeformer-v0.1.0.pth')
2022-09-26 14:29:50 +00:00
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)
2022-09-07 09:32:28 +00:00
checkpoint = torch.load(ckpt_path)['params_ema']
net.load_state_dict(checkpoint)
net.eval()
face_helper = FaceRestoreHelper(1, face_size=512, crop_ratio=(1, 1), det_model='retinaface_resnet50', save_ext='png', use_parse=True, device=devices.device_codeformer)
2022-09-07 09:32:28 +00:00
self.net = net
self.face_helper = face_helper
2022-09-07 09:32:28 +00:00
return net, face_helper
def send_model_to(self, device):
self.net.to(device)
self.face_helper.face_det.to(device)
self.face_helper.face_parse.to(device)
2022-09-07 10:35:02 +00:00
def restore(self, np_image, w=None):
2022-09-07 09:32:28 +00:00
np_image = np_image[:, :, ::-1]
original_resolution = np_image.shape[0:2]
self.create_models()
2022-09-26 14:29:50 +00:00
if self.net is None or self.face_helper is None:
return np_image
self.send_model_to(devices.device_codeformer)
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)
self.face_helper.align_warp_face()
2022-09-07 09:32:28 +00:00
for idx, cropped_face in enumerate(self.face_helper.cropped_faces):
2022-09-07 09:32:28 +00:00
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(devices.device_codeformer)
2022-09-07 09:32:28 +00:00
try:
with torch.no_grad():
output = self.net(cropped_face_t, w=w if w is not None else shared.opts.code_former_weight, adain=True)[0]
2022-09-07 09:32:28 +00:00
restored_face = tensor2img(output, rgb2bgr=True, min_max=(-1, 1))
del output
torch.cuda.empty_cache()
except Exception as error:
print(f'\tFailed inference for CodeFormer: {error}', file=sys.stderr)
restored_face = tensor2img(cropped_face_t, rgb2bgr=True, min_max=(-1, 1))
restored_face = restored_face.astype('uint8')
self.face_helper.add_restored_face(restored_face)
2022-09-07 09:32:28 +00:00
self.face_helper.get_inverse_affine(None)
2022-09-07 09:32:28 +00:00
restored_img = self.face_helper.paste_faces_to_input_image()
2022-09-07 09:32:28 +00:00
restored_img = restored_img[:, :, ::-1]
if original_resolution != restored_img.shape[0:2]:
restored_img = cv2.resize(restored_img, (0, 0), fx=original_resolution[1]/restored_img.shape[1], fy=original_resolution[0]/restored_img.shape[0], interpolation=cv2.INTER_LINEAR)
self.face_helper.clean_all()
if shared.opts.face_restoration_unload:
self.send_model_to(devices.cpu)
2022-09-07 09:32:28 +00:00
return restored_img
global have_codeformer
have_codeformer = True
2022-09-07 10:35:02 +00:00
global codeformer
2022-09-26 14:29:50 +00:00
codeformer = FaceRestorerCodeFormer(dirname)
2022-09-07 10:35:02 +00:00
shared.face_restorers.append(codeformer)
2022-09-07 09:32:28 +00:00
except Exception:
print("Error setting up CodeFormer:", file=sys.stderr)
print(traceback.format_exc(), file=sys.stderr)
# sys.path = stored_sys_path