Upload 203 files

.gitattributes

@@ -187,3 +187,4 @@ SimSwap/temp_results/frame_0000058.jpg filter=lfs diff=lfs merge=lfs -text
 SimSwap/temp_results/frame_0000059.jpg filter=lfs diff=lfs merge=lfs -text
 SimSwap/temp_results/frame_0000060.jpg filter=lfs diff=lfs merge=lfs -text
 SimSwap/temp_results/frame_0000061.jpg filter=lfs diff=lfs merge=lfs -text
+video-tetalking/third_part/GPEN/face_parse/test.png filter=lfs diff=lfs merge=lfs -text

video-tetalking/CODE_OF_CONDUCT.md

@@ -0,0 +1,43 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as contributors and maintainers pledge to make participation in our project and our community a harassment-free experience for everyone, regardless of age, body size, disability, ethnicity, gender identity and expression, level of experience, education, socio-economic status, nationality, personal appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment include:
+
+- Using welcoming and inclusive language
+- Being respectful of differing viewpoints and experiences
+- Gracefully accepting constructive criticism
+- Focusing on what is best for the community
+- Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+- The use of sexualized language or imagery and unwelcome sexual attention or advances
+- Trolling, insulting/derogatory comments, and personal or political attacks
+- Public or private harassment
+- Publishing others' private information, such as a physical or electronic address, without explicit permission
+- Other conduct that could reasonably be considered inappropriate in a professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable behavior and are expected to take appropriate and fair corrective action in response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or reject comments, commits, code, wiki edits, issues, and other contributions that are not aligned with this Code of Conduct, or to ban temporarily or permanently any contributor for other behaviors that they deem inappropriate, threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies both within project spaces and in public spaces when an individual is representing the project or its community. Examples of representing a project or community include using an official project email address, posting via an official social media account, or acting as an appointed representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be reported by contacting the project team  All complaints will be reviewed and investigated and will result in a response that is deemed necessary and appropriate to the circumstances. The project team is obligated to maintain confidentiality with regard to the reporter of an incident.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good faith may face temporary or permanent repercussions as determined by other members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the Contributor Covenant, version 2.0, available at [https://www.contributor-covenant.org/version/2/0/code_of_conduct.html](https://www.contributor-covenant.org/version/2/0/code_of_conduct.html).

video-tetalking/inference.py

@@ -0,0 +1,345 @@
+import numpy as np
+import cv2, os, sys, subprocess, platform, torch
+from tqdm import tqdm
+from PIL import Image
+from scipy.io import loadmat
+
+sys.path.insert(0, 'third_part')
+sys.path.insert(0, 'third_part/GPEN')
+sys.path.insert(0, 'third_part/GFPGAN')
+
+# 3dmm extraction
+from third_part.face3d.util.preprocess import align_img
+from third_part.face3d.util.load_mats import load_lm3d
+from third_part.face3d.extract_kp_videos import KeypointExtractor
+# face enhancement
+from third_part.GPEN.gpen_face_enhancer import FaceEnhancement
+from third_part.GFPGAN.gfpgan import GFPGANer
+# expression control
+from third_part.ganimation_replicate.model.ganimation import GANimationModel
+
+from utils import audio
+from utils.ffhq_preprocess import Croper
+from utils.alignment_stit import crop_faces, calc_alignment_coefficients, paste_image
+from utils.inference_utils import Laplacian_Pyramid_Blending_with_mask, face_detect, load_model, options, split_coeff, \
+                                  trans_image, transform_semantic, find_crop_norm_ratio, load_face3d_net, exp_aus_dict
+import warnings
+warnings.filterwarnings("ignore")
+
+args = options()
+
+def main():    
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    print('[Info] Using {} for inference.'.format(device))
+    os.makedirs(os.path.join('temp', args.tmp_dir), exist_ok=True)
+
+    enhancer = FaceEnhancement(base_dir='checkpoints', size=512, model='GPEN-BFR-512', use_sr=False, \
+                               sr_model='rrdb_realesrnet_psnr', channel_multiplier=2, narrow=1, device=device)
+    restorer = GFPGANer(model_path='checkpoints/GFPGANv1.3.pth', upscale=1, arch='clean', \
+                        channel_multiplier=2, bg_upsampler=None)
+
+    base_name = args.face.split('/')[-1]
+    if os.path.isfile(args.face) and args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+        args.static = True
+    if not os.path.isfile(args.face):
+        raise ValueError('--face argument must be a valid path to video/image file')
+    elif args.face.split('.')[1] in ['jpg', 'png', 'jpeg']:
+        full_frames = [cv2.imread(args.face)]
+        fps = args.fps
+    else:
+        video_stream = cv2.VideoCapture(args.face)
+        fps = video_stream.get(cv2.CAP_PROP_FPS)
+
+        full_frames = []
+        while True:
+            still_reading, frame = video_stream.read()
+            if not still_reading:
+                video_stream.release()
+                break
+            y1, y2, x1, x2 = args.crop
+            if x2 == -1: x2 = frame.shape[1]
+            if y2 == -1: y2 = frame.shape[0]
+            frame = frame[y1:y2, x1:x2]
+            full_frames.append(frame)
+
+    print ("[Step 0] Number of frames available for inference: "+str(len(full_frames)))
+    # face detection & cropping, cropping the first frame as the style of FFHQ
+    croper = Croper('checkpoints/shape_predictor_68_face_landmarks.dat')
+    full_frames_RGB = [cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) for frame in full_frames]
+    full_frames_RGB, crop, quad = croper.crop(full_frames_RGB, xsize=512)
+
+    clx, cly, crx, cry = crop
+    lx, ly, rx, ry = quad
+    lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry)
+    oy1, oy2, ox1, ox2 = cly+ly, min(cly+ry, full_frames[0].shape[0]), clx+lx, min(clx+rx, full_frames[0].shape[1])
+    # original_size = (ox2 - ox1, oy2 - oy1)
+    frames_pil = [Image.fromarray(cv2.resize(frame,(256,256))) for frame in full_frames_RGB]
+
+    # get the landmark according to the detected face.
+    if not os.path.isfile('temp/'+base_name+'_landmarks.txt') or args.re_preprocess:
+        print('[Step 1] Landmarks Extraction in Video.')
+        kp_extractor = KeypointExtractor()
+        lm = kp_extractor.extract_keypoint(frames_pil, './temp/'+base_name+'_landmarks.txt')
+    else:
+        print('[Step 1] Using saved landmarks.')
+        lm = np.loadtxt('temp/'+base_name+'_landmarks.txt').astype(np.float32)
+        lm = lm.reshape([len(full_frames), -1, 2])
+       
+    if not os.path.isfile('temp/'+base_name+'_coeffs.npy') or args.exp_img is not None or args.re_preprocess:
+        net_recon = load_face3d_net(args.face3d_net_path, device)
+        lm3d_std = load_lm3d('checkpoints/BFM')
+
+        video_coeffs = []
+        for idx in tqdm(range(len(frames_pil)), desc="[Step 2] 3DMM Extraction In Video:"):
+            frame = frames_pil[idx]
+            W, H = frame.size
+            lm_idx = lm[idx].reshape([-1, 2])
+            if np.mean(lm_idx) == -1:
+                lm_idx = (lm3d_std[:, :2]+1) / 2.
+                lm_idx = np.concatenate([lm_idx[:, :1] * W, lm_idx[:, 1:2] * H], 1)
+            else:
+                lm_idx[:, -1] = H - 1 - lm_idx[:, -1]
+
+            trans_params, im_idx, lm_idx, _ = align_img(frame, lm_idx, lm3d_std)
+            trans_params = np.array([float(item) for item in np.hsplit(trans_params, 5)]).astype(np.float32)
+            im_idx_tensor = torch.tensor(np.array(im_idx)/255., dtype=torch.float32).permute(2, 0, 1).to(device).unsqueeze(0) 
+            with torch.no_grad():
+                coeffs = split_coeff(net_recon(im_idx_tensor))
+
+            pred_coeff = {key:coeffs[key].cpu().numpy() for key in coeffs}
+            pred_coeff = np.concatenate([pred_coeff['id'], pred_coeff['exp'], pred_coeff['tex'], pred_coeff['angle'],\
+                                         pred_coeff['gamma'], pred_coeff['trans'], trans_params[None]], 1)
+            video_coeffs.append(pred_coeff)
+        semantic_npy = np.array(video_coeffs)[:,0]
+        np.save('temp/'+base_name+'_coeffs.npy', semantic_npy)
+    else:
+        print('[Step 2] Using saved coeffs.')
+        semantic_npy = np.load('temp/'+base_name+'_coeffs.npy').astype(np.float32)
+
+    # generate the 3dmm coeff from a single image
+    if args.exp_img is not None and ('.png' in args.exp_img or '.jpg' in args.exp_img):
+        print('extract the exp from',args.exp_img)
+        exp_pil = Image.open(args.exp_img).convert('RGB')
+        lm3d_std = load_lm3d('third_part/face3d/BFM')
+        
+        W, H = exp_pil.size
+        kp_extractor = KeypointExtractor()
+        lm_exp = kp_extractor.extract_keypoint([exp_pil], 'temp/'+base_name+'_temp.txt')[0]
+        if np.mean(lm_exp) == -1:
+            lm_exp = (lm3d_std[:, :2] + 1) / 2.
+            lm_exp = np.concatenate(
+                [lm_exp[:, :1] * W, lm_exp[:, 1:2] * H], 1)
+        else:
+            lm_exp[:, -1] = H - 1 - lm_exp[:, -1]
+
+        trans_params, im_exp, lm_exp, _ = align_img(exp_pil, lm_exp, lm3d_std)
+        trans_params = np.array([float(item) for item in np.hsplit(trans_params, 5)]).astype(np.float32)
+        im_exp_tensor = torch.tensor(np.array(im_exp)/255., dtype=torch.float32).permute(2, 0, 1).to(device).unsqueeze(0)
+        with torch.no_grad():
+            expression = split_coeff(net_recon(im_exp_tensor))['exp'][0]
+        del net_recon
+    elif args.exp_img == 'smile':
+        expression = torch.tensor(loadmat('checkpoints/expression.mat')['expression_mouth'])[0]
+    else:
+        print('using expression center')
+        expression = torch.tensor(loadmat('checkpoints/expression.mat')['expression_center'])[0]
+
+    # load DNet, model(LNet and ENet)
+    D_Net, model = load_model(args, device)
+
+    if not os.path.isfile('temp/'+base_name+'_stablized.npy') or args.re_preprocess:
+        imgs = []
+        for idx in tqdm(range(len(frames_pil)), desc="[Step 3] Stabilize the expression In Video:"):
+            if args.one_shot:
+                source_img = trans_image(frames_pil[0]).unsqueeze(0).to(device)
+                semantic_source_numpy = semantic_npy[0:1]
+            else:
+                source_img = trans_image(frames_pil[idx]).unsqueeze(0).to(device)
+                semantic_source_numpy = semantic_npy[idx:idx+1]
+            ratio = find_crop_norm_ratio(semantic_source_numpy, semantic_npy)
+            coeff = transform_semantic(semantic_npy, idx, ratio).unsqueeze(0).to(device)
+        
+            # hacking the new expression
+            coeff[:, :64, :] = expression[None, :64, None].to(device) 
+            with torch.no_grad():
+                output = D_Net(source_img, coeff)
+            img_stablized = np.uint8((output['fake_image'].squeeze(0).permute(1,2,0).cpu().clamp_(-1, 1).numpy() + 1 )/2. * 255)
+            imgs.append(cv2.cvtColor(img_stablized,cv2.COLOR_RGB2BGR)) 
+        np.save('temp/'+base_name+'_stablized.npy',imgs)
+        del D_Net
+    else:
+        print('[Step 3] Using saved stabilized video.')
+        imgs = np.load('temp/'+base_name+'_stablized.npy')
+    torch.cuda.empty_cache()
+
+    if not args.audio.endswith('.wav'):
+        command = 'ffmpeg -loglevel error -y -i {} -strict -2 {}'.format(args.audio, 'temp/{}/temp.wav'.format(args.tmp_dir))
+        subprocess.call(command, shell=True)
+        args.audio = 'temp/{}/temp.wav'.format(args.tmp_dir)
+    wav = audio.load_wav(args.audio, 16000)
+    mel = audio.melspectrogram(wav)
+    if np.isnan(mel.reshape(-1)).sum() > 0:
+        raise ValueError('Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again')
+
+    mel_step_size, mel_idx_multiplier, i, mel_chunks = 16, 80./fps, 0, []
+    while True:
+        start_idx = int(i * mel_idx_multiplier)
+        if start_idx + mel_step_size > len(mel[0]):
+            mel_chunks.append(mel[:, len(mel[0]) - mel_step_size:])
+            break
+        mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+        i += 1
+
+    print("[Step 4] Load audio; Length of mel chunks: {}".format(len(mel_chunks)))
+    imgs = imgs[:len(mel_chunks)]
+    full_frames = full_frames[:len(mel_chunks)]  
+    lm = lm[:len(mel_chunks)]
+    
+    imgs_enhanced = []
+    for idx in tqdm(range(len(imgs)), desc='[Step 5] Reference Enhancement'):
+        img = imgs[idx]
+        pred, _, _ = enhancer.process(img, img, face_enhance=True, possion_blending=False)
+        imgs_enhanced.append(pred)
+    gen = datagen(imgs_enhanced.copy(), mel_chunks, full_frames, None, (oy1,oy2,ox1,ox2))
+
+    frame_h, frame_w = full_frames[0].shape[:-1]
+    out = cv2.VideoWriter('temp/{}/result.mp4'.format(args.tmp_dir), cv2.VideoWriter_fourcc(*'mp4v'), fps, (frame_w, frame_h))
+    
+    if args.up_face != 'original':
+        instance = GANimationModel()
+        instance.initialize()
+        instance.setup()
+
+    kp_extractor = KeypointExtractor()
+    for i, (img_batch, mel_batch, frames, coords, img_original, f_frames) in enumerate(tqdm(gen, desc='[Step 6] Lip Synthesis:', total=int(np.ceil(float(len(mel_chunks)) / args.LNet_batch_size)))):
+        img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
+        mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
+        img_original = torch.FloatTensor(np.transpose(img_original, (0, 3, 1, 2))).to(device)/255. # BGR -> RGB
+        
+        with torch.no_grad():
+            incomplete, reference = torch.split(img_batch, 3, dim=1) 
+            pred, low_res = model(mel_batch, img_batch, reference)
+            pred = torch.clamp(pred, 0, 1)
+
+            if args.up_face in ['sad', 'angry', 'surprise']:
+                tar_aus = exp_aus_dict[args.up_face]
+            else:
+                pass
+            
+            if args.up_face == 'original':
+                cur_gen_faces = img_original
+            else:
+                test_batch = {'src_img': torch.nn.functional.interpolate((img_original * 2 - 1), size=(128, 128), mode='bilinear'), 
+                              'tar_aus': tar_aus.repeat(len(incomplete), 1)}
+                instance.feed_batch(test_batch)
+                instance.forward()
+                cur_gen_faces = torch.nn.functional.interpolate(instance.fake_img / 2. + 0.5, size=(384, 384), mode='bilinear')
+                
+            if args.without_rl1 is not False:
+                incomplete, reference = torch.split(img_batch, 3, dim=1)
+                mask = torch.where(incomplete==0, torch.ones_like(incomplete), torch.zeros_like(incomplete)) 
+                pred = pred * mask + cur_gen_faces * (1 - mask) 
+        
+        pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
+
+        torch.cuda.empty_cache()
+        for p, f, xf, c in zip(pred, frames, f_frames, coords):
+            y1, y2, x1, x2 = c
+            p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
+            
+            ff = xf.copy() 
+            ff[y1:y2, x1:x2] = p
+            
+            # month region enhancement by GFPGAN
+            cropped_faces, restored_faces, restored_img = restorer.enhance(
+                ff, has_aligned=False, only_center_face=True, paste_back=True)
+                # 0,   1,   2,   3,   4,   5,   6,   7,   8,  9, 10,  11,  12,
+            mm = [0,   0,   0,   0,   0,   0,   0,   0,   0,  0, 255, 255, 255, 0, 0, 0, 0, 0, 0]
+            mouse_mask = np.zeros_like(restored_img)
+            tmp_mask = enhancer.faceparser.process(restored_img[y1:y2, x1:x2], mm)[0]
+            mouse_mask[y1:y2, x1:x2]= cv2.resize(tmp_mask, (x2 - x1, y2 - y1))[:, :, np.newaxis] / 255.
+
+            height, width = ff.shape[:2]
+            restored_img, ff, full_mask = [cv2.resize(x, (512, 512)) for x in (restored_img, ff, np.float32(mouse_mask))]
+            img = Laplacian_Pyramid_Blending_with_mask(restored_img, ff, full_mask[:, :, 0], 10)
+            pp = np.uint8(cv2.resize(np.clip(img, 0 ,255), (width, height)))
+
+            pp, orig_faces, enhanced_faces = enhancer.process(pp, xf, bbox=c, face_enhance=False, possion_blending=True)
+            out.write(pp)
+    out.release()
+    
+    if not os.path.isdir(os.path.dirname(args.outfile)):
+        os.makedirs(os.path.dirname(args.outfile), exist_ok=True)
+    command = 'ffmpeg -loglevel error -y -i {} -i {} -strict -2 -q:v 1 {}'.format(args.audio, 'temp/{}/result.mp4'.format(args.tmp_dir), args.outfile)
+    subprocess.call(command, shell=platform.system() != 'Windows')
+    print('outfile:', args.outfile)
+
+
+# frames:256x256, full_frames: original size
+def datagen(frames, mels, full_frames, frames_pil, cox):
+    img_batch, mel_batch, frame_batch, coords_batch, ref_batch, full_frame_batch = [], [], [], [], [], []
+    base_name = args.face.split('/')[-1]
+    refs = []
+    image_size = 256 
+
+    # original frames
+    kp_extractor = KeypointExtractor()
+    fr_pil = [Image.fromarray(frame) for frame in frames]
+    lms = kp_extractor.extract_keypoint(fr_pil, 'temp/'+base_name+'x12_landmarks.txt')
+    frames_pil = [ (lm, frame) for frame,lm in zip(fr_pil, lms)] # frames is the croped version of modified face
+    crops, orig_images, quads  = crop_faces(image_size, frames_pil, scale=1.0, use_fa=True)
+    inverse_transforms = [calc_alignment_coefficients(quad + 0.5, [[0, 0], [0, image_size], [image_size, image_size], [image_size, 0]]) for quad in quads]
+    del kp_extractor.detector
+
+    oy1,oy2,ox1,ox2 = cox
+    face_det_results = face_detect(full_frames, args, jaw_correction=True)
+
+    for inverse_transform, crop, full_frame, face_det in zip(inverse_transforms, crops, full_frames, face_det_results):
+        imc_pil = paste_image(inverse_transform, crop, Image.fromarray(
+            cv2.resize(full_frame[int(oy1):int(oy2), int(ox1):int(ox2)], (256, 256))))
+
+        ff = full_frame.copy()
+        ff[int(oy1):int(oy2), int(ox1):int(ox2)] = cv2.resize(np.array(imc_pil.convert('RGB')), (ox2 - ox1, oy2 - oy1))
+        oface, coords = face_det
+        y1, y2, x1, x2 = coords
+        refs.append(ff[y1: y2, x1:x2])
+
+    for i, m in enumerate(mels):
+        idx = 0 if args.static else i % len(frames)
+        frame_to_save = frames[idx].copy()
+        face = refs[idx]
+        oface, coords = face_det_results[idx].copy()
+
+        face = cv2.resize(face, (args.img_size, args.img_size))
+        oface = cv2.resize(oface, (args.img_size, args.img_size))
+
+        img_batch.append(oface)
+        ref_batch.append(face) 
+        mel_batch.append(m)
+        coords_batch.append(coords)
+        frame_batch.append(frame_to_save)
+        full_frame_batch.append(full_frames[idx].copy())
+
+        if len(img_batch) >= args.LNet_batch_size:
+            img_batch, mel_batch, ref_batch = np.asarray(img_batch), np.asarray(mel_batch), np.asarray(ref_batch)
+            img_masked = img_batch.copy()
+            img_original = img_batch.copy()
+            img_masked[:, args.img_size//2:] = 0
+            img_batch = np.concatenate((img_masked, ref_batch), axis=3) / 255.
+            mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+
+            yield img_batch, mel_batch, frame_batch, coords_batch, img_original, full_frame_batch
+            img_batch, mel_batch, frame_batch, coords_batch, img_original, full_frame_batch, ref_batch  = [], [], [], [], [], [], []
+
+    if len(img_batch) > 0:
+        img_batch, mel_batch, ref_batch = np.asarray(img_batch), np.asarray(mel_batch), np.asarray(ref_batch)
+        img_masked = img_batch.copy()
+        img_original = img_batch.copy()
+        img_masked[:, args.img_size//2:] = 0
+        img_batch = np.concatenate((img_masked, ref_batch), axis=3) / 255.
+        mel_batch = np.reshape(mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
+        yield img_batch, mel_batch, frame_batch, coords_batch, img_original, full_frame_batch
+
+
+if __name__ == '__main__':
+    main()

video-tetalking/inference_videoretalking.sh

@@ -0,0 +1,4 @@
+python3 inference.py \
+  --face ./examples/face/1.mp4 \
+  --audio ./examples/audio/1.wav \
+  --outfile results/1_1.mp4

video-tetalking/models/DNet.py

@@ -0,0 +1,118 @@
+# TODO
+import functools
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from utils import flow_util
+from models.base_blocks import LayerNorm2d, ADAINHourglass, FineEncoder, FineDecoder
+
+# DNet
+class DNet(nn.Module):
+    def __init__(self):  
+        super(DNet, self).__init__()
+        self.mapping_net = MappingNet()
+        self.warpping_net = WarpingNet()
+        self.editing_net = EditingNet()
+ 
+    def forward(self, input_image, driving_source, stage=None):
+        if stage == 'warp':
+            descriptor = self.mapping_net(driving_source)
+            output = self.warpping_net(input_image, descriptor)
+        else:
+            descriptor = self.mapping_net(driving_source)
+            output = self.warpping_net(input_image, descriptor)
+            output['fake_image'] = self.editing_net(input_image, output['warp_image'], descriptor)
+        return output
+
+class MappingNet(nn.Module):
+    def __init__(self, coeff_nc=73, descriptor_nc=256, layer=3):
+        super( MappingNet, self).__init__()
+
+        self.layer = layer
+        nonlinearity = nn.LeakyReLU(0.1)
+
+        self.first = nn.Sequential(
+            torch.nn.Conv1d(coeff_nc, descriptor_nc, kernel_size=7, padding=0, bias=True))
+
+        for i in range(layer):
+            net = nn.Sequential(nonlinearity,
+                torch.nn.Conv1d(descriptor_nc, descriptor_nc, kernel_size=3, padding=0, dilation=3))
+            setattr(self, 'encoder' + str(i), net)   
+
+        self.pooling = nn.AdaptiveAvgPool1d(1)
+        self.output_nc = descriptor_nc
+
+    def forward(self, input_3dmm):
+        out = self.first(input_3dmm)
+        for i in range(self.layer):
+            model = getattr(self, 'encoder' + str(i))
+            out = model(out) + out[:,:,3:-3]
+        out = self.pooling(out)
+        return out   
+
+class WarpingNet(nn.Module):
+    def __init__(
+        self, 
+        image_nc=3, 
+        descriptor_nc=256, 
+        base_nc=32, 
+        max_nc=256, 
+        encoder_layer=5, 
+        decoder_layer=3, 
+        use_spect=False
+        ):
+        super( WarpingNet, self).__init__()
+
+        nonlinearity = nn.LeakyReLU(0.1)
+        norm_layer = functools.partial(LayerNorm2d, affine=True) 
+        kwargs = {'nonlinearity':nonlinearity, 'use_spect':use_spect}
+
+        self.descriptor_nc = descriptor_nc 
+        self.hourglass = ADAINHourglass(image_nc, self.descriptor_nc, base_nc,
+                                       max_nc, encoder_layer, decoder_layer, **kwargs)
+
+        self.flow_out = nn.Sequential(norm_layer(self.hourglass.output_nc), 
+                                      nonlinearity,
+                                      nn.Conv2d(self.hourglass.output_nc, 2, kernel_size=7, stride=1, padding=3))
+
+        self.pool = nn.AdaptiveAvgPool2d(1)
+
+    def forward(self, input_image, descriptor):
+        final_output={}
+        output = self.hourglass(input_image, descriptor)
+        final_output['flow_field'] = self.flow_out(output)
+
+        deformation = flow_util.convert_flow_to_deformation(final_output['flow_field'])
+        final_output['warp_image'] = flow_util.warp_image(input_image, deformation)
+        return final_output
+
+
+class EditingNet(nn.Module):
+    def __init__(
+        self, 
+        image_nc=3, 
+        descriptor_nc=256, 
+        layer=3, 
+        base_nc=64, 
+        max_nc=256, 
+        num_res_blocks=2, 
+        use_spect=False):  
+        super(EditingNet, self).__init__()
+
+        nonlinearity = nn.LeakyReLU(0.1)
+        norm_layer = functools.partial(LayerNorm2d, affine=True) 
+        kwargs = {'norm_layer':norm_layer, 'nonlinearity':nonlinearity, 'use_spect':use_spect}
+        self.descriptor_nc = descriptor_nc
+
+        # encoder part
+        self.encoder = FineEncoder(image_nc*2, base_nc, max_nc, layer, **kwargs)
+        self.decoder = FineDecoder(image_nc, self.descriptor_nc, base_nc, max_nc, layer, num_res_blocks, **kwargs)
+
+    def forward(self, input_image, warp_image, descriptor):
+        x = torch.cat([input_image, warp_image], 1)
+        x = self.encoder(x)
+        gen_image = self.decoder(x, descriptor)
+        return gen_image

video-tetalking/models/ENet.py

@@ -0,0 +1,139 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from models.base_blocks import ResBlock, StyleConv, ToRGB
+
+
+class ENet(nn.Module):
+    def __init__(
+        self, 
+        num_style_feat=512,
+        lnet=None,
+        concat=False
+        ):  
+        super(ENet, self).__init__()
+
+        self.low_res = lnet
+        for param in self.low_res.parameters():
+            param.requires_grad = False
+
+        channel_multiplier, narrow = 2, 1
+        channels = {
+            '4': int(512 * narrow),
+            '8': int(512 * narrow),
+            '16': int(512 * narrow),
+            '32': int(512 * narrow),
+            '64': int(256 * channel_multiplier * narrow),
+            '128': int(128 * channel_multiplier * narrow),
+            '256': int(64 * channel_multiplier * narrow),
+            '512': int(32 * channel_multiplier * narrow),
+            '1024': int(16 * channel_multiplier * narrow)
+        }
+
+        self.log_size = 8
+        first_out_size = 128
+        self.conv_body_first = nn.Conv2d(3, channels[f'{first_out_size}'], 1) # 256 -> 128
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(8, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels, mode='down'))
+            in_channels = out_channels
+
+        self.num_style_feat = num_style_feat
+        linear_out_channel = num_style_feat
+        self.final_linear = nn.Linear(channels['4'] * 4 * 4, linear_out_channel)
+        self.final_conv = nn.Conv2d(in_channels, channels['4'], 3, 1, 1)
+
+        self.style_convs = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+        
+        self.concat = concat
+        if concat:
+            in_channels = 3 + 32 # channels['64']
+        else:
+            in_channels = 3
+
+        for i in range(7, 9):  # 128, 256
+            out_channels = channels[f'{2**i}'] # 
+            self.style_convs.append(
+                StyleConv(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode='upsample'))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None))
+            self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
+            in_channels = out_channels
+
+    def forward(self, audio_sequences, face_sequences, gt_sequences):
+        B = audio_sequences.size(0)
+        input_dim_size = len(face_sequences.size())
+        inp, ref = torch.split(face_sequences,3,dim=1)
+
+        if input_dim_size > 4:
+            audio_sequences = torch.cat([audio_sequences[:, i] for i in range(audio_sequences.size(1))], dim=0)
+            inp = torch.cat([inp[:, :, i] for i in range(inp.size(2))], dim=0)
+            ref = torch.cat([ref[:, :, i] for i in range(ref.size(2))], dim=0)
+            gt_sequences = torch.cat([gt_sequences[:, :, i] for i in range(gt_sequences.size(2))], dim=0)
+        
+        # get the global style
+        feat = F.leaky_relu_(self.conv_body_first(F.interpolate(ref, size=(256,256), mode='bilinear')), negative_slope=0.2)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+        feat = F.leaky_relu_(self.final_conv(feat), negative_slope=0.2)
+
+        # style code
+        style_code = self.final_linear(feat.reshape(feat.size(0), -1))
+        style_code = style_code.reshape(style_code.size(0), -1, self.num_style_feat)
+        
+        LNet_input = torch.cat([inp, gt_sequences], dim=1)
+        LNet_input = F.interpolate(LNet_input, size=(96,96), mode='bilinear')
+        
+        if self.concat:
+            low_res_img, low_res_feat = self.low_res(audio_sequences, LNet_input)
+            low_res_img.detach()
+            low_res_feat.detach()
+            out = torch.cat([low_res_img, low_res_feat], dim=1) 
+
+        else:
+            low_res_img = self.low_res(audio_sequences, LNet_input)
+            low_res_img.detach()
+            # 96 x 96
+            out = low_res_img 
+        
+        p2d = (2,2,2,2)
+        out = F.pad(out, p2d, "reflect", 0)
+        skip = out
+
+        for conv1, conv2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], self.to_rgbs):
+            out = conv1(out, style_code)  # 96, 192, 384
+            out = conv2(out, style_code)
+            skip = to_rgb(out, style_code, skip)
+        _outputs = skip
+
+        # remove padding
+        _outputs = _outputs[:,:,8:-8,8:-8]
+
+        if input_dim_size > 4:
+            _outputs = torch.split(_outputs, B, dim=0)
+            outputs = torch.stack(_outputs, dim=2)
+            low_res_img = F.interpolate(low_res_img, outputs.size()[3:])
+            low_res_img = torch.split(low_res_img, B, dim=0) 
+            low_res_img = torch.stack(low_res_img, dim=2)
+        else:
+            outputs = _outputs
+        return outputs, low_res_img

video-tetalking/models/LNet.py

@@ -0,0 +1,139 @@
+import functools
+import torch
+import torch.nn as nn
+
+from models.transformer import RETURNX, Transformer
+from models.base_blocks import Conv2d, LayerNorm2d, FirstBlock2d, DownBlock2d, UpBlock2d, \
+                               FFCADAINResBlocks, Jump, FinalBlock2d
+
+
+class Visual_Encoder(nn.Module):
+    def __init__(self, image_nc, ngf, img_f, layers, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(Visual_Encoder, self).__init__()
+        self.layers = layers
+        self.first_inp = FirstBlock2d(image_nc, ngf, norm_layer, nonlinearity, use_spect)
+        self.first_ref = FirstBlock2d(image_nc, ngf, norm_layer, nonlinearity, use_spect)
+        for i in range(layers):
+            in_channels = min(ngf*(2**i), img_f)
+            out_channels = min(ngf*(2**(i+1)), img_f)
+            model_ref = DownBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
+            model_inp = DownBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
+            if i < 2:
+                ca_layer = RETURNX()
+            else:
+                ca_layer = Transformer(2**(i+1) * ngf,2,4,ngf,ngf*4)
+            setattr(self, 'ca' + str(i), ca_layer)
+            setattr(self, 'ref_down' + str(i), model_ref)
+            setattr(self, 'inp_down' + str(i), model_inp)
+        self.output_nc = out_channels * 2
+
+    def forward(self, maskGT, ref):
+        x_maskGT, x_ref = self.first_inp(maskGT), self.first_ref(ref)
+        out=[x_maskGT]
+        for i in range(self.layers):
+            model_ref = getattr(self, 'ref_down'+str(i))
+            model_inp = getattr(self, 'inp_down'+str(i))
+            ca_layer = getattr(self, 'ca'+str(i))
+            x_maskGT, x_ref = model_inp(x_maskGT), model_ref(x_ref)
+            x_maskGT = ca_layer(x_maskGT, x_ref)
+            if i < self.layers - 1:
+                out.append(x_maskGT)
+            else:           
+                out.append(torch.cat([x_maskGT, x_ref], dim=1)) # concat ref features !
+        return out
+
+
+class Decoder(nn.Module):
+    def __init__(self, image_nc, feature_nc, ngf, img_f, layers, num_block, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(Decoder, self).__init__()
+        self.layers = layers
+        for i in range(layers)[::-1]:
+            if  i == layers-1:
+                in_channels = ngf*(2**(i+1)) * 2
+            else:
+                in_channels = min(ngf*(2**(i+1)), img_f)
+            out_channels = min(ngf*(2**i), img_f)
+            up = UpBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
+            res = FFCADAINResBlocks(num_block, in_channels, feature_nc, norm_layer, nonlinearity, use_spect)
+            jump = Jump(out_channels, norm_layer, nonlinearity, use_spect)
+
+            setattr(self, 'up' + str(i), up)
+            setattr(self, 'res' + str(i), res)            
+            setattr(self, 'jump' + str(i), jump)
+
+        self.final = FinalBlock2d(out_channels, image_nc, use_spect, 'sigmoid')
+        self.output_nc = out_channels
+
+    def forward(self, x, z):
+        out = x.pop()
+        for i in range(self.layers)[::-1]:
+            res_model = getattr(self, 'res' + str(i))
+            up_model = getattr(self, 'up' + str(i))
+            jump_model = getattr(self, 'jump' + str(i))
+            out = res_model(out, z)
+            out = up_model(out)
+            out = jump_model(x.pop()) + out
+        out_image = self.final(out)
+        return out_image
+
+
+class LNet(nn.Module): 
+    def __init__(
+        self, 
+        image_nc=3, 
+        descriptor_nc=512, 
+        layer=3, 
+        base_nc=64, 
+        max_nc=512, 
+        num_res_blocks=9, 
+        use_spect=True,
+        encoder=Visual_Encoder,
+        decoder=Decoder
+        ):  
+        super(LNet, self).__init__()
+
+        nonlinearity = nn.LeakyReLU(0.1)
+        norm_layer = functools.partial(LayerNorm2d, affine=True) 
+        kwargs = {'norm_layer':norm_layer, 'nonlinearity':nonlinearity, 'use_spect':use_spect}
+        self.descriptor_nc = descriptor_nc
+
+        self.encoder = encoder(image_nc, base_nc, max_nc, layer, **kwargs)
+        self.decoder = decoder(image_nc, self.descriptor_nc, base_nc, max_nc, layer, num_res_blocks, **kwargs)
+        self.audio_encoder = nn.Sequential(
+            Conv2d(1, 32, kernel_size=3, stride=1, padding=1),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(32, 32, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(32, 64, kernel_size=3, stride=(3, 1), padding=1),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(64, 64, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(64, 128, kernel_size=3, stride=3, padding=1),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+            Conv2d(128, 128, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(128, 256, kernel_size=3, stride=(3, 2), padding=1),
+            Conv2d(256, 256, kernel_size=3, stride=1, padding=1, residual=True),
+
+            Conv2d(256, 512, kernel_size=3, stride=1, padding=0),
+            Conv2d(512, descriptor_nc, kernel_size=1, stride=1, padding=0),
+            )
+
+    def forward(self, audio_sequences, face_sequences):
+        B = audio_sequences.size(0)
+        input_dim_size = len(face_sequences.size())
+        if input_dim_size > 4:
+            audio_sequences = torch.cat([audio_sequences[:, i] for i in range(audio_sequences.size(1))], dim=0)
+            face_sequences = torch.cat([face_sequences[:, :, i] for i in range(face_sequences.size(2))], dim=0)
+        cropped, ref = torch.split(face_sequences, 3, dim=1)
+
+        vis_feat = self.encoder(cropped, ref)
+        audio_feat = self.audio_encoder(audio_sequences) 
+        _outputs = self.decoder(vis_feat, audio_feat)
+
+        if input_dim_size > 4:
+            _outputs = torch.split(_outputs, B, dim=0)
+            outputs = torch.stack(_outputs, dim=2) 
+        else:
+            outputs = _outputs
+        return outputs

video-tetalking/models/__init__.py

@@ -0,0 +1,37 @@
+import torch
+from models.DNet import DNet
+from models.LNet import LNet
+from models.ENet import ENet
+
+
+def _load(checkpoint_path):
+    map_location=None if torch.cuda.is_available() else torch.device('cpu')
+    checkpoint = torch.load(checkpoint_path, map_location=map_location)
+    return checkpoint
+
+def load_checkpoint(path, model):
+    print("Load checkpoint from: {}".format(path))
+    checkpoint = _load(path)
+    s = checkpoint["state_dict"] if 'arcface' not in path else checkpoint
+    new_s = {}
+    for k, v in s.items():
+        if 'low_res' in k:
+            continue
+        else:
+            new_s[k.replace('module.', '')] = v
+    model.load_state_dict(new_s, strict=False)
+    return model
+
+def load_network(args):
+    L_net = LNet()
+    L_net = load_checkpoint(args.LNet_path, L_net)
+    E_net = ENet(lnet=L_net)
+    model = load_checkpoint(args.ENet_path, E_net)
+    return model.eval()
+
+def load_DNet(args):
+    D_Net = DNet()
+    print("Load checkpoint from: {}".format(args.DNet_path))
+    checkpoint =  torch.load(args.DNet_path, map_location=lambda storage, loc: storage)
+    D_Net.load_state_dict(checkpoint['net_G_ema'], strict=False)
+    return D_Net.eval()

video-tetalking/models/base_blocks.py

@@ -0,0 +1,554 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.modules.batchnorm import BatchNorm2d
+from torch.nn.utils.spectral_norm import spectral_norm as SpectralNorm
+
+from models.ffc import FFC
+from basicsr.archs.arch_util import default_init_weights
+
+
+class Conv2d(nn.Module):
+    def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.conv_block = nn.Sequential(
+                            nn.Conv2d(cin, cout, kernel_size, stride, padding),
+                            nn.BatchNorm2d(cout)
+                            )
+        self.act = nn.ReLU()
+        self.residual = residual
+
+    def forward(self, x):
+        out = self.conv_block(x)
+        if self.residual:
+            out += x
+        return self.act(out)
+
+
+class ResBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, mode='down'):
+        super(ResBlock, self).__init__()
+        self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
+        self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
+        self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
+        if mode == 'down':
+            self.scale_factor = 0.5
+        elif mode == 'up':
+            self.scale_factor = 2
+
+    def forward(self, x):
+        out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
+        # upsample/downsample
+        out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
+        # skip
+        x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        skip = self.skip(x)
+        out = out + skip
+        return out
+
+
+class LayerNorm2d(nn.Module):
+    def __init__(self, n_out, affine=True):
+        super(LayerNorm2d, self).__init__()
+        self.n_out = n_out
+        self.affine = affine
+
+        if self.affine:
+          self.weight = nn.Parameter(torch.ones(n_out, 1, 1))
+          self.bias = nn.Parameter(torch.zeros(n_out, 1, 1))
+
+    def forward(self, x):
+        normalized_shape = x.size()[1:]
+        if self.affine:
+          return F.layer_norm(x, normalized_shape, \
+              self.weight.expand(normalized_shape), 
+              self.bias.expand(normalized_shape))    
+        else:
+          return F.layer_norm(x, normalized_shape)  
+
+
+def spectral_norm(module, use_spect=True):
+    if use_spect:
+        return SpectralNorm(module)
+    else:
+        return module
+
+
+class FirstBlock2d(nn.Module):
+    def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FirstBlock2d, self).__init__()
+        kwargs = {'kernel_size': 7, 'stride': 1, 'padding': 3}
+        conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
+
+        if type(norm_layer) == type(None):
+            self.model = nn.Sequential(conv, nonlinearity)
+        else:
+            self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
+
+    def forward(self, x):
+        out = self.model(x)
+        return out 
+
+
+class DownBlock2d(nn.Module):
+    def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(DownBlock2d, self).__init__()
+        kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+        conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
+        pool = nn.AvgPool2d(kernel_size=(2, 2))
+
+        if type(norm_layer) == type(None):
+            self.model = nn.Sequential(conv, nonlinearity, pool)
+        else:
+            self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity, pool)
+
+    def forward(self, x):
+        out = self.model(x)
+        return out 
+
+
+class UpBlock2d(nn.Module):
+    def __init__(self, input_nc, output_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(UpBlock2d, self).__init__()
+        kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+        conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
+        if type(norm_layer) == type(None):
+            self.model = nn.Sequential(conv, nonlinearity)
+        else:
+            self.model = nn.Sequential(conv, norm_layer(output_nc), nonlinearity)
+
+    def forward(self, x):
+        out = self.model(F.interpolate(x, scale_factor=2))
+        return out
+
+
+class ADAIN(nn.Module):
+    def __init__(self, norm_nc, feature_nc):
+        super().__init__()
+
+        self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False)
+
+        nhidden = 128
+        use_bias=True
+
+        self.mlp_shared = nn.Sequential(
+            nn.Linear(feature_nc, nhidden, bias=use_bias),            
+            nn.ReLU()
+        )
+        self.mlp_gamma = nn.Linear(nhidden, norm_nc, bias=use_bias)    
+        self.mlp_beta = nn.Linear(nhidden, norm_nc, bias=use_bias)    
+
+    def forward(self, x, feature):
+
+        # Part 1. generate parameter-free normalized activations
+        normalized = self.param_free_norm(x)
+        # Part 2. produce scaling and bias conditioned on feature
+        feature = feature.view(feature.size(0), -1)
+        actv = self.mlp_shared(feature)
+        gamma = self.mlp_gamma(actv)
+        beta = self.mlp_beta(actv)
+
+        # apply scale and bias
+        gamma = gamma.view(*gamma.size()[:2], 1,1)
+        beta = beta.view(*beta.size()[:2], 1,1)
+        out = normalized * (1 + gamma) + beta
+        return out
+
+
+class FineADAINResBlock2d(nn.Module):
+    """
+    Define an Residual block for different types
+    """
+    def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FineADAINResBlock2d, self).__init__()
+        kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+        self.conv1 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
+        self.conv2 = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
+        self.norm1 = ADAIN(input_nc, feature_nc)
+        self.norm2 = ADAIN(input_nc, feature_nc)
+        self.actvn = nonlinearity
+
+    def forward(self, x, z):
+        dx = self.actvn(self.norm1(self.conv1(x), z))
+        dx = self.norm2(self.conv2(x), z)
+        out = dx + x
+        return out  
+
+
+class FineADAINResBlocks(nn.Module):
+    def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FineADAINResBlocks, self).__init__()                                
+        self.num_block = num_block
+        for i in range(num_block):
+            model = FineADAINResBlock2d(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
+            setattr(self, 'res'+str(i), model)
+
+    def forward(self, x, z):
+        for i in range(self.num_block):
+            model = getattr(self, 'res'+str(i))
+            x = model(x, z)
+        return x   
+
+
+class ADAINEncoderBlock(nn.Module):       
+    def __init__(self, input_nc, output_nc, feature_nc, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(ADAINEncoderBlock, self).__init__()
+        kwargs_down = {'kernel_size': 4, 'stride': 2, 'padding': 1}
+        kwargs_fine = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+
+        self.conv_0 = spectral_norm(nn.Conv2d(input_nc,  output_nc, **kwargs_down), use_spect)
+        self.conv_1 = spectral_norm(nn.Conv2d(output_nc, output_nc, **kwargs_fine), use_spect)
+
+
+        self.norm_0 = ADAIN(input_nc, feature_nc)
+        self.norm_1 = ADAIN(output_nc, feature_nc)
+        self.actvn = nonlinearity
+
+    def forward(self, x, z):
+        x = self.conv_0(self.actvn(self.norm_0(x, z)))
+        x = self.conv_1(self.actvn(self.norm_1(x, z)))
+        return x
+
+
+class ADAINDecoderBlock(nn.Module):
+    def __init__(self, input_nc, output_nc, hidden_nc, feature_nc, use_transpose=True, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(ADAINDecoderBlock, self).__init__()        
+        # Attributes
+        self.actvn = nonlinearity
+        hidden_nc = min(input_nc, output_nc) if hidden_nc is None else hidden_nc
+
+        kwargs_fine = {'kernel_size':3, 'stride':1, 'padding':1}
+        if use_transpose:
+            kwargs_up = {'kernel_size':3, 'stride':2, 'padding':1, 'output_padding':1}
+        else:
+            kwargs_up = {'kernel_size':3, 'stride':1, 'padding':1}
+
+        # create conv layers
+        self.conv_0 = spectral_norm(nn.Conv2d(input_nc, hidden_nc, **kwargs_fine), use_spect)
+        if use_transpose:
+            self.conv_1 = spectral_norm(nn.ConvTranspose2d(hidden_nc, output_nc, **kwargs_up), use_spect)
+            self.conv_s = spectral_norm(nn.ConvTranspose2d(input_nc, output_nc, **kwargs_up), use_spect)
+        else:
+            self.conv_1 = nn.Sequential(spectral_norm(nn.Conv2d(hidden_nc, output_nc, **kwargs_up), use_spect),
+                                        nn.Upsample(scale_factor=2))
+            self.conv_s = nn.Sequential(spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs_up), use_spect),
+                                        nn.Upsample(scale_factor=2))
+        # define normalization layers
+        self.norm_0 = ADAIN(input_nc, feature_nc)
+        self.norm_1 = ADAIN(hidden_nc, feature_nc)
+        self.norm_s = ADAIN(input_nc, feature_nc)
+        
+    def forward(self, x, z):
+        x_s = self.shortcut(x, z)
+        dx = self.conv_0(self.actvn(self.norm_0(x, z)))
+        dx = self.conv_1(self.actvn(self.norm_1(dx, z)))
+        out = x_s + dx
+        return out
+
+    def shortcut(self, x, z):
+        x_s = self.conv_s(self.actvn(self.norm_s(x, z)))
+        return x_s   
+
+
+class FineEncoder(nn.Module):
+    """docstring for Encoder"""
+    def __init__(self, image_nc, ngf, img_f, layers, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FineEncoder, self).__init__()
+        self.layers = layers
+        self.first = FirstBlock2d(image_nc, ngf, norm_layer, nonlinearity, use_spect)
+        for i in range(layers):
+            in_channels = min(ngf*(2**i), img_f)
+            out_channels = min(ngf*(2**(i+1)), img_f)
+            model = DownBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
+            setattr(self, 'down' + str(i), model)
+        self.output_nc = out_channels
+
+    def forward(self, x):
+        x = self.first(x)
+        out=[x]
+        for i in range(self.layers):
+            model = getattr(self, 'down'+str(i))
+            x = model(x)
+            out.append(x)
+        return out
+
+
+class FineDecoder(nn.Module):
+    """docstring for FineDecoder"""
+    def __init__(self, image_nc, feature_nc, ngf, img_f, layers, num_block, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FineDecoder, self).__init__()
+        self.layers = layers
+        for i in range(layers)[::-1]:
+            in_channels = min(ngf*(2**(i+1)), img_f)
+            out_channels = min(ngf*(2**i), img_f)
+            up = UpBlock2d(in_channels, out_channels, norm_layer, nonlinearity, use_spect)
+            res = FineADAINResBlocks(num_block, in_channels, feature_nc, norm_layer, nonlinearity, use_spect)
+            jump = Jump(out_channels, norm_layer, nonlinearity, use_spect)
+            setattr(self, 'up' + str(i), up)
+            setattr(self, 'res' + str(i), res)            
+            setattr(self, 'jump' + str(i), jump)
+        self.final = FinalBlock2d(out_channels, image_nc, use_spect, 'tanh')
+        self.output_nc = out_channels
+
+    def forward(self, x, z):
+        out = x.pop()
+        for i in range(self.layers)[::-1]:
+            res_model = getattr(self, 'res' + str(i))
+            up_model = getattr(self, 'up' + str(i))
+            jump_model = getattr(self, 'jump' + str(i))
+            out = res_model(out, z)
+            out = up_model(out)
+            out = jump_model(x.pop()) + out
+        out_image = self.final(out)
+        return out_image
+
+
+class ADAINEncoder(nn.Module):
+    def __init__(self, image_nc, pose_nc, ngf, img_f, layers, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(ADAINEncoder, self).__init__()
+        self.layers = layers
+        self.input_layer = nn.Conv2d(image_nc, ngf, kernel_size=7, stride=1, padding=3)
+        for i in range(layers):
+            in_channels = min(ngf * (2**i), img_f)
+            out_channels = min(ngf *(2**(i+1)), img_f)
+            model = ADAINEncoderBlock(in_channels, out_channels, pose_nc, nonlinearity, use_spect)
+            setattr(self, 'encoder' + str(i), model)
+        self.output_nc = out_channels
+        
+    def forward(self, x, z):
+        out = self.input_layer(x)
+        out_list = [out]
+        for i in range(self.layers):
+            model = getattr(self, 'encoder' + str(i))
+            out = model(out, z)
+            out_list.append(out)
+        return out_list
+        
+        
+class ADAINDecoder(nn.Module):
+    """docstring for ADAINDecoder"""
+    def __init__(self, pose_nc, ngf, img_f, encoder_layers, decoder_layers, skip_connect=True, 
+                 nonlinearity=nn.LeakyReLU(), use_spect=False):
+
+        super(ADAINDecoder, self).__init__()
+        self.encoder_layers = encoder_layers
+        self.decoder_layers = decoder_layers
+        self.skip_connect = skip_connect
+        use_transpose = True
+        for i in range(encoder_layers-decoder_layers, encoder_layers)[::-1]:
+            in_channels = min(ngf * (2**(i+1)), img_f)
+            in_channels = in_channels*2 if i != (encoder_layers-1) and self.skip_connect else in_channels
+            out_channels = min(ngf * (2**i), img_f)
+            model = ADAINDecoderBlock(in_channels, out_channels, out_channels, pose_nc, use_transpose, nonlinearity, use_spect)
+            setattr(self, 'decoder' + str(i), model)
+        self.output_nc = out_channels*2 if self.skip_connect else out_channels
+
+    def forward(self, x, z):
+        out = x.pop() if self.skip_connect else x
+        for i in range(self.encoder_layers-self.decoder_layers, self.encoder_layers)[::-1]:
+            model = getattr(self, 'decoder' + str(i))
+            out = model(out, z)
+            out = torch.cat([out, x.pop()], 1) if self.skip_connect else out
+        return out
+
+
+class ADAINHourglass(nn.Module):
+    def __init__(self, image_nc, pose_nc, ngf, img_f, encoder_layers, decoder_layers, nonlinearity, use_spect):
+        super(ADAINHourglass, self).__init__()
+        self.encoder = ADAINEncoder(image_nc, pose_nc, ngf, img_f, encoder_layers, nonlinearity, use_spect)
+        self.decoder = ADAINDecoder(pose_nc, ngf, img_f, encoder_layers, decoder_layers, True, nonlinearity, use_spect)
+        self.output_nc = self.decoder.output_nc
+
+    def forward(self, x, z):
+        return self.decoder(self.encoder(x, z), z)        
+
+
+class FineADAINLama(nn.Module):
+    def __init__(self, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FineADAINLama, self).__init__()
+        kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+        self.actvn = nonlinearity
+        ratio_gin = 0.75
+        ratio_gout = 0.75        
+        self.ffc = FFC(input_nc, input_nc, 3,
+                       ratio_gin, ratio_gout, 1, 1, 1,
+                       1, False, False, padding_type='reflect')
+        global_channels = int(input_nc * ratio_gout)
+        self.bn_l = ADAIN(input_nc - global_channels, feature_nc)
+        self.bn_g = ADAIN(global_channels, feature_nc)
+
+    def forward(self, x, z):
+        x_l, x_g = self.ffc(x)
+        x_l = self.actvn(self.bn_l(x_l,z))
+        x_g = self.actvn(self.bn_g(x_g,z))
+        return x_l, x_g
+
+
+class FFCResnetBlock(nn.Module):
+    def __init__(self, dim, feature_dim, padding_type='reflect', norm_layer=BatchNorm2d, activation_layer=nn.ReLU, dilation=1,
+                 spatial_transform_kwargs=None, inline=False, **conv_kwargs):
+        super().__init__()
+        self.conv1 = FineADAINLama(dim, feature_dim, **conv_kwargs)
+        self.conv2 = FineADAINLama(dim, feature_dim, **conv_kwargs)
+        self.inline = True
+
+    def forward(self, x, z):
+        if self.inline:
+            x_l, x_g = x[:, :-self.conv1.ffc.global_in_num], x[:, -self.conv1.ffc.global_in_num:]
+        else:
+            x_l, x_g = x if type(x) is tuple else (x, 0)
+
+        id_l, id_g = x_l, x_g
+        x_l, x_g = self.conv1((x_l, x_g), z)
+        x_l, x_g = self.conv2((x_l, x_g), z)
+
+        x_l, x_g = id_l + x_l, id_g + x_g
+        out = x_l, x_g
+        if self.inline:
+            out = torch.cat(out, dim=1)
+        return out
+
+
+class FFCADAINResBlocks(nn.Module):
+    def __init__(self, num_block, input_nc, feature_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(FFCADAINResBlocks, self).__init__()                                
+        self.num_block = num_block
+        for i in range(num_block):
+            model = FFCResnetBlock(input_nc, feature_nc, norm_layer, nonlinearity, use_spect)
+            setattr(self, 'res'+str(i), model)
+
+    def forward(self, x, z):
+        for i in range(self.num_block):
+            model = getattr(self, 'res'+str(i))
+            x = model(x, z)
+        return x 
+
+
+class Jump(nn.Module):
+    def __init__(self, input_nc, norm_layer=nn.BatchNorm2d, nonlinearity=nn.LeakyReLU(), use_spect=False):
+        super(Jump, self).__init__()
+        kwargs = {'kernel_size': 3, 'stride': 1, 'padding': 1}
+        conv = spectral_norm(nn.Conv2d(input_nc, input_nc, **kwargs), use_spect)
+        if type(norm_layer) == type(None):
+            self.model = nn.Sequential(conv, nonlinearity)
+        else:
+            self.model = nn.Sequential(conv, norm_layer(input_nc), nonlinearity)
+
+    def forward(self, x):
+        out = self.model(x)
+        return out   
+
+
+class FinalBlock2d(nn.Module):
+    def __init__(self, input_nc, output_nc, use_spect=False, tanh_or_sigmoid='tanh'):
+        super(FinalBlock2d, self).__init__()
+        kwargs = {'kernel_size': 7, 'stride': 1, 'padding':3}
+        conv = spectral_norm(nn.Conv2d(input_nc, output_nc, **kwargs), use_spect)
+        if tanh_or_sigmoid == 'sigmoid':
+            out_nonlinearity = nn.Sigmoid()
+        else:
+            out_nonlinearity = nn.Tanh()            
+        self.model = nn.Sequential(conv, out_nonlinearity)
+
+    def forward(self, x):
+        out = self.model(x)
+        return out    
+
+
+class ModulatedConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 eps=1e-8):
+        super(ModulatedConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.demodulate = demodulate
+        self.sample_mode = sample_mode
+        self.eps = eps
+
+        # modulation inside each modulated conv
+        self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
+        # initialization
+        default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
+            math.sqrt(in_channels * kernel_size**2))
+        self.padding = kernel_size // 2
+
+    def forward(self, x, style):
+        b, c, h, w = x.shape   
+        style = self.modulation(style).view(b, 1, c, 1, 1)
+        weight = self.weight * style  
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
+            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
+
+        # upsample or downsample if necessary
+        if self.sample_mode == 'upsample':
+            x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
+        elif self.sample_mode == 'downsample':
+            x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
+
+        b, c, h, w = x.shape
+        x = x.view(1, b * c, h, w)
+        out = F.conv2d(x, weight, padding=self.padding, groups=b)
+        out = out.view(b, self.out_channels, *out.shape[2:4])
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
+
+
+class StyleConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
+        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
+        self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
+        self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x, style, noise=None):
+        # modulate
+        out = self.modulated_conv(x, style) * 2**0.5  # for conversion
+        # noise injection
+        if noise is None:
+            b, _, h, w = out.shape
+            noise = out.new_empty(b, 1, h, w).normal_()
+        out = out + self.weight * noise
+        # add bias
+        out = out + self.bias
+        # activation
+        out = self.activate(out)
+        return out
+
+
+class ToRGB(nn.Module):
+    def __init__(self, in_channels, num_style_feat, upsample=True):
+        super(ToRGB, self).__init__()
+        self.upsample = upsample
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, x, style, skip=None):
+        out = self.modulated_conv(x, style)
+        out = out + self.bias
+        if skip is not None:
+            if self.upsample:
+                skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
+            out = out + skip
+        return out

video-tetalking/models/ffc.py

@@ -0,0 +1,233 @@
+# Fast Fourier Convolution NeurIPS 2020
+# original implementation https://github.com/pkumivision/FFC/blob/main/model_zoo/ffc.py
+# paper https://proceedings.neurips.cc/paper/2020/file/2fd5d41ec6cfab47e32164d5624269b1-Paper.pdf
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+# from models.modules.squeeze_excitation import SELayer
+import torch.fft
+
+class SELayer(nn.Module):
+    def __init__(self, channel, reduction=16):
+        super(SELayer, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.fc = nn.Sequential(
+            nn.Linear(channel, channel // reduction, bias=False),
+            nn.ReLU(inplace=True),
+            nn.Linear(channel // reduction, channel, bias=False),
+            nn.Sigmoid()
+        )
+
+    def forward(self, x):
+        b, c, _, _ = x.size()
+        y = self.avg_pool(x).view(b, c)
+        y = self.fc(y).view(b, c, 1, 1)
+        res = x * y.expand_as(x)
+        return res
+
+
+class FFCSE_block(nn.Module):
+    def __init__(self, channels, ratio_g):
+        super(FFCSE_block, self).__init__()
+        in_cg = int(channels * ratio_g)
+        in_cl = channels - in_cg
+        r = 16
+
+        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
+        self.conv1 = nn.Conv2d(channels, channels // r,
+                               kernel_size=1, bias=True)
+        self.relu1 = nn.ReLU(inplace=True)
+        self.conv_a2l = None if in_cl == 0 else nn.Conv2d(
+            channels // r, in_cl, kernel_size=1, bias=True)
+        self.conv_a2g = None if in_cg == 0 else nn.Conv2d(
+            channels // r, in_cg, kernel_size=1, bias=True)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        x = x if type(x) is tuple else (x, 0)
+        id_l, id_g = x
+
+        x = id_l if type(id_g) is int else torch.cat([id_l, id_g], dim=1)
+        x = self.avgpool(x)
+        x = self.relu1(self.conv1(x))
+
+        x_l = 0 if self.conv_a2l is None else id_l * \
+            self.sigmoid(self.conv_a2l(x))
+        x_g = 0 if self.conv_a2g is None else id_g * \
+            self.sigmoid(self.conv_a2g(x))
+        return x_l, x_g
+
+
+class FourierUnit(nn.Module):
+
+    def __init__(self, in_channels, out_channels, groups=1, spatial_scale_factor=None, spatial_scale_mode='bilinear',
+                 spectral_pos_encoding=False, use_se=False, se_kwargs=None, ffc3d=False, fft_norm='ortho'):
+        # bn_layer not used
+        super(FourierUnit, self).__init__()
+        self.groups = groups
+
+        self.conv_layer = torch.nn.Conv2d(in_channels=in_channels * 2 + (2 if spectral_pos_encoding else 0),
+                                          out_channels=out_channels * 2,
+                                          kernel_size=1, stride=1, padding=0, groups=self.groups, bias=False)
+        self.bn = torch.nn.BatchNorm2d(out_channels * 2)
+        self.relu = torch.nn.ReLU(inplace=True)
+
+        # squeeze and excitation block
+        self.use_se = use_se
+        if use_se:
+            if se_kwargs is None:
+                se_kwargs = {}
+            self.se = SELayer(self.conv_layer.in_channels, **se_kwargs)
+
+        self.spatial_scale_factor = spatial_scale_factor
+        self.spatial_scale_mode = spatial_scale_mode
+        self.spectral_pos_encoding = spectral_pos_encoding
+        self.ffc3d = ffc3d
+        self.fft_norm = fft_norm
+
+    def forward(self, x):
+        batch = x.shape[0]
+
+        if self.spatial_scale_factor is not None:
+            orig_size = x.shape[-2:]
+            x = F.interpolate(x, scale_factor=self.spatial_scale_factor, mode=self.spatial_scale_mode, align_corners=False)
+
+        r_size = x.size()
+        # (batch, c, h, w/2+1, 2)
+        fft_dim = (-3, -2, -1) if self.ffc3d else (-2, -1)
+        ffted = torch.fft.rfftn(x, dim=fft_dim, norm=self.fft_norm)
+        ffted = torch.stack((ffted.real, ffted.imag), dim=-1)
+        ffted = ffted.permute(0, 1, 4, 2, 3).contiguous()  # (batch, c, 2, h, w/2+1)
+        ffted = ffted.view((batch, -1,) + ffted.size()[3:])
+
+        if self.spectral_pos_encoding:
+            height, width = ffted.shape[-2:]
+            coords_vert = torch.linspace(0, 1, height)[None, None, :, None].expand(batch, 1, height, width).to(ffted)
+            coords_hor = torch.linspace(0, 1, width)[None, None, None, :].expand(batch, 1, height, width).to(ffted)
+            ffted = torch.cat((coords_vert, coords_hor, ffted), dim=1)
+
+        if self.use_se:
+            ffted = self.se(ffted)
+
+        ffted = self.conv_layer(ffted)  # (batch, c*2, h, w/2+1)
+        ffted = self.relu(self.bn(ffted))
+
+        ffted = ffted.view((batch, -1, 2,) + ffted.size()[2:]).permute(
+            0, 1, 3, 4, 2).contiguous()  # (batch,c, t, h, w/2+1, 2)
+        ffted = torch.complex(ffted[..., 0], ffted[..., 1])
+
+        ifft_shape_slice = x.shape[-3:] if self.ffc3d else x.shape[-2:]
+        output = torch.fft.irfftn(ffted, s=ifft_shape_slice, dim=fft_dim, norm=self.fft_norm)
+
+        if self.spatial_scale_factor is not None:
+            output = F.interpolate(output, size=orig_size, mode=self.spatial_scale_mode, align_corners=False)
+
+        return output
+
+
+class SpectralTransform(nn.Module):
+    def __init__(self, in_channels, out_channels, stride=1, groups=1, enable_lfu=True, **fu_kwargs):
+        # bn_layer not used
+        super(SpectralTransform, self).__init__()
+        self.enable_lfu = enable_lfu
+        if stride == 2:
+            self.downsample = nn.AvgPool2d(kernel_size=(2, 2), stride=2)
+        else:
+            self.downsample = nn.Identity()
+
+        self.stride = stride
+        self.conv1 = nn.Sequential(
+            nn.Conv2d(in_channels, out_channels //
+                      2, kernel_size=1, groups=groups, bias=False),
+            nn.BatchNorm2d(out_channels // 2),
+            nn.ReLU(inplace=True)
+        )
+        self.fu = FourierUnit(
+            out_channels // 2, out_channels // 2, groups, **fu_kwargs)
+        if self.enable_lfu:
+            self.lfu = FourierUnit(
+                out_channels // 2, out_channels // 2, groups)
+        self.conv2 = torch.nn.Conv2d(
+            out_channels // 2, out_channels, kernel_size=1, groups=groups, bias=False)
+
+    def forward(self, x):
+        x = self.downsample(x)
+        x = self.conv1(x)
+        output = self.fu(x)
+
+        if self.enable_lfu:
+            n, c, h, w = x.shape
+            split_no = 2
+            split_s = h // split_no
+            xs = torch.cat(torch.split(
+                x[:, :c // 4], split_s, dim=-2), dim=1).contiguous()
+            xs = torch.cat(torch.split(xs, split_s, dim=-1),
+                           dim=1).contiguous()
+            xs = self.lfu(xs)
+            xs = xs.repeat(1, 1, split_no, split_no).contiguous()
+        else:
+            xs = 0
+
+        output = self.conv2(x + output + xs)
+        return output
+
+
+class FFC(nn.Module):
+
+    def __init__(self, in_channels, out_channels, kernel_size,
+                 ratio_gin, ratio_gout, stride=1, padding=0,
+                 dilation=1, groups=1, bias=False, enable_lfu=True,
+                 padding_type='reflect', gated=False, **spectral_kwargs):
+        super(FFC, self).__init__()
+
+        assert stride == 1 or stride == 2, "Stride should be 1 or 2."
+        self.stride = stride
+
+        in_cg = int(in_channels * ratio_gin)
+        in_cl = in_channels - in_cg
+        out_cg = int(out_channels * ratio_gout)
+        out_cl = out_channels - out_cg
+
+        self.ratio_gin = ratio_gin
+        self.ratio_gout = ratio_gout
+        self.global_in_num = in_cg
+
+        module = nn.Identity if in_cl == 0 or out_cl == 0 else nn.Conv2d
+        self.convl2l = module(in_cl, out_cl, kernel_size,
+                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
+        module = nn.Identity if in_cl == 0 or out_cg == 0 else nn.Conv2d
+        self.convl2g = module(in_cl, out_cg, kernel_size,
+                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
+        module = nn.Identity if in_cg == 0 or out_cl == 0 else nn.Conv2d
+        self.convg2l = module(in_cg, out_cl, kernel_size,
+                              stride, padding, dilation, groups, bias, padding_mode=padding_type)
+        module = nn.Identity if in_cg == 0 or out_cg == 0 else SpectralTransform
+        self.convg2g = module(
+            in_cg, out_cg, stride, 1 if groups == 1 else groups // 2, enable_lfu, **spectral_kwargs)
+
+        self.gated = gated
+        module = nn.Identity if in_cg == 0 or out_cl == 0 or not self.gated else nn.Conv2d
+        self.gate = module(in_channels, 2, 1)
+
+    def forward(self, x):
+        x_l, x_g = x if type(x) is tuple else (x, 0)
+        out_xl, out_xg = 0, 0
+
+        if self.gated:
+            total_input_parts = [x_l]
+            if torch.is_tensor(x_g):
+                total_input_parts.append(x_g)
+            total_input = torch.cat(total_input_parts, dim=1)
+
+            gates = torch.sigmoid(self.gate(total_input))
+            g2l_gate, l2g_gate = gates.chunk(2, dim=1)
+        else:
+            g2l_gate, l2g_gate = 1, 1
+
+        if self.ratio_gout != 1:
+            out_xl = self.convl2l(x_l) + self.convg2l(x_g) * g2l_gate
+        if self.ratio_gout != 0:
+            out_xg = self.convl2g(x_l) * l2g_gate + self.convg2g(x_g)
+
+        return out_xl, out_xg

video-tetalking/models/transformer.py

@@ -0,0 +1,119 @@
+import torch
+from torch import nn
+
+from einops import rearrange
+
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+
+class GELU(nn.Module):
+    def __init__(self):
+        super(GELU, self).__init__()
+    def forward(self, x):
+        return 0.5*x*(1+F.tanh(np.sqrt(2/np.pi)*(x+0.044715*torch.pow(x,3))))
+
+# helpers
+
+def pair(t):
+    return t if isinstance(t, tuple) else (t, t)
+
+# classes
+
+class PreNorm(nn.Module):
+    def __init__(self, dim, fn):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim)
+        self.fn = fn
+    def forward(self, x, **kwargs):
+        return self.fn(self.norm(x), **kwargs)
+
+class DualPreNorm(nn.Module):
+    def __init__(self, dim, fn):
+        super().__init__()
+        self.normx = nn.LayerNorm(dim)
+        self.normy = nn.LayerNorm(dim)
+        self.fn = fn
+    def forward(self, x, y, **kwargs):
+        return self.fn(self.normx(x), self.normy(y), **kwargs)
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout = 0.):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout)
+        )
+    def forward(self, x):
+        return self.net(x)
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
+        super().__init__()
+        inner_dim = dim_head *  heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+
+        self.attend = nn.Softmax(dim = -1)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias = False)
+        self.to_k = nn.Linear(dim, inner_dim, bias = False)
+        self.to_v = nn.Linear(dim, inner_dim, bias = False)
+
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim),
+            nn.Dropout(dropout)
+        ) if project_out else nn.Identity()
+
+    def forward(self, x, y):
+        # qk = self.to_qk(x).chunk(2, dim = -1) #
+        q = rearrange(self.to_q(x), 'b n (h d) -> b h n d', h = self.heads) # q,k from the zero feature
+        k = rearrange(self.to_k(x), 'b n (h d) -> b h n d', h = self.heads) # v from the reference features
+        v = rearrange(self.to_v(y), 'b n (h d) -> b h n d', h = self.heads) 
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+class Transformer(nn.Module):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(nn.ModuleList([
+                DualPreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
+                PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
+            ]))
+
+
+    def forward(self, x, y): # x is the cropped, y is the foreign reference
+        bs,c,h,w = x.size()
+
+        # img to embedding
+        x = x.view(bs,c,-1).permute(0,2,1)
+        y = y.view(bs,c,-1).permute(0,2,1)
+
+        for attn, ff in self.layers:
+            x = attn(x, y) + x
+            x = ff(x) + x
+
+        x = x.view(bs,h,w,c).permute(0,3,1,2)
+        return x
+
+class RETURNX(nn.Module):
+    def __init__(self,):
+        super().__init__()
+
+    def forward(self, x, y): # x is the cropped, y is the foreign reference 
+        return x

video-tetalking/predict.py

@@ -0,0 +1,552 @@
+# Prediction interface for Cog ⚙️
+# https://github.com/replicate/cog/blob/main/docs/python.md
+
+import os
+import sys
+import argparse
+import subprocess
+import numpy as np
+from tqdm import tqdm
+from PIL import Image
+from scipy.io import loadmat
+import torch
+import cv2
+from cog import BasePredictor, Input, Path
+
+sys.path.insert(0, "third_part")
+sys.path.insert(0, "third_part/GPEN")
+sys.path.insert(0, "third_part/GFPGAN")
+
+# 3dmm extraction
+from third_part.face3d.util.preprocess import align_img
+from third_part.face3d.util.load_mats import load_lm3d
+from third_part.face3d.extract_kp_videos import KeypointExtractor
+
+# face enhancement
+from third_part.GPEN.gpen_face_enhancer import FaceEnhancement
+from third_part.GFPGAN.gfpgan import GFPGANer
+
+# expression control
+from third_part.ganimation_replicate.model.ganimation import GANimationModel
+
+from utils import audio
+from utils.ffhq_preprocess import Croper
+from utils.alignment_stit import crop_faces, calc_alignment_coefficients, paste_image
+from utils.inference_utils import (
+    Laplacian_Pyramid_Blending_with_mask,
+    face_detect,
+    load_model,
+    options,
+    split_coeff,
+    trans_image,
+    transform_semantic,
+    find_crop_norm_ratio,
+    load_face3d_net,
+    exp_aus_dict,
+)
+
+
+class Predictor(BasePredictor):
+    def setup(self) -> None:
+        """Load the model into memory to make running multiple predictions efficient"""
+        self.enhancer = FaceEnhancement(
+            base_dir="checkpoints",
+            size=512,
+            model="GPEN-BFR-512",
+            use_sr=False,
+            sr_model="rrdb_realesrnet_psnr",
+            channel_multiplier=2,
+            narrow=1,
+            device="cuda",
+        )
+        self.restorer = GFPGANer(
+            model_path="checkpoints/GFPGANv1.3.pth",
+            upscale=1,
+            arch="clean",
+            channel_multiplier=2,
+            bg_upsampler=None,
+        )
+        self.croper = Croper("checkpoints/shape_predictor_68_face_landmarks.dat")
+        self.kp_extractor = KeypointExtractor()
+
+        face3d_net_path = "checkpoints/face3d_pretrain_epoch_20.pth"
+
+        self.net_recon = load_face3d_net(face3d_net_path, "cuda")
+        self.lm3d_std = load_lm3d("checkpoints/BFM")
+
+    def predict(
+        self,
+        face: Path = Input(description="Input video file of a talking-head."),
+        input_audio: Path = Input(description="Input audio file."),
+    ) -> Path:
+        """Run a single prediction on the model"""
+        device = "cuda"
+        args = argparse.Namespace(
+            DNet_path="checkpoints/DNet.pt",
+            LNet_path="checkpoints/LNet.pth",
+            ENet_path="checkpoints/ENet.pth",
+            face3d_net_path="checkpoints/face3d_pretrain_epoch_20.pth",
+            face=str(face),
+            audio=str(input_audio),
+            exp_img="neutral",
+            outfile=None,
+            fps=25,
+            pads=[0, 20, 0, 0],
+            face_det_batch_size=4,
+            LNet_batch_size=16,
+            img_size=384,
+            crop=[0, -1, 0, -1],
+            box=[-1, -1, -1, -1],
+            nosmooth=False,
+            static=False,
+            up_face="original",
+            one_shot=False,
+            without_rl1=False,
+            tmp_dir="temp",
+            re_preprocess=False,
+        )
+
+        base_name = args.face.split("/")[-1]
+
+        if args.face.split(".")[1] in ["jpg", "png", "jpeg"]:
+            full_frames = [cv2.imread(args.face)]
+            args.static = True
+            fps = args.fps
+        else:
+            video_stream = cv2.VideoCapture(args.face)
+            fps = video_stream.get(cv2.CAP_PROP_FPS)
+            full_frames = []
+            while True:
+                still_reading, frame = video_stream.read()
+                if not still_reading:
+                    video_stream.release()
+                    break
+                y1, y2, x1, x2 = args.crop
+                if x2 == -1:
+                    x2 = frame.shape[1]
+                if y2 == -1:
+                    y2 = frame.shape[0]
+                frame = frame[y1:y2, x1:x2]
+                full_frames.append(frame)
+
+        full_frames_RGB = [
+            cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) for frame in full_frames
+        ]
+        full_frames_RGB, crop, quad = self.croper.crop(full_frames_RGB, xsize=512)
+
+        clx, cly, crx, cry = crop
+        lx, ly, rx, ry = quad
+        lx, ly, rx, ry = int(lx), int(ly), int(rx), int(ry)
+        oy1, oy2, ox1, ox2 = (
+            cly + ly,
+            min(cly + ry, full_frames[0].shape[0]),
+            clx + lx,
+            min(clx + rx, full_frames[0].shape[1]),
+        )
+        # original_size = (ox2 - ox1, oy2 - oy1)
+        frames_pil = [
+            Image.fromarray(cv2.resize(frame, (256, 256))) for frame in full_frames_RGB
+        ]
+
+        # get the landmark according to the detected face.
+        if (
+            not os.path.isfile("temp/" + base_name + "_landmarks.txt")
+            or args.re_preprocess
+        ):
+            print("[Step 1] Landmarks Extraction in Video.")
+            lm = self.kp_extractor.extract_keypoint(
+                frames_pil, "./temp/" + base_name + "_landmarks.txt"
+            )
+        else:
+            print("[Step 1] Using saved landmarks.")
+            lm = np.loadtxt("temp/" + base_name + "_landmarks.txt").astype(np.float32)
+            lm = lm.reshape([len(full_frames), -1, 2])
+
+        if (
+            not os.path.isfile("temp/" + base_name + "_coeffs.npy")
+            or args.exp_img is not None
+            or args.re_preprocess
+        ):
+            video_coeffs = []
+            for idx in tqdm(
+                range(len(frames_pil)), desc="[Step 2] 3DMM Extraction In Video:"
+            ):
+                frame = frames_pil[idx]
+                W, H = frame.size
+                lm_idx = lm[idx].reshape([-1, 2])
+                if np.mean(lm_idx) == -1:
+                    lm_idx = (self.lm3d_std[:, :2] + 1) / 2.0
+                    lm_idx = np.concatenate([lm_idx[:, :1] * W, lm_idx[:, 1:2] * H], 1)
+                else:
+                    lm_idx[:, -1] = H - 1 - lm_idx[:, -1]
+
+                trans_params, im_idx, lm_idx, _ = align_img(
+                    frame, lm_idx, self.lm3d_std
+                )
+                trans_params = np.array(
+                    [float(item) for item in np.hsplit(trans_params, 5)]
+                ).astype(np.float32)
+                im_idx_tensor = (
+                    torch.tensor(np.array(im_idx) / 255.0, dtype=torch.float32)
+                    .permute(2, 0, 1)
+                    .to(device)
+                    .unsqueeze(0)
+                )
+                with torch.no_grad():
+                    coeffs = split_coeff(self.net_recon(im_idx_tensor))
+
+                pred_coeff = {key: coeffs[key].cpu().numpy() for key in coeffs}
+                pred_coeff = np.concatenate(
+                    [
+                        pred_coeff["id"],
+                        pred_coeff["exp"],
+                        pred_coeff["tex"],
+                        pred_coeff["angle"],
+                        pred_coeff["gamma"],
+                        pred_coeff["trans"],
+                        trans_params[None],
+                    ],
+                    1,
+                )
+                video_coeffs.append(pred_coeff)
+            semantic_npy = np.array(video_coeffs)[:, 0]
+            np.save("temp/" + base_name + "_coeffs.npy", semantic_npy)
+        else:
+            print("[Step 2] Using saved coeffs.")
+            semantic_npy = np.load("temp/" + base_name + "_coeffs.npy").astype(
+                np.float32
+            )
+
+        # generate the 3dmm coeff from a single image
+        if args.exp_img == "smile":
+            expression = torch.tensor(
+                loadmat("checkpoints/expression.mat")["expression_mouth"]
+            )[0]
+        else:
+            print("using expression center")
+            expression = torch.tensor(
+                loadmat("checkpoints/expression.mat")["expression_center"]
+            )[0]
+
+        # load DNet, model(LNet and ENet)
+        D_Net, model = load_model(args, device)
+
+        if (
+            not os.path.isfile("temp/" + base_name + "_stablized.npy")
+            or args.re_preprocess
+        ):
+            imgs = []
+            for idx in tqdm(
+                range(len(frames_pil)),
+                desc="[Step 3] Stabilize the expression In Video:",
+            ):
+                if args.one_shot:
+                    source_img = trans_image(frames_pil[0]).unsqueeze(0).to(device)
+                    semantic_source_numpy = semantic_npy[0:1]
+                else:
+                    source_img = trans_image(frames_pil[idx]).unsqueeze(0).to(device)
+                    semantic_source_numpy = semantic_npy[idx : idx + 1]
+                ratio = find_crop_norm_ratio(semantic_source_numpy, semantic_npy)
+                coeff = (
+                    transform_semantic(semantic_npy, idx, ratio).unsqueeze(0).to(device)
+                )
+
+                # hacking the new expression
+                coeff[:, :64, :] = expression[None, :64, None].to(device)
+                with torch.no_grad():
+                    output = D_Net(source_img, coeff)
+                img_stablized = np.uint8(
+                    (
+                        output["fake_image"]
+                        .squeeze(0)
+                        .permute(1, 2, 0)
+                        .cpu()
+                        .clamp_(-1, 1)
+                        .numpy()
+                        + 1
+                    )
+                    / 2.0
+                    * 255
+                )
+                imgs.append(cv2.cvtColor(img_stablized, cv2.COLOR_RGB2BGR))
+            np.save("temp/" + base_name + "_stablized.npy", imgs)
+            del D_Net
+        else:
+            print("[Step 3] Using saved stabilized video.")
+            imgs = np.load("temp/" + base_name + "_stablized.npy")
+        torch.cuda.empty_cache()
+
+        if not args.audio.endswith(".wav"):
+            command = "ffmpeg -loglevel error -y -i {} -strict -2 {}".format(
+                args.audio, "temp/{}/temp.wav".format(args.tmp_dir)
+            )
+            subprocess.call(command, shell=True)
+            args.audio = "temp/{}/temp.wav".format(args.tmp_dir)
+        wav = audio.load_wav(args.audio, 16000)
+        mel = audio.melspectrogram(wav)
+        if np.isnan(mel.reshape(-1)).sum() > 0:
+            raise ValueError(
+                "Mel contains nan! Using a TTS voice? Add a small epsilon noise to the wav file and try again"
+            )
+
+        mel_step_size, mel_idx_multiplier, i, mel_chunks = 16, 80.0 / fps, 0, []
+        while True:
+            start_idx = int(i * mel_idx_multiplier)
+            if start_idx + mel_step_size > len(mel[0]):
+                mel_chunks.append(mel[:, len(mel[0]) - mel_step_size :])
+                break
+            mel_chunks.append(mel[:, start_idx : start_idx + mel_step_size])
+            i += 1
+
+        print("[Step 4] Load audio; Length of mel chunks: {}".format(len(mel_chunks)))
+        imgs = imgs[: len(mel_chunks)]
+        full_frames = full_frames[: len(mel_chunks)]
+        lm = lm[: len(mel_chunks)]
+
+        imgs_enhanced = []
+        for idx in tqdm(range(len(imgs)), desc="[Step 5] Reference Enhancement"):
+            img = imgs[idx]
+            pred, _, _ = self.enhancer.process(
+                img, img, face_enhance=True, possion_blending=False
+            )
+            imgs_enhanced.append(pred)
+        gen = datagen(
+            imgs_enhanced.copy(), mel_chunks, full_frames, args, (oy1, oy2, ox1, ox2)
+        )
+
+        frame_h, frame_w = full_frames[0].shape[:-1]
+        out = cv2.VideoWriter(
+            "temp/{}/result.mp4".format(args.tmp_dir),
+            cv2.VideoWriter_fourcc(*"mp4v"),
+            fps,
+            (frame_w, frame_h),
+        )
+
+        if args.up_face != "original":
+            instance = GANimationModel()
+            instance.initialize()
+            instance.setup()
+
+        # kp_extractor = KeypointExtractor()
+        for i, (
+            img_batch,
+            mel_batch,
+            frames,
+            coords,
+            img_original,
+            f_frames,
+        ) in enumerate(
+            tqdm(
+                gen,
+                desc="[Step 6] Lip Synthesis:",
+                total=int(np.ceil(float(len(mel_chunks)) / args.LNet_batch_size)),
+            )
+        ):
+            img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(
+                device
+            )
+            mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(
+                device
+            )
+            img_original = (
+                torch.FloatTensor(np.transpose(img_original, (0, 3, 1, 2))).to(device)
+                / 255.0
+            )  # BGR -> RGB
+
+            with torch.no_grad():
+                incomplete, reference = torch.split(img_batch, 3, dim=1)
+                pred, low_res = model(mel_batch, img_batch, reference)
+                pred = torch.clamp(pred, 0, 1)
+
+                if args.up_face in ["sad", "angry", "surprise"]:
+                    tar_aus = exp_aus_dict[args.up_face]
+                else:
+                    pass
+
+                if args.up_face == "original":
+                    cur_gen_faces = img_original
+                else:
+                    test_batch = {
+                        "src_img": torch.nn.functional.interpolate(
+                            (img_original * 2 - 1), size=(128, 128), mode="bilinear"
+                        ),
+                        "tar_aus": tar_aus.repeat(len(incomplete), 1),
+                    }
+                    instance.feed_batch(test_batch)
+                    instance.forward()
+                    cur_gen_faces = torch.nn.functional.interpolate(
+                        instance.fake_img / 2.0 + 0.5, size=(384, 384), mode="bilinear"
+                    )
+
+                if args.without_rl1 is not False:
+                    incomplete, reference = torch.split(img_batch, 3, dim=1)
+                    mask = torch.where(
+                        incomplete == 0,
+                        torch.ones_like(incomplete),
+                        torch.zeros_like(incomplete),
+                    )
+                    pred = pred * mask + cur_gen_faces * (1 - mask)
+
+            pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.0
+
+            torch.cuda.empty_cache()
+            for p, f, xf, c in zip(pred, frames, f_frames, coords):
+                y1, y2, x1, x2 = c
+                p = cv2.resize(p.astype(np.uint8), (x2 - x1, y2 - y1))
+
+                ff = xf.copy()
+                ff[y1:y2, x1:x2] = p
+
+                # month region enhancement by GFPGAN
+                cropped_faces, restored_faces, restored_img = self.restorer.enhance(
+                    ff, has_aligned=False, only_center_face=True, paste_back=True
+                )
+                # 0,   1,   2,   3,   4,   5,   6,   7,   8,  9, 10,  11,  12,
+                mm = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 255, 255, 255, 0, 0, 0, 0, 0, 0]
+                mouse_mask = np.zeros_like(restored_img)
+                tmp_mask = self.enhancer.faceparser.process(
+                    restored_img[y1:y2, x1:x2], mm
+                )[0]
+                mouse_mask[y1:y2, x1:x2] = (
+                    cv2.resize(tmp_mask, (x2 - x1, y2 - y1))[:, :, np.newaxis] / 255.0
+                )
+
+                height, width = ff.shape[:2]
+                restored_img, ff, full_mask = [
+                    cv2.resize(x, (512, 512))
+                    for x in (restored_img, ff, np.float32(mouse_mask))
+                ]
+                img = Laplacian_Pyramid_Blending_with_mask(
+                    restored_img, ff, full_mask[:, :, 0], 10
+                )
+                pp = np.uint8(cv2.resize(np.clip(img, 0, 255), (width, height)))
+
+                pp, orig_faces, enhanced_faces = self.enhancer.process(
+                    pp, xf, bbox=c, face_enhance=False, possion_blending=True
+                )
+                out.write(pp)
+        out.release()
+
+        output_file = "/tmp/output.mp4"
+        command = "ffmpeg -loglevel error -y -i {} -i {} -strict -2 -q:v 1 {}".format(
+            args.audio, "temp/{}/result.mp4".format(args.tmp_dir), output_file
+        )
+        subprocess.call(command, shell=True)
+
+        return Path(output_file)
+
+
+# frames:256x256, full_frames: original size
+def datagen(frames, mels, full_frames, args, cox):
+    img_batch, mel_batch, frame_batch, coords_batch, ref_batch, full_frame_batch = (
+        [],
+        [],
+        [],
+        [],
+        [],
+        [],
+    )
+    base_name = args.face.split("/")[-1]
+    refs = []
+    image_size = 256
+
+    # original frames
+    kp_extractor = KeypointExtractor()
+    fr_pil = [Image.fromarray(frame) for frame in frames]
+    lms = kp_extractor.extract_keypoint(
+        fr_pil, "temp/" + base_name + "x12_landmarks.txt"
+    )
+    frames_pil = [
+        (lm, frame) for frame, lm in zip(fr_pil, lms)
+    ]  # frames is the croped version of modified face
+    crops, orig_images, quads = crop_faces(
+        image_size, frames_pil, scale=1.0, use_fa=True
+    )
+    inverse_transforms = [
+        calc_alignment_coefficients(
+            quad + 0.5,
+            [[0, 0], [0, image_size], [image_size, image_size], [image_size, 0]],
+        )
+        for quad in quads
+    ]
+    del kp_extractor.detector
+
+    oy1, oy2, ox1, ox2 = cox
+    face_det_results = face_detect(full_frames, args, jaw_correction=True)
+
+    for inverse_transform, crop, full_frame, face_det in zip(
+        inverse_transforms, crops, full_frames, face_det_results
+    ):
+        imc_pil = paste_image(
+            inverse_transform,
+            crop,
+            Image.fromarray(
+                cv2.resize(
+                    full_frame[int(oy1) : int(oy2), int(ox1) : int(ox2)], (256, 256)
+                )
+            ),
+        )
+
+        ff = full_frame.copy()
+        ff[int(oy1) : int(oy2), int(ox1) : int(ox2)] = cv2.resize(
+            np.array(imc_pil.convert("RGB")), (ox2 - ox1, oy2 - oy1)
+        )
+        oface, coords = face_det
+        y1, y2, x1, x2 = coords
+        refs.append(ff[y1:y2, x1:x2])
+
+    for i, m in enumerate(mels):
+        idx = 0 if args.static else i % len(frames)
+        frame_to_save = frames[idx].copy()
+        face = refs[idx]
+        oface, coords = face_det_results[idx].copy()
+
+        face = cv2.resize(face, (args.img_size, args.img_size))
+        oface = cv2.resize(oface, (args.img_size, args.img_size))
+
+        img_batch.append(oface)
+        ref_batch.append(face)
+        mel_batch.append(m)
+        coords_batch.append(coords)
+        frame_batch.append(frame_to_save)
+        full_frame_batch.append(full_frames[idx].copy())
+
+        if len(img_batch) >= args.LNet_batch_size:
+            img_batch, mel_batch, ref_batch = (
+                np.asarray(img_batch),
+                np.asarray(mel_batch),
+                np.asarray(ref_batch),
+            )
+            img_masked = img_batch.copy()
+            img_original = img_batch.copy()
+            img_masked[:, args.img_size // 2 :] = 0
+            img_batch = np.concatenate((img_masked, ref_batch), axis=3) / 255.0
+            mel_batch = np.reshape(
+                mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1]
+            )
+
+            yield img_batch, mel_batch, frame_batch, coords_batch, img_original, full_frame_batch
+            (
+                img_batch,
+                mel_batch,
+                frame_batch,
+                coords_batch,
+                img_original,
+                full_frame_batch,
+                ref_batch,
+            ) = ([], [], [], [], [], [], [])
+
+    if len(img_batch) > 0:
+        img_batch, mel_batch, ref_batch = (
+            np.asarray(img_batch),
+            np.asarray(mel_batch),
+            np.asarray(ref_batch),
+        )
+        img_masked = img_batch.copy()
+        img_original = img_batch.copy()
+        img_masked[:, args.img_size // 2 :] = 0
+        img_batch = np.concatenate((img_masked, ref_batch), axis=3) / 255.0
+        mel_batch = np.reshape(
+            mel_batch, [len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1]
+        )
+        yield img_batch, mel_batch, frame_batch, coords_batch, img_original, full_frame_batch

video-tetalking/quick_demo.ipynb

@@ -0,0 +1,293 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 0,
+  "metadata": {
+    "colab": {
+      "provenance": [],
+      "authorship_tag": "ABX9TyMPin07iIA2oewCCP9ZTz6w",
+      "include_colab_link": true
+    },
+    "kernelspec": {
+      "name": "python3",
+      "display_name": "Python 3"
+    },
+    "language_info": {
+      "name": "python"
+    },
+    "accelerator": "GPU",
+    "gpuClass": "standard"
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "view-in-github",
+        "colab_type": "text"
+      },
+      "source": [
+        "<a href=\"https://colab.research.google.com/github/vinthony/video-retalking/blob/main/quick_demo.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "## VideoReTalking：Audio-based Lip Synchronization for Talking Head Video Editing In the Wild\n",
+        "\n",
+        "[Arxiv](https://arxiv.org/abs/2211.14758) | [Project](https://vinthony.github.io/video-retalking/) | [Github](https://github.com/vinthony/video-retalking)\n",
+        "\n",
+        "Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, Nannan Wang\n",
+        "\n",
+        "Xidian University, Tencent AI Lab, Tsinghua University\n",
+        "\n",
+        "*SIGGRAPH Asia 2022 Conferenence Track*\n",
+        "\n"
+      ],
+      "metadata": {
+        "id": "NVfkv2BXSpr3"
+      }
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "**Installation** (30s)"
+      ],
+      "metadata": {
+        "id": "u9hdPaH6UL_F"
+      }
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "PnKT9goiQ3Hk"
+      },
+      "outputs": [],
+      "source": [
+        "#@title\n",
+        "### make sure that CUDA is available in Edit -> Nootbook settings -> GPU\n",
+        "!nvidia-smi\n",
+        "\n",
+        "!python --version  \n",
+        "!apt-get update\n",
+        "!apt install ffmpeg &> /dev/null \n",
+        "\n",
+        "print('Git clone project and install requirements...')\n",
+        "!git clone https://github.com/vinthony/video-retalking.git &> /dev/null\n",
+        "%cd video-retalking\n",
+        "# !pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 -f https://download.pytorch.org/whl/torch_stable.html\n",
+        "!pip install -r requirements.txt"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "**Download Pretrained Models**"
+      ],
+      "metadata": {
+        "id": "uwJS0eaM61Cq"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title\n",
+        "print('Download pre-trained models...')\n",
+        "!mkdir ./checkpoints  \n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/30_net_gen.pth -O ./checkpoints/30_net_gen.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/BFM.zip -O ./checkpoints/BFM.zip\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/DNet.pt -O ./checkpoints/DNet.pt\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/ENet.pth -O ./checkpoints/ENet.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/expression.mat -O ./checkpoints/expression.mat\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/face3d_pretrain_epoch_20.pth -O ./checkpoints/face3d_pretrain_epoch_20.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/GFPGANv1.3.pth -O ./checkpoints/GFPGANv1.3.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/GPEN-BFR-512.pth -O ./checkpoints/GPEN-BFR-512.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/LNet.pth -O ./checkpoints/LNet.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/ParseNet-latest.pth -O ./checkpoints/ParseNet-latest.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/RetinaFace-R50.pth -O ./checkpoints/RetinaFace-R50.pth\n",
+        "!wget https://github.com/vinthony/video-retalking/releases/download/v0.0.1/shape_predictor_68_face_landmarks.dat -O ./checkpoints/shape_predictor_68_face_landmarks.dat\n",
+        "!unzip -d ./checkpoints/BFM ./checkpoints/BFM.zip"
+      ],
+      "metadata": {
+        "id": "x18qYuQY678E"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "**Inference**\n",
+        "\n",
+        "`--face`: Input video.\n",
+        "\n",
+        "`--audio`: Input audio. Both *.wav* and *.mp4* files are supported.\n",
+        "\n",
+        "You can choose our provided data from `./examples` folder or upload from your local computer.\n",
+        "\n",
+        "\n",
+        "\n",
+        "\n"
+      ],
+      "metadata": {
+        "id": "QJRTF4U8UOjv"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title\n",
+        "import glob, os, sys\n",
+        "import ipywidgets as widgets\n",
+        "from IPython.display import HTML\n",
+        "from base64 import b64encode\n",
+        "\n",
+        "print(\"Choose the Video name to edit: (saved in folder 'examples/face')\")\n",
+        "vid_list = glob.glob1('examples/face/', '*.mp4')\n",
+        "vid_list.sort()\n",
+        "default_vid_name = widgets.Dropdown(options=vid_list, value='1.mp4')\n",
+        "display(default_vid_name)\n",
+        "\n",
+        "print(\"Choose the Audio name to edit: (saved in folder 'examples/audio')\")\n",
+        "audio_list = glob.glob1('examples/audio/', '*')\n",
+        "audio_list.sort()\n",
+        "default_audio_name = widgets.Dropdown(options=audio_list, value='1.wav')\n",
+        "display(default_audio_name)\n"
+      ],
+      "metadata": {
+        "id": "U-IY-cBSporP",
+        "cellView": "form"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "Visualize the input video and audio:"
+      ],
+      "metadata": {
+        "id": "-MtI_R1bLJ-f"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title\n",
+        "input_video_name = './examples/face/{}'.format(default_vid_name.value)\n",
+        "input_video_mp4 = open('{}'.format(input_video_name),'rb').read()\n",
+        "input_video_data_url = \"data:video/x-m4v;base64,\" + b64encode(input_video_mp4).decode()\n",
+        "print('Display input video: {}'.format(input_video_name), file=sys.stderr)\n",
+        "display(HTML(\"\"\"\n",
+        "  <video width=400 controls>\n",
+        "        <source src=\"%s\" type=\"video/mp4\">\n",
+        "  </video>\n",
+        "  \"\"\" % input_video_data_url))\n",
+        "\n",
+        "input_audio_name = './examples/audio/{}'.format(default_audio_name.value)\n",
+        "input_audio_mp4 = open('{}'.format(input_audio_name),'rb').read()\n",
+        "input_audio_data_url = \"data:audio/wav;base64,\" + b64encode(input_audio_mp4).decode()\n",
+        "print('Display input audio: {}'.format(input_audio_name), file=sys.stderr)\n",
+        "display(HTML(\"\"\"\n",
+        "  <audio width=400 controls>\n",
+        "        <source src=\"%s\" type=\"audio/wav\">\n",
+        "  </audio>\n",
+        "  \"\"\" % input_audio_data_url))\n"
+      ],
+      "metadata": {
+        "id": "ljbScdofJyGO",
+        "cellView": "form"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "input_video_path = 'examples/face/{}'.format(default_vid_name.value)\n",
+        "input_audio_path = 'examples/audio/{}'.format(default_audio_name.value)\n",
+        "\n",
+        "!python3 inference.py \\\n",
+        "  --face {input_video_path} \\\n",
+        "  --audio {input_audio_path} \\\n",
+        "  --outfile results/output.mp4"
+      ],
+      "metadata": {
+        "id": "D7hUwRCyUYEA"
+      },
+      "execution_count": null,
+      "outputs": []
+    },
+    {
+      "cell_type": "markdown",
+      "source": [
+        "Visualize the output video:"
+      ],
+      "metadata": {
+        "id": "JB5RbKc-njkB"
+      }
+    },
+    {
+      "cell_type": "code",
+      "source": [
+        "#@title\n",
+        "# visualize code from makeittalk\n",
+        "from IPython.display import HTML\n",
+        "from base64 import b64encode\n",
+        "import os, sys, glob, cv2, subprocess, platform\n",
+        "\n",
+        "def read_video(vid_name):\n",
+        "  video_stream = cv2.VideoCapture(vid_name)\n",
+        "  fps = video_stream.get(cv2.CAP_PROP_FPS)\n",
+        "  full_frames = []\n",
+        "  while True:\n",
+        "    still_reading, frame = video_stream.read()\n",
+        "    if not still_reading:\n",
+        "        video_stream.release()\n",
+        "        break\n",
+        "    full_frames.append(frame)\n",
+        "  return full_frames, fps\n",
+        "\n",
+        "input_video_frames, fps = read_video(input_video_path)\n",
+        "output_video_frames, _ = read_video('./results/output.mp4')\n",
+        "\n",
+        "frame_h, frame_w = input_video_frames[0].shape[:-1]\n",
+        "out_concat = cv2.VideoWriter('./temp/temp/result_concat.mp4', cv2.VideoWriter_fourcc(*'mp4v'), fps, (frame_w*2, frame_h))\n",
+        "for i in range(len(output_video_frames)):\n",
+        "  frame_input = input_video_frames[i % len(input_video_frames)]\n",
+        "  frame_output = output_video_frames[i]\n",
+        "  out_concat.write(cv2.hconcat([frame_input, frame_output]))\n",
+        "out_concat.release()\n",
+        "\n",
+        "command = 'ffmpeg -loglevel error -y -i {} -i {} -strict -2 -q:v 1 {}'.format(input_audio_path, './temp/temp/result_concat.mp4', './results/output_concat_input.mp4')\n",
+        "subprocess.call(command, shell=platform.system() != 'Windows')\n",
+        "\n",
+        "\n",
+        "output_video_name = './results/output.mp4'\n",
+        "output_video_mp4 = open('{}'.format(output_video_name),'rb').read()\n",
+        "output_video_data_url = \"data:video/mp4;base64,\" + b64encode(output_video_mp4).decode()\n",
+        "print('Display lip-syncing video: {}'.format(output_video_name), file=sys.stderr)\n",
+        "display(HTML(\"\"\"\n",
+        "  <video height=400 controls>\n",
+        "        <source src=\"%s\" type=\"video/mp4\">\n",
+        "  </video>\n",
+        "  \"\"\" % output_video_data_url))\n",
+        "\n",
+        "output_concat_video_name = './results/output_concat_input.mp4'\n",
+        "output_concat_video_mp4 = open('{}'.format(output_concat_video_name),'rb').read()\n",
+        "output_concat_video_data_url = \"data:video/mp4;base64,\" + b64encode(output_concat_video_mp4).decode()\n",
+        "print('Display input video and lip-syncing video: {}'.format(output_concat_video_name), file=sys.stderr)\n",
+        "display(HTML(\"\"\"\n",
+        "  <video height=400 controls>\n",
+        "        <source src=\"%s\" type=\"video/mp4\">\n",
+        "  </video>\n",
+        "  \"\"\" % output_concat_video_data_url))\n"
+      ],
+      "metadata": {
+        "id": "ravs9UDucMfy",
+        "cellView": "form"
+      },
+      "execution_count": null,
+      "outputs": []
+    }
+  ]
+}

video-tetalking/requirements.txt

@@ -0,0 +1,10 @@
+basicsr==1.4.2
+kornia==0.5.1
+face-alignment==1.3.4
+ninja==1.10.2.3
+einops==0.4.1
+facexlib==0.2.5
+librosa==0.9.2
+dlib==19.24.0
+gradio>=3.7.0
+numpy==1.23.4

video-tetalking/third_part/GFPGAN/LICENSE

@@ -0,0 +1,351 @@
+Tencent is pleased to support the open source community by making GFPGAN available.
+
+Copyright (C) 2021 THL A29 Limited, a Tencent company.  All rights reserved.
+
+GFPGAN is licensed under the Apache License Version 2.0 except for the third-party components listed below.
+
+
+Terms of the Apache License Version 2.0:
+---------------------------------------------
+Apache License
+
+Version 2.0, January 2004
+
+http://www.apache.org/licenses/
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+1. Definitions.
+
+“License” shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.
+
+“Licensor” shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
+
+“Legal Entity” shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, “control” means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
+
+“You” (or “Your”) shall mean an individual or Legal Entity exercising permissions granted by this License.
+
+“Source” form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
+
+“Object” form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+
+“Work” shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).
+
+“Derivative Works” shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
+
+“Contribution” shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, “submitted” means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as “Not a Contribution.”
+
+“Contributor” shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
+
+3. Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
+
+4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
+
+You must give any other recipients of the Work or Derivative Works a copy of this License; and
+
+You must cause any modified files to carry prominent notices stating that You changed the files; and
+
+You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
+
+If the Work includes a “NOTICE” text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
+
+You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
+
+5. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
+
+6. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
+
+7. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
+
+8. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+
+9. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS
+
+
+
+Other  dependencies and licenses:
+
+
+Open Source Software licensed under the Apache 2.0 license and Other Licenses of the Third-Party Components therein:
+---------------------------------------------
+1. basicsr
+Copyright 2018-2020 BasicSR Authors
+
+
+This BasicSR project is released under the Apache 2.0 license.
+
+A copy of Apache 2.0 is included in this file.
+
+StyleGAN2
+The codes are modified from the repository stylegan2-pytorch. Many thanks to the author - Kim Seonghyeon 😊 for translating from the official TensorFlow codes to PyTorch ones. Here is the license of stylegan2-pytorch.
+The official repository is https://github.com/NVlabs/stylegan2, and here is the NVIDIA license.
+DFDNet
+The codes are largely modified from the repository DFDNet. Their license is Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.
+
+Terms of the Nvidia License:
+---------------------------------------------
+
+1. Definitions
+
+"Licensor" means any person or entity that distributes its Work.
+
+"Software" means the original work of authorship made available under
+this License.
+
+"Work" means the Software and any additions to or derivative works of
+the Software that are made available under this License.
+
+"Nvidia Processors" means any central processing unit (CPU), graphics
+processing unit (GPU), field-programmable gate array (FPGA),
+application-specific integrated circuit (ASIC) or any combination
+thereof designed, made, sold, or provided by Nvidia or its affiliates.
+
+The terms "reproduce," "reproduction," "derivative works," and
+"distribution" have the meaning as provided under U.S. copyright law;
+provided, however, that for the purposes of this License, derivative
+works shall not include works that remain separable from, or merely
+link (or bind by name) to the interfaces of, the Work.
+
+Works, including the Software, are "made available" under this License
+by including in or with the Work either (a) a copyright notice
+referencing the applicability of this License to the Work, or (b) a
+copy of this License.
+
+2. License Grants
+
+    2.1 Copyright Grant. Subject to the terms and conditions of this
+    License, each Licensor grants to you a perpetual, worldwide,
+    non-exclusive, royalty-free, copyright license to reproduce,
+    prepare derivative works of, publicly display, publicly perform,
+    sublicense and distribute its Work and any resulting derivative
+    works in any form.
+
+3. Limitations
+
+    3.1 Redistribution. You may reproduce or distribute the Work only
+    if (a) you do so under this License, (b) you include a complete
+    copy of this License with your distribution, and (c) you retain
+    without modification any copyright, patent, trademark, or
+    attribution notices that are present in the Work.
+
+    3.2 Derivative Works. You may specify that additional or different
+    terms apply to the use, reproduction, and distribution of your
+    derivative works of the Work ("Your Terms") only if (a) Your Terms
+    provide that the use limitation in Section 3.3 applies to your
+    derivative works, and (b) you identify the specific derivative
+    works that are subject to Your Terms. Notwithstanding Your Terms,
+    this License (including the redistribution requirements in Section
+    3.1) will continue to apply to the Work itself.
+
+    3.3 Use Limitation. The Work and any derivative works thereof only
+    may be used or intended for use non-commercially. The Work or
+    derivative works thereof may be used or intended for use by Nvidia
+    or its affiliates commercially or non-commercially. As used herein,
+    "non-commercially" means for research or evaluation purposes only.
+
+    3.4 Patent Claims. If you bring or threaten to bring a patent claim
+    against any Licensor (including any claim, cross-claim or
+    counterclaim in a lawsuit) to enforce any patents that you allege
+    are infringed by any Work, then your rights under this License from
+    such Licensor (including the grants in Sections 2.1 and 2.2) will
+    terminate immediately.
+
+    3.5 Trademarks. This License does not grant any rights to use any
+    Licensor's or its affiliates' names, logos, or trademarks, except
+    as necessary to reproduce the notices described in this License.
+
+    3.6 Termination. If you violate any term of this License, then your
+    rights under this License (including the grants in Sections 2.1 and
+    2.2) will terminate immediately.
+
+4. Disclaimer of Warranty.
+
+THE WORK IS PROVIDED "AS IS" WITHOUT WARRANTIES OR CONDITIONS OF ANY
+KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WARRANTIES OR CONDITIONS OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE OR
+NON-INFRINGEMENT. YOU BEAR THE RISK OF UNDERTAKING ANY ACTIVITIES UNDER
+THIS LICENSE.
+
+5. Limitation of Liability.
+
+EXCEPT AS PROHIBITED BY APPLICABLE LAW, IN NO EVENT AND UNDER NO LEGAL
+THEORY, WHETHER IN TORT (INCLUDING NEGLIGENCE), CONTRACT, OR OTHERWISE
+SHALL ANY LICENSOR BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY DIRECT,
+INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
+OR RELATED TO THIS LICENSE, THE USE OR INABILITY TO USE THE WORK
+(INCLUDING BUT NOT LIMITED TO LOSS OF GOODWILL, BUSINESS INTERRUPTION,
+LOST PROFITS OR DATA, COMPUTER FAILURE OR MALFUNCTION, OR ANY OTHER
+COMMERCIAL DAMAGES OR LOSSES), EVEN IF THE LICENSOR HAS BEEN ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGES.
+
+MIT License
+
+Copyright (c) 2019 Kim Seonghyeon
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+
+
+Open Source Software licensed under the BSD 3-Clause license:
+---------------------------------------------
+1. torchvision
+Copyright (c) Soumith Chintala 2016,
+All rights reserved.
+
+2. torch
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+
+
+Terms of the BSD 3-Clause License:
+---------------------------------------------
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+
+Open Source Software licensed under the BSD 3-Clause License and Other Licenses of the Third-Party Components therein:
+---------------------------------------------
+1. numpy
+Copyright (c) 2005-2020, NumPy Developers.
+All rights reserved.
+
+A copy of BSD 3-Clause License is included in this file.
+
+The NumPy repository and source distributions bundle several libraries that are
+compatibly licensed.  We list these here.
+
+Name: Numpydoc
+Files: doc/sphinxext/numpydoc/*
+License: BSD-2-Clause
+  For details, see doc/sphinxext/LICENSE.txt
+
+Name: scipy-sphinx-theme
+Files: doc/scipy-sphinx-theme/*
+License: BSD-3-Clause AND PSF-2.0 AND Apache-2.0
+  For details, see doc/scipy-sphinx-theme/LICENSE.txt
+
+Name: lapack-lite
+Files: numpy/linalg/lapack_lite/*
+License: BSD-3-Clause
+  For details, see numpy/linalg/lapack_lite/LICENSE.txt
+
+Name: tempita
+Files: tools/npy_tempita/*
+License: MIT
+  For details, see tools/npy_tempita/license.txt
+
+Name: dragon4
+Files: numpy/core/src/multiarray/dragon4.c
+License: MIT
+  For license text, see numpy/core/src/multiarray/dragon4.c
+
+
+
+Open Source Software licensed under the MIT license:
+---------------------------------------------
+1. facexlib
+Copyright (c) 2020 Xintao Wang
+
+2. opencv-python
+Copyright (c) Olli-Pekka Heinisuo
+Please note that only files in cv2 package are used.
+
+
+Terms of the MIT License:
+---------------------------------------------
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+
+
+Open Source Software licensed under the MIT license and Other Licenses of the Third-Party Components therein:
+---------------------------------------------
+1. tqdm
+Copyright (c) 2013 noamraph
+
+`tqdm` is a product of collaborative work.
+Unless otherwise stated, all authors (see commit logs) retain copyright
+for their respective work, and release the work under the MIT licence
+(text below).
+
+Exceptions or notable authors are listed below
+in reverse chronological order:
+
+* files: *
+  MPLv2.0 2015-2020 (c) Casper da Costa-Luis
+  [casperdcl](https://github.com/casperdcl).
+* files: tqdm/_tqdm.py
+  MIT 2016 (c) [PR #96] on behalf of Google Inc.
+* files: tqdm/_tqdm.py setup.py README.rst MANIFEST.in .gitignore
+  MIT 2013 (c) Noam Yorav-Raphael, original author.
+
+[PR #96]: https://github.com/tqdm/tqdm/pull/96
+
+
+Mozilla Public Licence (MPL) v. 2.0 - Exhibit A
+-----------------------------------------------
+
+This Source Code Form is subject to the terms of the
+Mozilla Public License, v. 2.0.
+If a copy of the MPL was not distributed with this file,
+You can obtain one at https://mozilla.org/MPL/2.0/.
+
+
+MIT License (MIT)
+-----------------
+
+Copyright (c) 2013 noamraph
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+the Software, and to permit persons to whom the Software is furnished to do so,
+subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

video-tetalking/third_part/GFPGAN/gfpgan/__init__.py

@@ -0,0 +1,8 @@
+# flake8: noqa
+
+from .archs import *
+from .data import *
+from .models import *
+from .utils import *
+
+# from .version import *

video-tetalking/third_part/GFPGAN/gfpgan/archs/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import arch modules for registry
+# scan all the files that end with '_arch.py' under the archs folder
+arch_folder = osp.dirname(osp.abspath(__file__))
+arch_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(arch_folder) if v.endswith('_arch.py')]
+# import all the arch modules
+_arch_modules = [importlib.import_module(f'gfpgan.archs.{file_name}') for file_name in arch_filenames]

video-tetalking/third_part/GFPGAN/gfpgan/archs/arcface_arch.py

@@ -0,0 +1,245 @@
+import torch.nn as nn
+from basicsr.utils.registry import ARCH_REGISTRY
+
+
+def conv3x3(inplanes, outplanes, stride=1):
+    """A simple wrapper for 3x3 convolution with padding.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        outplanes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+    """
+    return nn.Conv2d(inplanes, outplanes, kernel_size=3, stride=stride, padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    """Basic residual block used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+    """
+    expansion = 1  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class IRBlock(nn.Module):
+    """Improved residual block (IR Block) used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+        use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
+    """
+    expansion = 1  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True):
+        super(IRBlock, self).__init__()
+        self.bn0 = nn.BatchNorm2d(inplanes)
+        self.conv1 = conv3x3(inplanes, inplanes)
+        self.bn1 = nn.BatchNorm2d(inplanes)
+        self.prelu = nn.PReLU()
+        self.conv2 = conv3x3(inplanes, planes, stride)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+        self.use_se = use_se
+        if self.use_se:
+            self.se = SEBlock(planes)
+
+    def forward(self, x):
+        residual = x
+        out = self.bn0(x)
+        out = self.conv1(out)
+        out = self.bn1(out)
+        out = self.prelu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        if self.use_se:
+            out = self.se(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.prelu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    """Bottleneck block used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+    """
+    expansion = 4  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class SEBlock(nn.Module):
+    """The squeeze-and-excitation block (SEBlock) used in the IRBlock.
+
+    Args:
+        channel (int): Channel number of inputs.
+        reduction (int): Channel reduction ration. Default: 16.
+    """
+
+    def __init__(self, channel, reduction=16):
+        super(SEBlock, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)  # pool to 1x1 without spatial information
+        self.fc = nn.Sequential(
+            nn.Linear(channel, channel // reduction), nn.PReLU(), nn.Linear(channel // reduction, channel),
+            nn.Sigmoid())
+
+    def forward(self, x):
+        b, c, _, _ = x.size()
+        y = self.avg_pool(x).view(b, c)
+        y = self.fc(y).view(b, c, 1, 1)
+        return x * y
+
+
+@ARCH_REGISTRY.register()
+class ResNetArcFace(nn.Module):
+    """ArcFace with ResNet architectures.
+
+    Ref: ArcFace: Additive Angular Margin Loss for Deep Face Recognition.
+
+    Args:
+        block (str): Block used in the ArcFace architecture.
+        layers (tuple(int)): Block numbers in each layer.
+        use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
+    """
+
+    def __init__(self, block, layers, use_se=True):
+        if block == 'IRBlock':
+            block = IRBlock
+        self.inplanes = 64
+        self.use_se = use_se
+        super(ResNetArcFace, self).__init__()
+
+        self.conv1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.prelu = nn.PReLU()
+        self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+        self.bn4 = nn.BatchNorm2d(512)
+        self.dropout = nn.Dropout()
+        self.fc5 = nn.Linear(512 * 8 * 8, 512)
+        self.bn5 = nn.BatchNorm1d(512)
+
+        # initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.xavier_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.xavier_normal_(m.weight)
+                nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, planes, num_blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample, use_se=self.use_se))
+        self.inplanes = planes
+        for _ in range(1, num_blocks):
+            layers.append(block(self.inplanes, planes, use_se=self.use_se))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.prelu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.bn4(x)
+        x = self.dropout(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc5(x)
+        x = self.bn5(x)
+
+        return x

video-tetalking/third_part/GFPGAN/gfpgan/archs/gfpgan_bilinear_arch.py

@@ -0,0 +1,312 @@
+import math
+import random
+import torch
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+
+from .gfpganv1_arch import ResUpBlock
+from .stylegan2_bilinear_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
+                                      StyleGAN2GeneratorBilinear)
+
+
+class StyleGAN2GeneratorBilinearSFT(StyleGAN2GeneratorBilinear):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    It is the bilinear version. It does not use the complicated UpFirDnSmooth function that is not friendly for
+    deployment. It can be easily converted to the clean version: StyleGAN2GeneratorCSFT.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 lr_mlp=0.01,
+                 narrow=1,
+                 sft_half=False):
+        super(StyleGAN2GeneratorBilinearSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            lr_mlp=lr_mlp,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorBilinearSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+@ARCH_REGISTRY.register()
+class GFPGANBilinear(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    It is the bilinear version and it does not use the complicated UpFirDnSmooth function that is not friendly for
+    deployment. It can be easily converted to the clean version: GFPGANv1Clean.
+
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            lr_mlp=0.01,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANBilinear, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = EqualLinear(
+            channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorBilinearSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            lr_mlp=lr_mlp,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True):
+        """Forward function for GFPGANBilinear.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = self.conv_body_first(x)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+
+        feat = self.final_conv(feat)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs

video-tetalking/third_part/GFPGAN/gfpgan/archs/gfpganv1_arch.py

@@ -0,0 +1,439 @@
+import math
+import random
+import torch
+from basicsr.archs.stylegan2_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
+                                          StyleGAN2Generator)
+from basicsr.ops.fused_act import FusedLeakyReLU
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class StyleGAN2GeneratorSFT(StyleGAN2Generator):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
+            applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 resample_kernel=(1, 3, 3, 1),
+                 lr_mlp=0.01,
+                 narrow=1,
+                 sft_half=False):
+        super(StyleGAN2GeneratorSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            resample_kernel=resample_kernel,
+            lr_mlp=lr_mlp,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ConvUpLayer(nn.Module):
+    """Convolutional upsampling layer. It uses bilinear upsampler + Conv.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        stride (int): Stride of the convolution. Default: 1
+        padding (int): Zero-padding added to both sides of the input. Default: 0.
+        bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+        activate (bool): Whether use activateion. Default: True.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 bias=True,
+                 bias_init_val=0,
+                 activate=True):
+        super(ConvUpLayer, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        # self.scale is used to scale the convolution weights, which is related to the common initializations.
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
+
+        if bias and not activate:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+        # activation
+        if activate:
+            if bias:
+                self.activation = FusedLeakyReLU(out_channels)
+            else:
+                self.activation = ScaledLeakyReLU(0.2)
+        else:
+            self.activation = None
+
+    def forward(self, x):
+        # bilinear upsample
+        out = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
+        # conv
+        out = F.conv2d(
+            out,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+        # activation
+        if self.activation is not None:
+            out = self.activation(out)
+        return out
+
+
+class ResUpBlock(nn.Module):
+    """Residual block with upsampling.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+    """
+
+    def __init__(self, in_channels, out_channels):
+        super(ResUpBlock, self).__init__()
+
+        self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
+        self.conv2 = ConvUpLayer(in_channels, out_channels, 3, stride=1, padding=1, bias=True, activate=True)
+        self.skip = ConvUpLayer(in_channels, out_channels, 1, bias=False, activate=False)
+
+    def forward(self, x):
+        out = self.conv1(x)
+        out = self.conv2(out)
+        skip = self.skip(x)
+        out = (out + skip) / math.sqrt(2)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class GFPGANv1(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
+            applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            resample_kernel=(1, 3, 3, 1),
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            lr_mlp=0.01,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANv1, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels, resample_kernel))
+            in_channels = out_channels
+
+        self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = EqualLinear(
+            channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            resample_kernel=resample_kernel,
+            lr_mlp=lr_mlp,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True):
+        """Forward function for GFPGANv1.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = self.conv_body_first(x)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+
+        feat = self.final_conv(feat)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs
+
+
+@ARCH_REGISTRY.register()
+class FacialComponentDiscriminator(nn.Module):
+    """Facial component (eyes, mouth, noise) discriminator used in GFPGAN.
+    """
+
+    def __init__(self):
+        super(FacialComponentDiscriminator, self).__init__()
+        # It now uses a VGG-style architectrue with fixed model size
+        self.conv1 = ConvLayer(3, 64, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv2 = ConvLayer(64, 128, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv3 = ConvLayer(128, 128, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv4 = ConvLayer(128, 256, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv5 = ConvLayer(256, 256, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.final_conv = ConvLayer(256, 1, 3, bias=True, activate=False)
+
+    def forward(self, x, return_feats=False):
+        """Forward function for FacialComponentDiscriminator.
+
+        Args:
+            x (Tensor): Input images.
+            return_feats (bool): Whether to return intermediate features. Default: False.
+        """
+        feat = self.conv1(x)
+        feat = self.conv3(self.conv2(feat))
+        rlt_feats = []
+        if return_feats:
+            rlt_feats.append(feat.clone())
+        feat = self.conv5(self.conv4(feat))
+        if return_feats:
+            rlt_feats.append(feat.clone())
+        out = self.final_conv(feat)
+
+        if return_feats:
+            return out, rlt_feats
+        else:
+            return out, None

video-tetalking/third_part/GFPGAN/gfpgan/archs/gfpganv1_clean_arch.py

@@ -0,0 +1,324 @@
+import math
+import random
+import torch
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+from .stylegan2_clean_arch import StyleGAN2GeneratorClean
+
+
+class StyleGAN2GeneratorCSFT(StyleGAN2GeneratorClean):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1, sft_half=False):
+        super(StyleGAN2GeneratorCSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorCSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ResBlock(nn.Module):
+    """Residual block with bilinear upsampling/downsampling.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        mode (str): Upsampling/downsampling mode. Options: down | up. Default: down.
+    """
+
+    def __init__(self, in_channels, out_channels, mode='down'):
+        super(ResBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
+        self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
+        self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
+        if mode == 'down':
+            self.scale_factor = 0.5
+        elif mode == 'up':
+            self.scale_factor = 2
+
+    def forward(self, x):
+        out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
+        # upsample/downsample
+        out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
+        # skip
+        x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        skip = self.skip(x)
+        out = out + skip
+        return out
+
+
+@ARCH_REGISTRY.register()
+class GFPGANv1Clean(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANv1Clean, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = nn.Conv2d(3, channels[f'{first_out_size}'], 1)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels, mode='down'))
+            in_channels = out_channels
+
+        self.final_conv = nn.Conv2d(in_channels, channels['4'], 3, 1, 1)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResBlock(in_channels, out_channels, mode='up'))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(nn.Conv2d(channels[f'{2**i}'], 3, 1))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = nn.Linear(channels['4'] * 4 * 4, linear_out_channel)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorCSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
+                    nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
+                    nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True):
+        """Forward function for GFPGANv1Clean.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = F.leaky_relu_(self.conv_body_first(x), negative_slope=0.2)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+        feat = F.leaky_relu_(self.final_conv(feat), negative_slope=0.2)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs

video-tetalking/third_part/GFPGAN/gfpgan/archs/stylegan2_bilinear_arch.py

@@ -0,0 +1,613 @@
+import math
+import random
+import torch
+from basicsr.ops.fused_act import FusedLeakyReLU, fused_leaky_relu
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class NormStyleCode(nn.Module):
+
+    def forward(self, x):
+        """Normalize the style codes.
+
+        Args:
+            x (Tensor): Style codes with shape (b, c).
+
+        Returns:
+            Tensor: Normalized tensor.
+        """
+        return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
+
+
+class EqualLinear(nn.Module):
+    """Equalized Linear as StyleGAN2.
+
+    Args:
+        in_channels (int): Size of each sample.
+        out_channels (int): Size of each output sample.
+        bias (bool): If set to ``False``, the layer will not learn an additive
+            bias. Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+        lr_mul (float): Learning rate multiplier. Default: 1.
+        activation (None | str): The activation after ``linear`` operation.
+            Supported: 'fused_lrelu', None. Default: None.
+    """
+
+    def __init__(self, in_channels, out_channels, bias=True, bias_init_val=0, lr_mul=1, activation=None):
+        super(EqualLinear, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.lr_mul = lr_mul
+        self.activation = activation
+        if self.activation not in ['fused_lrelu', None]:
+            raise ValueError(f'Wrong activation value in EqualLinear: {activation}'
+                             "Supported ones are: ['fused_lrelu', None].")
+        self.scale = (1 / math.sqrt(in_channels)) * lr_mul
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+    def forward(self, x):
+        if self.bias is None:
+            bias = None
+        else:
+            bias = self.bias * self.lr_mul
+        if self.activation == 'fused_lrelu':
+            out = F.linear(x, self.weight * self.scale)
+            out = fused_leaky_relu(out, bias)
+        else:
+            out = F.linear(x, self.weight * self.scale, bias=bias)
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, bias={self.bias is not None})')
+
+
+class ModulatedConv2d(nn.Module):
+    """Modulated Conv2d used in StyleGAN2.
+
+    There is no bias in ModulatedConv2d.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether to demodulate in the conv layer.
+            Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
+            Default: None.
+        eps (float): A value added to the denominator for numerical stability.
+            Default: 1e-8.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 eps=1e-8,
+                 interpolation_mode='bilinear'):
+        super(ModulatedConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.demodulate = demodulate
+        self.sample_mode = sample_mode
+        self.eps = eps
+        self.interpolation_mode = interpolation_mode
+        if self.interpolation_mode == 'nearest':
+            self.align_corners = None
+        else:
+            self.align_corners = False
+
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+        # modulation inside each modulated conv
+        self.modulation = EqualLinear(
+            num_style_feat, in_channels, bias=True, bias_init_val=1, lr_mul=1, activation=None)
+
+        self.weight = nn.Parameter(torch.randn(1, out_channels, in_channels, kernel_size, kernel_size))
+        self.padding = kernel_size // 2
+
+    def forward(self, x, style):
+        """Forward function.
+
+        Args:
+            x (Tensor): Tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+
+        Returns:
+            Tensor: Modulated tensor after convolution.
+        """
+        b, c, h, w = x.shape  # c = c_in
+        # weight modulation
+        style = self.modulation(style).view(b, 1, c, 1, 1)
+        # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
+        weight = self.scale * self.weight * style  # (b, c_out, c_in, k, k)
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
+            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
+
+        if self.sample_mode == 'upsample':
+            x = F.interpolate(x, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
+        elif self.sample_mode == 'downsample':
+            x = F.interpolate(x, scale_factor=0.5, mode=self.interpolation_mode, align_corners=self.align_corners)
+
+        b, c, h, w = x.shape
+        x = x.view(1, b * c, h, w)
+        # weight: (b*c_out, c_in, k, k), groups=b
+        out = F.conv2d(x, weight, padding=self.padding, groups=b)
+        out = out.view(b, self.out_channels, *out.shape[2:4])
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size}, '
+                f'demodulate={self.demodulate}, sample_mode={self.sample_mode})')
+
+
+class StyleConv(nn.Module):
+    """Style conv.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
+            Default: None.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 interpolation_mode='bilinear'):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            num_style_feat,
+            demodulate=demodulate,
+            sample_mode=sample_mode,
+            interpolation_mode=interpolation_mode)
+        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
+        self.activate = FusedLeakyReLU(out_channels)
+
+    def forward(self, x, style, noise=None):
+        # modulate
+        out = self.modulated_conv(x, style)
+        # noise injection
+        if noise is None:
+            b, _, h, w = out.shape
+            noise = out.new_empty(b, 1, h, w).normal_()
+        out = out + self.weight * noise
+        # activation (with bias)
+        out = self.activate(out)
+        return out
+
+
+class ToRGB(nn.Module):
+    """To RGB from features.
+
+    Args:
+        in_channels (int): Channel number of input.
+        num_style_feat (int): Channel number of style features.
+        upsample (bool): Whether to upsample. Default: True.
+    """
+
+    def __init__(self, in_channels, num_style_feat, upsample=True, interpolation_mode='bilinear'):
+        super(ToRGB, self).__init__()
+        self.upsample = upsample
+        self.interpolation_mode = interpolation_mode
+        if self.interpolation_mode == 'nearest':
+            self.align_corners = None
+        else:
+            self.align_corners = False
+        self.modulated_conv = ModulatedConv2d(
+            in_channels,
+            3,
+            kernel_size=1,
+            num_style_feat=num_style_feat,
+            demodulate=False,
+            sample_mode=None,
+            interpolation_mode=interpolation_mode)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, x, style, skip=None):
+        """Forward function.
+
+        Args:
+            x (Tensor): Feature tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+            skip (Tensor): Base/skip tensor. Default: None.
+
+        Returns:
+            Tensor: RGB images.
+        """
+        out = self.modulated_conv(x, style)
+        out = out + self.bias
+        if skip is not None:
+            if self.upsample:
+                skip = F.interpolate(
+                    skip, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
+            out = out + skip
+        return out
+
+
+class ConstantInput(nn.Module):
+    """Constant input.
+
+    Args:
+        num_channel (int): Channel number of constant input.
+        size (int): Spatial size of constant input.
+    """
+
+    def __init__(self, num_channel, size):
+        super(ConstantInput, self).__init__()
+        self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
+
+    def forward(self, batch):
+        out = self.weight.repeat(batch, 1, 1, 1)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class StyleGAN2GeneratorBilinear(nn.Module):
+    """StyleGAN2 Generator.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of
+            StyleGAN2. Default: 2.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): Narrow ratio for channels. Default: 1.0.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 lr_mlp=0.01,
+                 narrow=1,
+                 interpolation_mode='bilinear'):
+        super(StyleGAN2GeneratorBilinear, self).__init__()
+        # Style MLP layers
+        self.num_style_feat = num_style_feat
+        style_mlp_layers = [NormStyleCode()]
+        for i in range(num_mlp):
+            style_mlp_layers.append(
+                EqualLinear(
+                    num_style_feat, num_style_feat, bias=True, bias_init_val=0, lr_mul=lr_mlp,
+                    activation='fused_lrelu'))
+        self.style_mlp = nn.Sequential(*style_mlp_layers)
+
+        channels = {
+            '4': int(512 * narrow),
+            '8': int(512 * narrow),
+            '16': int(512 * narrow),
+            '32': int(512 * narrow),
+            '64': int(256 * channel_multiplier * narrow),
+            '128': int(128 * channel_multiplier * narrow),
+            '256': int(64 * channel_multiplier * narrow),
+            '512': int(32 * channel_multiplier * narrow),
+            '1024': int(16 * channel_multiplier * narrow)
+        }
+        self.channels = channels
+
+        self.constant_input = ConstantInput(channels['4'], size=4)
+        self.style_conv1 = StyleConv(
+            channels['4'],
+            channels['4'],
+            kernel_size=3,
+            num_style_feat=num_style_feat,
+            demodulate=True,
+            sample_mode=None,
+            interpolation_mode=interpolation_mode)
+        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False, interpolation_mode=interpolation_mode)
+
+        self.log_size = int(math.log(out_size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+        self.num_latent = self.log_size * 2 - 2
+
+        self.style_convs = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channels = channels['4']
+        # noise
+        for layer_idx in range(self.num_layers):
+            resolution = 2**((layer_idx + 5) // 2)
+            shape = [1, 1, resolution, resolution]
+            self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
+        # style convs and to_rgbs
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.style_convs.append(
+                StyleConv(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode='upsample',
+                    interpolation_mode=interpolation_mode))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None,
+                    interpolation_mode=interpolation_mode))
+            self.to_rgbs.append(
+                ToRGB(out_channels, num_style_feat, upsample=True, interpolation_mode=interpolation_mode))
+            in_channels = out_channels
+
+    def make_noise(self):
+        """Make noise for noise injection."""
+        device = self.constant_input.weight.device
+        noises = [torch.randn(1, 1, 4, 4, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
+
+        return noises
+
+    def get_latent(self, x):
+        return self.style_mlp(x)
+
+    def mean_latent(self, num_latent):
+        latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
+        latent = self.style_mlp(latent_in).mean(0, keepdim=True)
+        return latent
+
+    def forward(self,
+                styles,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2Generator.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            input_is_latent (bool): Whether input is latent style.
+                Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is
+                False. Default: True.
+            truncation (float): TODO. Default: 1.
+            truncation_latent (Tensor | None): TODO. Default: None.
+            inject_index (int | None): The injection index for mixing noise.
+                Default: None.
+            return_latents (bool): Whether to return style latents.
+                Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latent with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ScaledLeakyReLU(nn.Module):
+    """Scaled LeakyReLU.
+
+    Args:
+        negative_slope (float): Negative slope. Default: 0.2.
+    """
+
+    def __init__(self, negative_slope=0.2):
+        super(ScaledLeakyReLU, self).__init__()
+        self.negative_slope = negative_slope
+
+    def forward(self, x):
+        out = F.leaky_relu(x, negative_slope=self.negative_slope)
+        return out * math.sqrt(2)
+
+
+class EqualConv2d(nn.Module):
+    """Equalized Linear as StyleGAN2.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        stride (int): Stride of the convolution. Default: 1
+        padding (int): Zero-padding added to both sides of the input.
+            Default: 0.
+        bias (bool): If ``True``, adds a learnable bias to the output.
+            Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+    """
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, bias=True, bias_init_val=0):
+        super(EqualConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+    def forward(self, x):
+        out = F.conv2d(
+            x,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size},'
+                f' stride={self.stride}, padding={self.padding}, '
+                f'bias={self.bias is not None})')
+
+
+class ConvLayer(nn.Sequential):
+    """Conv Layer used in StyleGAN2 Discriminator.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Kernel size.
+        downsample (bool): Whether downsample by a factor of 2.
+            Default: False.
+        bias (bool): Whether with bias. Default: True.
+        activate (bool): Whether use activateion. Default: True.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 downsample=False,
+                 bias=True,
+                 activate=True,
+                 interpolation_mode='bilinear'):
+        layers = []
+        self.interpolation_mode = interpolation_mode
+        # downsample
+        if downsample:
+            if self.interpolation_mode == 'nearest':
+                self.align_corners = None
+            else:
+                self.align_corners = False
+
+            layers.append(
+                torch.nn.Upsample(scale_factor=0.5, mode=interpolation_mode, align_corners=self.align_corners))
+        stride = 1
+        self.padding = kernel_size // 2
+        # conv
+        layers.append(
+            EqualConv2d(
+                in_channels, out_channels, kernel_size, stride=stride, padding=self.padding, bias=bias
+                and not activate))
+        # activation
+        if activate:
+            if bias:
+                layers.append(FusedLeakyReLU(out_channels))
+            else:
+                layers.append(ScaledLeakyReLU(0.2))
+
+        super(ConvLayer, self).__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    """Residual block used in StyleGAN2 Discriminator.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+    """
+
+    def __init__(self, in_channels, out_channels, interpolation_mode='bilinear'):
+        super(ResBlock, self).__init__()
+
+        self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
+        self.conv2 = ConvLayer(
+            in_channels,
+            out_channels,
+            3,
+            downsample=True,
+            interpolation_mode=interpolation_mode,
+            bias=True,
+            activate=True)
+        self.skip = ConvLayer(
+            in_channels,
+            out_channels,
+            1,
+            downsample=True,
+            interpolation_mode=interpolation_mode,
+            bias=False,
+            activate=False)
+
+    def forward(self, x):
+        out = self.conv1(x)
+        out = self.conv2(out)
+        skip = self.skip(x)
+        out = (out + skip) / math.sqrt(2)
+        return out

video-tetalking/third_part/GFPGAN/gfpgan/archs/stylegan2_clean_arch.py

@@ -0,0 +1,368 @@
+import math
+import random
+import torch
+from basicsr.archs.arch_util import default_init_weights
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class NormStyleCode(nn.Module):
+
+    def forward(self, x):
+        """Normalize the style codes.
+
+        Args:
+            x (Tensor): Style codes with shape (b, c).
+
+        Returns:
+            Tensor: Normalized tensor.
+        """
+        return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
+
+
+class ModulatedConv2d(nn.Module):
+    """Modulated Conv2d used in StyleGAN2.
+
+    There is no bias in ModulatedConv2d.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether to demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
+        eps (float): A value added to the denominator for numerical stability. Default: 1e-8.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 eps=1e-8):
+        super(ModulatedConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.demodulate = demodulate
+        self.sample_mode = sample_mode
+        self.eps = eps
+
+        # modulation inside each modulated conv
+        self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
+        # initialization
+        default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
+            math.sqrt(in_channels * kernel_size**2))
+        self.padding = kernel_size // 2
+
+    def forward(self, x, style):
+        """Forward function.
+
+        Args:
+            x (Tensor): Tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+
+        Returns:
+            Tensor: Modulated tensor after convolution.
+        """
+        b, c, h, w = x.shape  # c = c_in
+        # weight modulation
+        style = self.modulation(style).view(b, 1, c, 1, 1)
+        # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
+        weight = self.weight * style  # (b, c_out, c_in, k, k)
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
+            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
+
+        # upsample or downsample if necessary
+        if self.sample_mode == 'upsample':
+            x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
+        elif self.sample_mode == 'downsample':
+            x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
+
+        b, c, h, w = x.shape
+        x = x.view(1, b * c, h, w)
+        # weight: (b*c_out, c_in, k, k), groups=b
+        out = F.conv2d(x, weight, padding=self.padding, groups=b)
+        out = out.view(b, self.out_channels, *out.shape[2:4])
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
+
+
+class StyleConv(nn.Module):
+    """Style conv used in StyleGAN2.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
+    """
+
+    def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
+        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
+        self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
+        self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x, style, noise=None):
+        # modulate
+        out = self.modulated_conv(x, style) * 2**0.5  # for conversion
+        # noise injection
+        if noise is None:
+            b, _, h, w = out.shape
+            noise = out.new_empty(b, 1, h, w).normal_()
+        out = out + self.weight * noise
+        # add bias
+        out = out + self.bias
+        # activation
+        out = self.activate(out)
+        return out
+
+
+class ToRGB(nn.Module):
+    """To RGB (image space) from features.
+
+    Args:
+        in_channels (int): Channel number of input.
+        num_style_feat (int): Channel number of style features.
+        upsample (bool): Whether to upsample. Default: True.
+    """
+
+    def __init__(self, in_channels, num_style_feat, upsample=True):
+        super(ToRGB, self).__init__()
+        self.upsample = upsample
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, x, style, skip=None):
+        """Forward function.
+
+        Args:
+            x (Tensor): Feature tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+            skip (Tensor): Base/skip tensor. Default: None.
+
+        Returns:
+            Tensor: RGB images.
+        """
+        out = self.modulated_conv(x, style)
+        out = out + self.bias
+        if skip is not None:
+            if self.upsample:
+                skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
+            out = out + skip
+        return out
+
+
+class ConstantInput(nn.Module):
+    """Constant input.
+
+    Args:
+        num_channel (int): Channel number of constant input.
+        size (int): Spatial size of constant input.
+    """
+
+    def __init__(self, num_channel, size):
+        super(ConstantInput, self).__init__()
+        self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
+
+    def forward(self, batch):
+        out = self.weight.repeat(batch, 1, 1, 1)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class StyleGAN2GeneratorClean(nn.Module):
+    """Clean version of StyleGAN2 Generator.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        narrow (float): Narrow ratio for channels. Default: 1.0.
+    """
+
+    def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1):
+        super(StyleGAN2GeneratorClean, self).__init__()
+        # Style MLP layers
+        self.num_style_feat = num_style_feat
+        style_mlp_layers = [NormStyleCode()]
+        for i in range(num_mlp):
+            style_mlp_layers.extend(
+                [nn.Linear(num_style_feat, num_style_feat, bias=True),
+                 nn.LeakyReLU(negative_slope=0.2, inplace=True)])
+        self.style_mlp = nn.Sequential(*style_mlp_layers)
+        # initialization
+        default_init_weights(self.style_mlp, scale=1, bias_fill=0, a=0.2, mode='fan_in', nonlinearity='leaky_relu')
+
+        # channel list
+        channels = {
+            '4': int(512 * narrow),
+            '8': int(512 * narrow),
+            '16': int(512 * narrow),
+            '32': int(512 * narrow),
+            '64': int(256 * channel_multiplier * narrow),
+            '128': int(128 * channel_multiplier * narrow),
+            '256': int(64 * channel_multiplier * narrow),
+            '512': int(32 * channel_multiplier * narrow),
+            '1024': int(16 * channel_multiplier * narrow)
+        }
+        self.channels = channels
+
+        self.constant_input = ConstantInput(channels['4'], size=4)
+        self.style_conv1 = StyleConv(
+            channels['4'],
+            channels['4'],
+            kernel_size=3,
+            num_style_feat=num_style_feat,
+            demodulate=True,
+            sample_mode=None)
+        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False)
+
+        self.log_size = int(math.log(out_size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+        self.num_latent = self.log_size * 2 - 2
+
+        self.style_convs = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channels = channels['4']
+        # noise
+        for layer_idx in range(self.num_layers):
+            resolution = 2**((layer_idx + 5) // 2)
+            shape = [1, 1, resolution, resolution]
+            self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
+        # style convs and to_rgbs
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.style_convs.append(
+                StyleConv(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode='upsample'))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None))
+            self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
+            in_channels = out_channels
+
+    def make_noise(self):
+        """Make noise for noise injection."""
+        device = self.constant_input.weight.device
+        noises = [torch.randn(1, 1, 4, 4, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
+
+        return noises
+
+    def get_latent(self, x):
+        return self.style_mlp(x)
+
+    def mean_latent(self, num_latent):
+        latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
+        latent = self.style_mlp(latent_in).mean(0, keepdim=True)
+        return latent
+
+    def forward(self,
+                styles,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorClean.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None

video-tetalking/third_part/GFPGAN/gfpgan/data/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import dataset modules for registry
+# scan all the files that end with '_dataset.py' under the data folder
+data_folder = osp.dirname(osp.abspath(__file__))
+dataset_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(data_folder) if v.endswith('_dataset.py')]
+# import all the dataset modules
+_dataset_modules = [importlib.import_module(f'gfpgan.data.{file_name}') for file_name in dataset_filenames]

video-tetalking/third_part/GFPGAN/gfpgan/data/ffhq_degradation_dataset.py

@@ -0,0 +1,230 @@
+import cv2
+import math
+import numpy as np
+import os.path as osp
+import torch
+import torch.utils.data as data
+from basicsr.data import degradations as degradations
+from basicsr.data.data_util import paths_from_folder
+from basicsr.data.transforms import augment
+from basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor
+from basicsr.utils.registry import DATASET_REGISTRY
+from torchvision.transforms.functional import (adjust_brightness, adjust_contrast, adjust_hue, adjust_saturation,
+                                               normalize)
+
+
+@DATASET_REGISTRY.register()
+class FFHQDegradationDataset(data.Dataset):
+    """FFHQ dataset for GFPGAN.
+
+    It reads high resolution images, and then generate low-quality (LQ) images on-the-fly.
+
+    Args:
+        opt (dict): Config for train datasets. It contains the following keys:
+            dataroot_gt (str): Data root path for gt.
+            io_backend (dict): IO backend type and other kwarg.
+            mean (list | tuple): Image mean.
+            std (list | tuple): Image std.
+            use_hflip (bool): Whether to horizontally flip.
+            Please see more options in the codes.
+    """
+
+    def __init__(self, opt):
+        super(FFHQDegradationDataset, self).__init__()
+        self.opt = opt
+        # file client (io backend)
+        self.file_client = None
+        self.io_backend_opt = opt['io_backend']
+
+        self.gt_folder = opt['dataroot_gt']
+        self.mean = opt['mean']
+        self.std = opt['std']
+        self.out_size = opt['out_size']
+
+        self.crop_components = opt.get('crop_components', False)  # facial components
+        self.eye_enlarge_ratio = opt.get('eye_enlarge_ratio', 1)  # whether enlarge eye regions
+
+        if self.crop_components:
+            # load component list from a pre-process pth files
+            self.components_list = torch.load(opt.get('component_path'))
+
+        # file client (lmdb io backend)
+        if self.io_backend_opt['type'] == 'lmdb':
+            self.io_backend_opt['db_paths'] = self.gt_folder
+            if not self.gt_folder.endswith('.lmdb'):
+                raise ValueError(f"'dataroot_gt' should end with '.lmdb', but received {self.gt_folder}")
+            with open(osp.join(self.gt_folder, 'meta_info.txt')) as fin:
+                self.paths = [line.split('.')[0] for line in fin]
+        else:
+            # disk backend: scan file list from a folder
+            self.paths = paths_from_folder(self.gt_folder)
+
+        # degradation configurations
+        self.blur_kernel_size = opt['blur_kernel_size']
+        self.kernel_list = opt['kernel_list']
+        self.kernel_prob = opt['kernel_prob']
+        self.blur_sigma = opt['blur_sigma']
+        self.downsample_range = opt['downsample_range']
+        self.noise_range = opt['noise_range']
+        self.jpeg_range = opt['jpeg_range']
+
+        # color jitter
+        self.color_jitter_prob = opt.get('color_jitter_prob')
+        self.color_jitter_pt_prob = opt.get('color_jitter_pt_prob')
+        self.color_jitter_shift = opt.get('color_jitter_shift', 20)
+        # to gray
+        self.gray_prob = opt.get('gray_prob')
+
+        logger = get_root_logger()
+        logger.info(f'Blur: blur_kernel_size {self.blur_kernel_size}, sigma: [{", ".join(map(str, self.blur_sigma))}]')
+        logger.info(f'Downsample: downsample_range [{", ".join(map(str, self.downsample_range))}]')
+        logger.info(f'Noise: [{", ".join(map(str, self.noise_range))}]')
+        logger.info(f'JPEG compression: [{", ".join(map(str, self.jpeg_range))}]')
+
+        if self.color_jitter_prob is not None:
+            logger.info(f'Use random color jitter. Prob: {self.color_jitter_prob}, shift: {self.color_jitter_shift}')
+        if self.gray_prob is not None:
+            logger.info(f'Use random gray. Prob: {self.gray_prob}')
+        self.color_jitter_shift /= 255.
+
+    @staticmethod
+    def color_jitter(img, shift):
+        """jitter color: randomly jitter the RGB values, in numpy formats"""
+        jitter_val = np.random.uniform(-shift, shift, 3).astype(np.float32)
+        img = img + jitter_val
+        img = np.clip(img, 0, 1)
+        return img
+
+    @staticmethod
+    def color_jitter_pt(img, brightness, contrast, saturation, hue):
+        """jitter color: randomly jitter the brightness, contrast, saturation, and hue, in torch Tensor formats"""
+        fn_idx = torch.randperm(4)
+        for fn_id in fn_idx:
+            if fn_id == 0 and brightness is not None:
+                brightness_factor = torch.tensor(1.0).uniform_(brightness[0], brightness[1]).item()
+                img = adjust_brightness(img, brightness_factor)
+
+            if fn_id == 1 and contrast is not None:
+                contrast_factor = torch.tensor(1.0).uniform_(contrast[0], contrast[1]).item()
+                img = adjust_contrast(img, contrast_factor)
+
+            if fn_id == 2 and saturation is not None:
+                saturation_factor = torch.tensor(1.0).uniform_(saturation[0], saturation[1]).item()
+                img = adjust_saturation(img, saturation_factor)
+
+            if fn_id == 3 and hue is not None:
+                hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
+                img = adjust_hue(img, hue_factor)
+        return img
+
+    def get_component_coordinates(self, index, status):
+        """Get facial component (left_eye, right_eye, mouth) coordinates from a pre-loaded pth file"""
+        components_bbox = self.components_list[f'{index:08d}']
+        if status[0]:  # hflip
+            # exchange right and left eye
+            tmp = components_bbox['left_eye']
+            components_bbox['left_eye'] = components_bbox['right_eye']
+            components_bbox['right_eye'] = tmp
+            # modify the width coordinate
+            components_bbox['left_eye'][0] = self.out_size - components_bbox['left_eye'][0]
+            components_bbox['right_eye'][0] = self.out_size - components_bbox['right_eye'][0]
+            components_bbox['mouth'][0] = self.out_size - components_bbox['mouth'][0]
+
+        # get coordinates
+        locations = []
+        for part in ['left_eye', 'right_eye', 'mouth']:
+            mean = components_bbox[part][0:2]
+            half_len = components_bbox[part][2]
+            if 'eye' in part:
+                half_len *= self.eye_enlarge_ratio
+            loc = np.hstack((mean - half_len + 1, mean + half_len))
+            loc = torch.from_numpy(loc).float()
+            locations.append(loc)
+        return locations
+
+    def __getitem__(self, index):
+        if self.file_client is None:
+            self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
+
+        # load gt image
+        # Shape: (h, w, c); channel order: BGR; image range: [0, 1], float32.
+        gt_path = self.paths[index]
+        img_bytes = self.file_client.get(gt_path)
+        img_gt = imfrombytes(img_bytes, float32=True)
+
+        # random horizontal flip
+        img_gt, status = augment(img_gt, hflip=self.opt['use_hflip'], rotation=False, return_status=True)
+        h, w, _ = img_gt.shape
+
+        # get facial component coordinates
+        if self.crop_components:
+            locations = self.get_component_coordinates(index, status)
+            loc_left_eye, loc_right_eye, loc_mouth = locations
+
+        # ------------------------ generate lq image ------------------------ #
+        # blur
+        kernel = degradations.random_mixed_kernels(
+            self.kernel_list,
+            self.kernel_prob,
+            self.blur_kernel_size,
+            self.blur_sigma,
+            self.blur_sigma, [-math.pi, math.pi],
+            noise_range=None)
+        img_lq = cv2.filter2D(img_gt, -1, kernel)
+        # downsample
+        scale = np.random.uniform(self.downsample_range[0], self.downsample_range[1])
+        img_lq = cv2.resize(img_lq, (int(w // scale), int(h // scale)), interpolation=cv2.INTER_LINEAR)
+        # noise
+        if self.noise_range is not None:
+            img_lq = degradations.random_add_gaussian_noise(img_lq, self.noise_range)
+        # jpeg compression
+        if self.jpeg_range is not None:
+            img_lq = degradations.random_add_jpg_compression(img_lq, self.jpeg_range)
+
+        # resize to original size
+        img_lq = cv2.resize(img_lq, (w, h), interpolation=cv2.INTER_LINEAR)
+
+        # random color jitter (only for lq)
+        if self.color_jitter_prob is not None and (np.random.uniform() < self.color_jitter_prob):
+            img_lq = self.color_jitter(img_lq, self.color_jitter_shift)
+        # random to gray (only for lq)
+        if self.gray_prob and np.random.uniform() < self.gray_prob:
+            img_lq = cv2.cvtColor(img_lq, cv2.COLOR_BGR2GRAY)
+            img_lq = np.tile(img_lq[:, :, None], [1, 1, 3])
+            if self.opt.get('gt_gray'):  # whether convert GT to gray images
+                img_gt = cv2.cvtColor(img_gt, cv2.COLOR_BGR2GRAY)
+                img_gt = np.tile(img_gt[:, :, None], [1, 1, 3])  # repeat the color channels
+
+        # BGR to RGB, HWC to CHW, numpy to tensor
+        img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
+
+        # random color jitter (pytorch version) (only for lq)
+        if self.color_jitter_pt_prob is not None and (np.random.uniform() < self.color_jitter_pt_prob):
+            brightness = self.opt.get('brightness', (0.5, 1.5))
+            contrast = self.opt.get('contrast', (0.5, 1.5))
+            saturation = self.opt.get('saturation', (0, 1.5))
+            hue = self.opt.get('hue', (-0.1, 0.1))
+            img_lq = self.color_jitter_pt(img_lq, brightness, contrast, saturation, hue)
+
+        # round and clip
+        img_lq = torch.clamp((img_lq * 255.0).round(), 0, 255) / 255.
+
+        # normalize
+        normalize(img_gt, self.mean, self.std, inplace=True)
+        normalize(img_lq, self.mean, self.std, inplace=True)
+
+        if self.crop_components:
+            return_dict = {
+                'lq': img_lq,
+                'gt': img_gt,
+                'gt_path': gt_path,
+                'loc_left_eye': loc_left_eye,
+                'loc_right_eye': loc_right_eye,
+                'loc_mouth': loc_mouth
+            }
+            return return_dict
+        else:
+            return {'lq': img_lq, 'gt': img_gt, 'gt_path': gt_path}
+
+    def __len__(self):
+        return len(self.paths)

video-tetalking/third_part/GFPGAN/gfpgan/models/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import model modules for registry
+# scan all the files that end with '_model.py' under the model folder
+model_folder = osp.dirname(osp.abspath(__file__))
+model_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(model_folder) if v.endswith('_model.py')]
+# import all the model modules
+_model_modules = [importlib.import_module(f'gfpgan.models.{file_name}') for file_name in model_filenames]

video-tetalking/third_part/GFPGAN/gfpgan/models/gfpgan_model.py

@@ -0,0 +1,580 @@
+import math
+import os.path as osp
+import torch
+from basicsr.archs import build_network
+from basicsr.losses import build_loss
+# from basicsr.losses.losses import r1_penalty
+from basicsr.losses import r1_penalty
+from basicsr.metrics import calculate_metric
+from basicsr.models.base_model import BaseModel
+from basicsr.utils import get_root_logger, imwrite, tensor2img
+from basicsr.utils.registry import MODEL_REGISTRY
+from collections import OrderedDict
+from torch.nn import functional as F
+from torchvision.ops import roi_align
+from tqdm import tqdm
+
+
+@MODEL_REGISTRY.register()
+class GFPGANModel(BaseModel):
+    """The GFPGAN model for Towards real-world blind face restoratin with generative facial prior"""
+
+    def __init__(self, opt):
+        super(GFPGANModel, self).__init__(opt)
+        self.idx = 0  # it is used for saving data for check
+
+        # define network
+        self.net_g = build_network(opt['network_g'])
+        self.net_g = self.model_to_device(self.net_g)
+        self.print_network(self.net_g)
+
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_g', None)
+        if load_path is not None:
+            param_key = self.opt['path'].get('param_key_g', 'params')
+            self.load_network(self.net_g, load_path, self.opt['path'].get('strict_load_g', True), param_key)
+
+        self.log_size = int(math.log(self.opt['network_g']['out_size'], 2))
+
+        if self.is_train:
+            self.init_training_settings()
+
+    def init_training_settings(self):
+        train_opt = self.opt['train']
+
+        # ----------- define net_d ----------- #
+        self.net_d = build_network(self.opt['network_d'])
+        self.net_d = self.model_to_device(self.net_d)
+        self.print_network(self.net_d)
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_d', None)
+        if load_path is not None:
+            self.load_network(self.net_d, load_path, self.opt['path'].get('strict_load_d', True))
+
+        # ----------- define net_g with Exponential Moving Average (EMA) ----------- #
+        # net_g_ema only used for testing on one GPU and saving. There is no need to wrap with DistributedDataParallel
+        self.net_g_ema = build_network(self.opt['network_g']).to(self.device)
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_g', None)
+        if load_path is not None:
+            self.load_network(self.net_g_ema, load_path, self.opt['path'].get('strict_load_g', True), 'params_ema')
+        else:
+            self.model_ema(0)  # copy net_g weight
+
+        self.net_g.train()
+        self.net_d.train()
+        self.net_g_ema.eval()
+
+        # ----------- facial component networks ----------- #
+        if ('network_d_left_eye' in self.opt and 'network_d_right_eye' in self.opt and 'network_d_mouth' in self.opt):
+            self.use_facial_disc = True
+        else:
+            self.use_facial_disc = False
+
+        if self.use_facial_disc:
+            # left eye
+            self.net_d_left_eye = build_network(self.opt['network_d_left_eye'])
+            self.net_d_left_eye = self.model_to_device(self.net_d_left_eye)
+            self.print_network(self.net_d_left_eye)
+            load_path = self.opt['path'].get('pretrain_network_d_left_eye')
+            if load_path is not None:
+                self.load_network(self.net_d_left_eye, load_path, True, 'params')
+            # right eye
+            self.net_d_right_eye = build_network(self.opt['network_d_right_eye'])
+            self.net_d_right_eye = self.model_to_device(self.net_d_right_eye)
+            self.print_network(self.net_d_right_eye)
+            load_path = self.opt['path'].get('pretrain_network_d_right_eye')
+            if load_path is not None:
+                self.load_network(self.net_d_right_eye, load_path, True, 'params')
+            # mouth
+            self.net_d_mouth = build_network(self.opt['network_d_mouth'])
+            self.net_d_mouth = self.model_to_device(self.net_d_mouth)
+            self.print_network(self.net_d_mouth)
+            load_path = self.opt['path'].get('pretrain_network_d_mouth')
+            if load_path is not None:
+                self.load_network(self.net_d_mouth, load_path, True, 'params')
+
+            self.net_d_left_eye.train()
+            self.net_d_right_eye.train()
+            self.net_d_mouth.train()
+
+            # ----------- define facial component gan loss ----------- #
+            self.cri_component = build_loss(train_opt['gan_component_opt']).to(self.device)
+
+        # ----------- define losses ----------- #
+        # pixel loss
+        if train_opt.get('pixel_opt'):
+            self.cri_pix = build_loss(train_opt['pixel_opt']).to(self.device)
+        else:
+            self.cri_pix = None
+
+        # perceptual loss
+        if train_opt.get('perceptual_opt'):
+            self.cri_perceptual = build_loss(train_opt['perceptual_opt']).to(self.device)
+        else:
+            self.cri_perceptual = None
+
+        # L1 loss is used in pyramid loss, component style loss and identity loss
+        self.cri_l1 = build_loss(train_opt['L1_opt']).to(self.device)
+
+        # gan loss (wgan)
+        self.cri_gan = build_loss(train_opt['gan_opt']).to(self.device)
+
+        # ----------- define identity loss ----------- #
+        if 'network_identity' in self.opt:
+            self.use_identity = True
+        else:
+            self.use_identity = False
+
+        if self.use_identity:
+            # define identity network
+            self.network_identity = build_network(self.opt['network_identity'])
+            self.network_identity = self.model_to_device(self.network_identity)
+            self.print_network(self.network_identity)
+            load_path = self.opt['path'].get('pretrain_network_identity')
+            if load_path is not None:
+                self.load_network(self.network_identity, load_path, True, None)
+            self.network_identity.eval()
+            for param in self.network_identity.parameters():
+                param.requires_grad = False
+
+        # regularization weights
+        self.r1_reg_weight = train_opt['r1_reg_weight']  # for discriminator
+        self.net_d_iters = train_opt.get('net_d_iters', 1)
+        self.net_d_init_iters = train_opt.get('net_d_init_iters', 0)
+        self.net_d_reg_every = train_opt['net_d_reg_every']
+
+        # set up optimizers and schedulers
+        self.setup_optimizers()
+        self.setup_schedulers()
+
+    def setup_optimizers(self):
+        train_opt = self.opt['train']
+
+        # ----------- optimizer g ----------- #
+        net_g_reg_ratio = 1
+        normal_params = []
+        for _, param in self.net_g.named_parameters():
+            normal_params.append(param)
+        optim_params_g = [{  # add normal params first
+            'params': normal_params,
+            'lr': train_opt['optim_g']['lr']
+        }]
+        optim_type = train_opt['optim_g'].pop('type')
+        lr = train_opt['optim_g']['lr'] * net_g_reg_ratio
+        betas = (0**net_g_reg_ratio, 0.99**net_g_reg_ratio)
+        self.optimizer_g = self.get_optimizer(optim_type, optim_params_g, lr, betas=betas)
+        self.optimizers.append(self.optimizer_g)
+
+        # ----------- optimizer d ----------- #
+        net_d_reg_ratio = self.net_d_reg_every / (self.net_d_reg_every + 1)
+        normal_params = []
+        for _, param in self.net_d.named_parameters():
+            normal_params.append(param)
+        optim_params_d = [{  # add normal params first
+            'params': normal_params,
+            'lr': train_opt['optim_d']['lr']
+        }]
+        optim_type = train_opt['optim_d'].pop('type')
+        lr = train_opt['optim_d']['lr'] * net_d_reg_ratio
+        betas = (0**net_d_reg_ratio, 0.99**net_d_reg_ratio)
+        self.optimizer_d = self.get_optimizer(optim_type, optim_params_d, lr, betas=betas)
+        self.optimizers.append(self.optimizer_d)
+
+        # ----------- optimizers for facial component networks ----------- #
+        if self.use_facial_disc:
+            # setup optimizers for facial component discriminators
+            optim_type = train_opt['optim_component'].pop('type')
+            lr = train_opt['optim_component']['lr']
+            # left eye
+            self.optimizer_d_left_eye = self.get_optimizer(
+                optim_type, self.net_d_left_eye.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_left_eye)
+            # right eye
+            self.optimizer_d_right_eye = self.get_optimizer(
+                optim_type, self.net_d_right_eye.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_right_eye)
+            # mouth
+            self.optimizer_d_mouth = self.get_optimizer(
+                optim_type, self.net_d_mouth.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_mouth)
+
+    def feed_data(self, data):
+        self.lq = data['lq'].to(self.device)
+        if 'gt' in data:
+            self.gt = data['gt'].to(self.device)
+
+        if 'loc_left_eye' in data:
+            # get facial component locations, shape (batch, 4)
+            self.loc_left_eyes = data['loc_left_eye']
+            self.loc_right_eyes = data['loc_right_eye']
+            self.loc_mouths = data['loc_mouth']
+
+        # uncomment to check data
+        # import torchvision
+        # if self.opt['rank'] == 0:
+        #     import os
+        #     os.makedirs('tmp/gt', exist_ok=True)
+        #     os.makedirs('tmp/lq', exist_ok=True)
+        #     print(self.idx)
+        #     torchvision.utils.save_image(
+        #         self.gt, f'tmp/gt/gt_{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
+        #     torchvision.utils.save_image(
+        #         self.lq, f'tmp/lq/lq{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
+        #     self.idx = self.idx + 1
+
+    def construct_img_pyramid(self):
+        """Construct image pyramid for intermediate restoration loss"""
+        pyramid_gt = [self.gt]
+        down_img = self.gt
+        for _ in range(0, self.log_size - 3):
+            down_img = F.interpolate(down_img, scale_factor=0.5, mode='bilinear', align_corners=False)
+            pyramid_gt.insert(0, down_img)
+        return pyramid_gt
+
+    def get_roi_regions(self, eye_out_size=80, mouth_out_size=120):
+        face_ratio = int(self.opt['network_g']['out_size'] / 512)
+        eye_out_size *= face_ratio
+        mouth_out_size *= face_ratio
+
+        rois_eyes = []
+        rois_mouths = []
+        for b in range(self.loc_left_eyes.size(0)):  # loop for batch size
+            # left eye and right eye
+            img_inds = self.loc_left_eyes.new_full((2, 1), b)
+            bbox = torch.stack([self.loc_left_eyes[b, :], self.loc_right_eyes[b, :]], dim=0)  # shape: (2, 4)
+            rois = torch.cat([img_inds, bbox], dim=-1)  # shape: (2, 5)
+            rois_eyes.append(rois)
+            # mouse
+            img_inds = self.loc_left_eyes.new_full((1, 1), b)
+            rois = torch.cat([img_inds, self.loc_mouths[b:b + 1, :]], dim=-1)  # shape: (1, 5)
+            rois_mouths.append(rois)
+
+        rois_eyes = torch.cat(rois_eyes, 0).to(self.device)
+        rois_mouths = torch.cat(rois_mouths, 0).to(self.device)
+
+        # real images
+        all_eyes = roi_align(self.gt, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
+        self.left_eyes_gt = all_eyes[0::2, :, :, :]
+        self.right_eyes_gt = all_eyes[1::2, :, :, :]
+        self.mouths_gt = roi_align(self.gt, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
+        # output
+        all_eyes = roi_align(self.output, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
+        self.left_eyes = all_eyes[0::2, :, :, :]
+        self.right_eyes = all_eyes[1::2, :, :, :]
+        self.mouths = roi_align(self.output, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
+
+    def _gram_mat(self, x):
+        """Calculate Gram matrix.
+
+        Args:
+            x (torch.Tensor): Tensor with shape of (n, c, h, w).
+
+        Returns:
+            torch.Tensor: Gram matrix.
+        """
+        n, c, h, w = x.size()
+        features = x.view(n, c, w * h)
+        features_t = features.transpose(1, 2)
+        gram = features.bmm(features_t) / (c * h * w)
+        return gram
+
+    def gray_resize_for_identity(self, out, size=128):
+        out_gray = (0.2989 * out[:, 0, :, :] + 0.5870 * out[:, 1, :, :] + 0.1140 * out[:, 2, :, :])
+        out_gray = out_gray.unsqueeze(1)
+        out_gray = F.interpolate(out_gray, (size, size), mode='bilinear', align_corners=False)
+        return out_gray
+
+    def optimize_parameters(self, current_iter):
+        # optimize net_g
+        for p in self.net_d.parameters():
+            p.requires_grad = False
+        self.optimizer_g.zero_grad()
+
+        # do not update facial component net_d
+        if self.use_facial_disc:
+            for p in self.net_d_left_eye.parameters():
+                p.requires_grad = False
+            for p in self.net_d_right_eye.parameters():
+                p.requires_grad = False
+            for p in self.net_d_mouth.parameters():
+                p.requires_grad = False
+
+        # image pyramid loss weight
+        pyramid_loss_weight = self.opt['train'].get('pyramid_loss_weight', 0)
+        if pyramid_loss_weight > 0 and current_iter > self.opt['train'].get('remove_pyramid_loss', float('inf')):
+            pyramid_loss_weight = 1e-12  # very small weight to avoid unused param error
+        if pyramid_loss_weight > 0:
+            self.output, out_rgbs = self.net_g(self.lq, return_rgb=True)
+            pyramid_gt = self.construct_img_pyramid()
+        else:
+            self.output, out_rgbs = self.net_g(self.lq, return_rgb=False)
+
+        # get roi-align regions
+        if self.use_facial_disc:
+            self.get_roi_regions(eye_out_size=80, mouth_out_size=120)
+
+        l_g_total = 0
+        loss_dict = OrderedDict()
+        if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):
+            # pixel loss
+            if self.cri_pix:
+                l_g_pix = self.cri_pix(self.output, self.gt)
+                l_g_total += l_g_pix
+                loss_dict['l_g_pix'] = l_g_pix
+
+            # image pyramid loss
+            if pyramid_loss_weight > 0:
+                for i in range(0, self.log_size - 2):
+                    l_pyramid = self.cri_l1(out_rgbs[i], pyramid_gt[i]) * pyramid_loss_weight
+                    l_g_total += l_pyramid
+                    loss_dict[f'l_p_{2**(i+3)}'] = l_pyramid
+
+            # perceptual loss
+            if self.cri_perceptual:
+                l_g_percep, l_g_style = self.cri_perceptual(self.output, self.gt)
+                if l_g_percep is not None:
+                    l_g_total += l_g_percep
+                    loss_dict['l_g_percep'] = l_g_percep
+                if l_g_style is not None:
+                    l_g_total += l_g_style
+                    loss_dict['l_g_style'] = l_g_style
+
+            # gan loss
+            fake_g_pred = self.net_d(self.output)
+            l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)
+            l_g_total += l_g_gan
+            loss_dict['l_g_gan'] = l_g_gan
+
+            # facial component loss
+            if self.use_facial_disc:
+                # left eye
+                fake_left_eye, fake_left_eye_feats = self.net_d_left_eye(self.left_eyes, return_feats=True)
+                l_g_gan = self.cri_component(fake_left_eye, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_left_eye'] = l_g_gan
+                # right eye
+                fake_right_eye, fake_right_eye_feats = self.net_d_right_eye(self.right_eyes, return_feats=True)
+                l_g_gan = self.cri_component(fake_right_eye, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_right_eye'] = l_g_gan
+                # mouth
+                fake_mouth, fake_mouth_feats = self.net_d_mouth(self.mouths, return_feats=True)
+                l_g_gan = self.cri_component(fake_mouth, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_mouth'] = l_g_gan
+
+                if self.opt['train'].get('comp_style_weight', 0) > 0:
+                    # get gt feat
+                    _, real_left_eye_feats = self.net_d_left_eye(self.left_eyes_gt, return_feats=True)
+                    _, real_right_eye_feats = self.net_d_right_eye(self.right_eyes_gt, return_feats=True)
+                    _, real_mouth_feats = self.net_d_mouth(self.mouths_gt, return_feats=True)
+
+                    def _comp_style(feat, feat_gt, criterion):
+                        return criterion(self._gram_mat(feat[0]), self._gram_mat(
+                            feat_gt[0].detach())) * 0.5 + criterion(
+                                self._gram_mat(feat[1]), self._gram_mat(feat_gt[1].detach()))
+
+                    # facial component style loss
+                    comp_style_loss = 0
+                    comp_style_loss += _comp_style(fake_left_eye_feats, real_left_eye_feats, self.cri_l1)
+                    comp_style_loss += _comp_style(fake_right_eye_feats, real_right_eye_feats, self.cri_l1)
+                    comp_style_loss += _comp_style(fake_mouth_feats, real_mouth_feats, self.cri_l1)
+                    comp_style_loss = comp_style_loss * self.opt['train']['comp_style_weight']
+                    l_g_total += comp_style_loss
+                    loss_dict['l_g_comp_style_loss'] = comp_style_loss
+
+            # identity loss
+            if self.use_identity:
+                identity_weight = self.opt['train']['identity_weight']
+                # get gray images and resize
+                out_gray = self.gray_resize_for_identity(self.output)
+                gt_gray = self.gray_resize_for_identity(self.gt)
+
+                identity_gt = self.network_identity(gt_gray).detach()
+                identity_out = self.network_identity(out_gray)
+                l_identity = self.cri_l1(identity_out, identity_gt) * identity_weight
+                l_g_total += l_identity
+                loss_dict['l_identity'] = l_identity
+
+            l_g_total.backward()
+            self.optimizer_g.step()
+
+        # EMA
+        self.model_ema(decay=0.5**(32 / (10 * 1000)))
+
+        # ----------- optimize net_d ----------- #
+        for p in self.net_d.parameters():
+            p.requires_grad = True
+        self.optimizer_d.zero_grad()
+        if self.use_facial_disc:
+            for p in self.net_d_left_eye.parameters():
+                p.requires_grad = True
+            for p in self.net_d_right_eye.parameters():
+                p.requires_grad = True
+            for p in self.net_d_mouth.parameters():
+                p.requires_grad = True
+            self.optimizer_d_left_eye.zero_grad()
+            self.optimizer_d_right_eye.zero_grad()
+            self.optimizer_d_mouth.zero_grad()
+
+        fake_d_pred = self.net_d(self.output.detach())
+        real_d_pred = self.net_d(self.gt)
+        l_d = self.cri_gan(real_d_pred, True, is_disc=True) + self.cri_gan(fake_d_pred, False, is_disc=True)
+        loss_dict['l_d'] = l_d
+        # In WGAN, real_score should be positive and fake_score should be negative
+        loss_dict['real_score'] = real_d_pred.detach().mean()
+        loss_dict['fake_score'] = fake_d_pred.detach().mean()
+        l_d.backward()
+
+        # regularization loss
+        if current_iter % self.net_d_reg_every == 0:
+            self.gt.requires_grad = True
+            real_pred = self.net_d(self.gt)
+            l_d_r1 = r1_penalty(real_pred, self.gt)
+            l_d_r1 = (self.r1_reg_weight / 2 * l_d_r1 * self.net_d_reg_every + 0 * real_pred[0])
+            loss_dict['l_d_r1'] = l_d_r1.detach().mean()
+            l_d_r1.backward()
+
+        self.optimizer_d.step()
+
+        # optimize facial component discriminators
+        if self.use_facial_disc:
+            # left eye
+            fake_d_pred, _ = self.net_d_left_eye(self.left_eyes.detach())
+            real_d_pred, _ = self.net_d_left_eye(self.left_eyes_gt)
+            l_d_left_eye = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_left_eye'] = l_d_left_eye
+            l_d_left_eye.backward()
+            # right eye
+            fake_d_pred, _ = self.net_d_right_eye(self.right_eyes.detach())
+            real_d_pred, _ = self.net_d_right_eye(self.right_eyes_gt)
+            l_d_right_eye = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_right_eye'] = l_d_right_eye
+            l_d_right_eye.backward()
+            # mouth
+            fake_d_pred, _ = self.net_d_mouth(self.mouths.detach())
+            real_d_pred, _ = self.net_d_mouth(self.mouths_gt)
+            l_d_mouth = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_mouth'] = l_d_mouth
+            l_d_mouth.backward()
+
+            self.optimizer_d_left_eye.step()
+            self.optimizer_d_right_eye.step()
+            self.optimizer_d_mouth.step()
+
+        self.log_dict = self.reduce_loss_dict(loss_dict)
+
+    def test(self):
+        with torch.no_grad():
+            if hasattr(self, 'net_g_ema'):
+                self.net_g_ema.eval()
+                self.output, _ = self.net_g_ema(self.lq)
+            else:
+                logger = get_root_logger()
+                logger.warning('Do not have self.net_g_ema, use self.net_g.')
+                self.net_g.eval()
+                self.output, _ = self.net_g(self.lq)
+                self.net_g.train()
+
+    def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        if self.opt['rank'] == 0:
+            self.nondist_validation(dataloader, current_iter, tb_logger, save_img)
+
+    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        dataset_name = dataloader.dataset.opt['name']
+        with_metrics = self.opt['val'].get('metrics') is not None
+        use_pbar = self.opt['val'].get('pbar', False)
+
+        if with_metrics:
+            if not hasattr(self, 'metric_results'):  # only execute in the first run
+                self.metric_results = {metric: 0 for metric in self.opt['val']['metrics'].keys()}
+            # initialize the best metric results for each dataset_name (supporting multiple validation datasets)
+            self._initialize_best_metric_results(dataset_name)
+            # zero self.metric_results
+            self.metric_results = {metric: 0 for metric in self.metric_results}
+
+        metric_data = dict()
+        if use_pbar:
+            pbar = tqdm(total=len(dataloader), unit='image')
+
+        for idx, val_data in enumerate(dataloader):
+            img_name = osp.splitext(osp.basename(val_data['lq_path'][0]))[0]
+            self.feed_data(val_data)
+            self.test()
+
+            sr_img = tensor2img(self.output.detach().cpu(), min_max=(-1, 1))
+            metric_data['img'] = sr_img
+            if hasattr(self, 'gt'):
+                gt_img = tensor2img(self.gt.detach().cpu(), min_max=(-1, 1))
+                metric_data['img2'] = gt_img
+                del self.gt
+
+            # tentative for out of GPU memory
+            del self.lq
+            del self.output
+            torch.cuda.empty_cache()
+
+            if save_img:
+                if self.opt['is_train']:
+                    save_img_path = osp.join(self.opt['path']['visualization'], img_name,
+                                             f'{img_name}_{current_iter}.png')
+                else:
+                    if self.opt['val']['suffix']:
+                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
+                                                 f'{img_name}_{self.opt["val"]["suffix"]}.png')
+                    else:
+                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
+                                                 f'{img_name}_{self.opt["name"]}.png')
+                imwrite(sr_img, save_img_path)
+
+            if with_metrics:
+                # calculate metrics
+                for name, opt_ in self.opt['val']['metrics'].items():
+                    self.metric_results[name] += calculate_metric(metric_data, opt_)
+            if use_pbar:
+                pbar.update(1)
+                pbar.set_description(f'Test {img_name}')
+        if use_pbar:
+            pbar.close()
+
+        if with_metrics:
+            for metric in self.metric_results.keys():
+                self.metric_results[metric] /= (idx + 1)
+                # update the best metric result
+                self._update_best_metric_result(dataset_name, metric, self.metric_results[metric], current_iter)
+
+            self._log_validation_metric_values(current_iter, dataset_name, tb_logger)
+
+    def _log_validation_metric_values(self, current_iter, dataset_name, tb_logger):
+        log_str = f'Validation {dataset_name}\n'
+        for metric, value in self.metric_results.items():
+            log_str += f'\t # {metric}: {value:.4f}'
+            if hasattr(self, 'best_metric_results'):
+                log_str += (f'\tBest: {self.best_metric_results[dataset_name][metric]["val"]:.4f} @ '
+                            f'{self.best_metric_results[dataset_name][metric]["iter"]} iter')
+            log_str += '\n'
+
+        logger = get_root_logger()
+        logger.info(log_str)
+        if tb_logger:
+            for metric, value in self.metric_results.items():
+                tb_logger.add_scalar(f'metrics/{dataset_name}/{metric}', value, current_iter)
+
+    def save(self, epoch, current_iter):
+        # save net_g and net_d
+        self.save_network([self.net_g, self.net_g_ema], 'net_g', current_iter, param_key=['params', 'params_ema'])
+        self.save_network(self.net_d, 'net_d', current_iter)
+        # save component discriminators
+        if self.use_facial_disc:
+            self.save_network(self.net_d_left_eye, 'net_d_left_eye', current_iter)
+            self.save_network(self.net_d_right_eye, 'net_d_right_eye', current_iter)
+            self.save_network(self.net_d_mouth, 'net_d_mouth', current_iter)
+        # save training state
+        self.save_training_state(epoch, current_iter)

video-tetalking/third_part/GFPGAN/gfpgan/train.py

@@ -0,0 +1,11 @@
+# flake8: noqa
+import os.path as osp
+from basicsr.train import train_pipeline
+
+import gfpgan.archs
+import gfpgan.data
+import gfpgan.models
+
+if __name__ == '__main__':
+    root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))
+    train_pipeline(root_path)

video-tetalking/third_part/GFPGAN/gfpgan/utils.py

@@ -0,0 +1,143 @@
+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 GFPGANer():
+    """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, upscale=2, arch='clean', channel_multiplier=2, bg_upsampler=None):
+        self.upscale = upscale
+        self.bg_upsampler = bg_upsampler
+
+        # initialize model
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        # 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)
+        # initialize face helper
+        self.face_helper = FaceRestoreHelper(
+            upscale,
+            face_size=512,
+            crop_ratio=(1, 1),
+            det_model='retinaface_resnet50',
+            save_ext='png',
+            device=self.device)
+
+        if model_path.startswith('https://'):
+            model_path = load_file_from_url(
+                url=model_path, model_dir=os.path.join(ROOT_DIR, 'gfpgan/weights'), 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):
+        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)[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

video-tetalking/third_part/GFPGAN/gfpgan/version.py

@@ -0,0 +1,5 @@
+# GENERATED VERSION FILE
+# TIME: Wed Apr 20 14:43:06 2022
+__version__ = '1.3.2'
+__gitsha__ = '924ce47'
+version_info = (1, 3, 2)

video-tetalking/third_part/GFPGAN/gfpgan/weights/README.md

@@ -0,0 +1,3 @@
+# Weights
+
+Put the downloaded weights to this folder.

video-tetalking/third_part/GFPGAN/options/train_gfpgan_v1.yml

@@ -0,0 +1,216 @@
+# general settings
+name: train_GFPGANv1_512
+model_type: GFPGANModel
+num_gpu: auto  # officially, we use 4 GPUs
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: FFHQ
+    type: FFHQDegradationDataset
+    # dataroot_gt: datasets/ffhq/ffhq_512.lmdb
+    dataroot_gt: datasets/ffhq/ffhq_512
+    io_backend:
+      # type: lmdb
+      type: disk
+
+    use_hflip: true
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+    out_size: 512
+
+    blur_kernel_size: 41
+    kernel_list: ['iso', 'aniso']
+    kernel_prob: [0.5, 0.5]
+    blur_sigma: [0.1, 10]
+    downsample_range: [0.8, 8]
+    noise_range: [0, 20]
+    jpeg_range: [60, 100]
+
+    # color jitter and gray
+    color_jitter_prob: 0.3
+    color_jitter_shift: 20
+    color_jitter_pt_prob: 0.3
+    gray_prob: 0.01
+
+    # If you do not want colorization, please set
+    # color_jitter_prob: ~
+    # color_jitter_pt_prob: ~
+    # gray_prob: 0.01
+    # gt_gray: True
+
+    crop_components: true
+    component_path: experiments/pretrained_models/FFHQ_eye_mouth_landmarks_512.pth
+    eye_enlarge_ratio: 1.4
+
+    # data loader
+    use_shuffle: true
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 3
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  val:
+    # Please modify accordingly to use your own validation
+    # Or comment the val block if do not need validation during training
+    name: validation
+    type: PairedImageDataset
+    dataroot_lq: datasets/faces/validation/input
+    dataroot_gt: datasets/faces/validation/reference
+    io_backend:
+      type: disk
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+    scale: 1
+
+# network structures
+network_g:
+  type: GFPGANv1
+  out_size: 512
+  num_style_feat: 512
+  channel_multiplier: 1
+  resample_kernel: [1, 3, 3, 1]
+  decoder_load_path: experiments/pretrained_models/StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth
+  fix_decoder: true
+  num_mlp: 8
+  lr_mlp: 0.01
+  input_is_latent: true
+  different_w: true
+  narrow: 1
+  sft_half: true
+
+network_d:
+  type: StyleGAN2Discriminator
+  out_size: 512
+  channel_multiplier: 1
+  resample_kernel: [1, 3, 3, 1]
+
+network_d_left_eye:
+  type: FacialComponentDiscriminator
+
+network_d_right_eye:
+  type: FacialComponentDiscriminator
+
+network_d_mouth:
+  type: FacialComponentDiscriminator
+
+network_identity:
+  type: ResNetArcFace
+  block: IRBlock
+  layers: [2, 2, 2, 2]
+  use_se: False
+
+# path
+path:
+  pretrain_network_g: ~
+  param_key_g: params_ema
+  strict_load_g: ~
+  pretrain_network_d: ~
+  pretrain_network_d_left_eye: ~
+  pretrain_network_d_right_eye: ~
+  pretrain_network_d_mouth: ~
+  pretrain_network_identity: experiments/pretrained_models/arcface_resnet18.pth
+  # resume
+  resume_state: ~
+  ignore_resume_networks: ['network_identity']
+
+# training settings
+train:
+  optim_g:
+    type: Adam
+    lr: !!float 2e-3
+  optim_d:
+    type: Adam
+    lr: !!float 2e-3
+  optim_component:
+    type: Adam
+    lr: !!float 2e-3
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [600000, 700000]
+    gamma: 0.5
+
+  total_iter: 800000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  # pixel loss
+  pixel_opt:
+    type: L1Loss
+    loss_weight: !!float 1e-1
+    reduction: mean
+  # L1 loss used in pyramid loss, component style loss and identity loss
+  L1_opt:
+    type: L1Loss
+    loss_weight: 1
+    reduction: mean
+
+  # image pyramid loss
+  pyramid_loss_weight: 1
+  remove_pyramid_loss: 50000
+  # perceptual loss (content and style losses)
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      # before relu
+      'conv1_2': 0.1
+      'conv2_2': 0.1
+      'conv3_4': 1
+      'conv4_4': 1
+      'conv5_4': 1
+    vgg_type: vgg19
+    use_input_norm: true
+    perceptual_weight: !!float 1
+    style_weight: 50
+    range_norm: true
+    criterion: l1
+  # gan loss
+  gan_opt:
+    type: GANLoss
+    gan_type: wgan_softplus
+    loss_weight: !!float 1e-1
+  # r1 regularization for discriminator
+  r1_reg_weight: 10
+  # facial component loss
+  gan_component_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 1
+  comp_style_weight: 200
+  # identity loss
+  identity_weight: 10
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+  net_d_reg_every: 16
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: true
+
+  metrics:
+    psnr: # metric name
+      type: calculate_psnr
+      crop_border: 0
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500
+
+find_unused_parameters: true

video-tetalking/third_part/GFPGAN/options/train_gfpgan_v1_simple.yml

@@ -0,0 +1,182 @@
+# general settings
+name: train_GFPGANv1_512_simple
+model_type: GFPGANModel
+num_gpu: auto  # officially, we use 4 GPUs
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: FFHQ
+    type: FFHQDegradationDataset
+    # dataroot_gt: datasets/ffhq/ffhq_512.lmdb
+    dataroot_gt: datasets/ffhq/ffhq_512
+    io_backend:
+      # type: lmdb
+      type: disk
+
+    use_hflip: true
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+    out_size: 512
+
+    blur_kernel_size: 41
+    kernel_list: ['iso', 'aniso']
+    kernel_prob: [0.5, 0.5]
+    blur_sigma: [0.1, 10]
+    downsample_range: [0.8, 8]
+    noise_range: [0, 20]
+    jpeg_range: [60, 100]
+
+    # color jitter and gray
+    color_jitter_prob: 0.3
+    color_jitter_shift: 20
+    color_jitter_pt_prob: 0.3
+    gray_prob: 0.01
+
+    # If you do not want colorization, please set
+    # color_jitter_prob: ~
+    # color_jitter_pt_prob: ~
+    # gray_prob: 0.01
+    # gt_gray: True
+
+    # data loader
+    use_shuffle: true
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 3
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  val:
+    # Please modify accordingly to use your own validation
+    # Or comment the val block if do not need validation during training
+    name: validation
+    type: PairedImageDataset
+    dataroot_lq: datasets/faces/validation/input
+    dataroot_gt: datasets/faces/validation/reference
+    io_backend:
+      type: disk
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+    scale: 1
+
+# network structures
+network_g:
+  type: GFPGANv1
+  out_size: 512
+  num_style_feat: 512
+  channel_multiplier: 1
+  resample_kernel: [1, 3, 3, 1]
+  decoder_load_path: experiments/pretrained_models/StyleGAN2_512_Cmul1_FFHQ_B12G4_scratch_800k.pth
+  fix_decoder: true
+  num_mlp: 8
+  lr_mlp: 0.01
+  input_is_latent: true
+  different_w: true
+  narrow: 1
+  sft_half: true
+
+network_d:
+  type: StyleGAN2Discriminator
+  out_size: 512
+  channel_multiplier: 1
+  resample_kernel: [1, 3, 3, 1]
+
+
+# path
+path:
+  pretrain_network_g: ~
+  param_key_g: params_ema
+  strict_load_g: ~
+  pretrain_network_d: ~
+  resume_state: ~
+
+# training settings
+train:
+  optim_g:
+    type: Adam
+    lr: !!float 2e-3
+  optim_d:
+    type: Adam
+    lr: !!float 2e-3
+  optim_component:
+    type: Adam
+    lr: !!float 2e-3
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [600000, 700000]
+    gamma: 0.5
+
+  total_iter: 800000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  # pixel loss
+  pixel_opt:
+    type: L1Loss
+    loss_weight: !!float 1e-1
+    reduction: mean
+  # L1 loss used in pyramid loss, component style loss and identity loss
+  L1_opt:
+    type: L1Loss
+    loss_weight: 1
+    reduction: mean
+
+  # image pyramid loss
+  pyramid_loss_weight: 1
+  remove_pyramid_loss: 50000
+  # perceptual loss (content and style losses)
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      # before relu
+      'conv1_2': 0.1
+      'conv2_2': 0.1
+      'conv3_4': 1
+      'conv4_4': 1
+      'conv5_4': 1
+    vgg_type: vgg19
+    use_input_norm: true
+    perceptual_weight: !!float 1
+    style_weight: 50
+    range_norm: true
+    criterion: l1
+  # gan loss
+  gan_opt:
+    type: GANLoss
+    gan_type: wgan_softplus
+    loss_weight: !!float 1e-1
+  # r1 regularization for discriminator
+  r1_reg_weight: 10
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+  net_d_reg_every: 16
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: true
+
+  metrics:
+    psnr: # metric name
+      type: calculate_psnr
+      crop_border: 0
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500
+
+find_unused_parameters: true

video-tetalking/third_part/GPEN/align_faces.py

@@ -0,0 +1,271 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Mon Apr 24 15:43:29 2017
+@author: zhaoy
+"""
+"""
+@Modified by yangxy (yangtao9009@gmail.com)
+"""
+import cv2
+import numpy as np
+from skimage import transform as trans
+
+# reference facial points, a list of coordinates (x,y)
+REFERENCE_FACIAL_POINTS = [
+    [30.29459953, 51.69630051],
+    [65.53179932, 51.50139999],
+    [48.02519989, 71.73660278],
+    [33.54930115, 92.3655014],
+    [62.72990036, 92.20410156]
+]
+
+DEFAULT_CROP_SIZE = (96, 112)
+
+
+def _umeyama(src, dst, estimate_scale=True, scale=1.0):
+    """Estimate N-D similarity transformation with or without scaling.
+    Parameters
+    ----------
+    src : (M, N) array
+        Source coordinates.
+    dst : (M, N) array
+        Destination coordinates.
+    estimate_scale : bool
+        Whether to estimate scaling factor.
+    Returns
+    -------
+    T : (N + 1, N + 1)
+        The homogeneous similarity transformation matrix. The matrix contains
+        NaN values only if the problem is not well-conditioned.
+    References
+    ----------
+    .. [1] "Least-squares estimation of transformation parameters between two
+            point patterns", Shinji Umeyama, PAMI 1991, :DOI:`10.1109/34.88573`
+    """
+
+    num = src.shape[0]
+    dim = src.shape[1]
+
+    # Compute mean of src and dst.
+    src_mean = src.mean(axis=0)
+    dst_mean = dst.mean(axis=0)
+
+    # Subtract mean from src and dst.
+    src_demean = src - src_mean
+    dst_demean = dst - dst_mean
+
+    # Eq. (38).
+    A = dst_demean.T @ src_demean / num
+
+    # Eq. (39).
+    d = np.ones((dim,), dtype=np.double)
+    if np.linalg.det(A) < 0:
+        d[dim - 1] = -1
+
+    T = np.eye(dim + 1, dtype=np.double)
+
+    U, S, V = np.linalg.svd(A)
+
+    # Eq. (40) and (43).
+    rank = np.linalg.matrix_rank(A)
+    if rank == 0:
+        return np.nan * T
+    elif rank == dim - 1:
+        if np.linalg.det(U) * np.linalg.det(V) > 0:
+            T[:dim, :dim] = U @ V
+        else:
+            s = d[dim - 1]
+            d[dim - 1] = -1
+            T[:dim, :dim] = U @ np.diag(d) @ V
+            d[dim - 1] = s
+    else:
+        T[:dim, :dim] = U @ np.diag(d) @ V
+
+    if estimate_scale:
+        # Eq. (41) and (42).
+        scale = 1.0 / src_demean.var(axis=0).sum() * (S @ d)
+    else:
+        scale = scale
+
+    T[:dim, dim] = dst_mean - scale * (T[:dim, :dim] @ src_mean.T)
+    T[:dim, :dim] *= scale
+
+    return T, scale
+
+
+class FaceWarpException(Exception):
+    def __str__(self):
+        return 'In File {}:{}'.format(
+            __file__, super.__str__(self))
+
+
+def get_reference_facial_points(output_size=None,
+                                inner_padding_factor=0.0,
+                                outer_padding=(0, 0),
+                                default_square=False):
+    tmp_5pts = np.array(REFERENCE_FACIAL_POINTS)
+    tmp_crop_size = np.array(DEFAULT_CROP_SIZE)
+
+    # 0) make the inner region a square
+    if default_square:
+        size_diff = max(tmp_crop_size) - tmp_crop_size
+        tmp_5pts += size_diff / 2
+        tmp_crop_size += size_diff
+
+    if (output_size and
+            output_size[0] == tmp_crop_size[0] and
+            output_size[1] == tmp_crop_size[1]):
+        print('output_size == DEFAULT_CROP_SIZE {}: return default reference points'.format(tmp_crop_size))
+        return tmp_5pts
+
+    if (inner_padding_factor == 0 and
+            outer_padding == (0, 0)):
+        if output_size is None:
+            print('No paddings to do: return default reference points')
+            return tmp_5pts
+        else:
+            raise FaceWarpException(
+                'No paddings to do, output_size must be None or {}'.format(tmp_crop_size))
+
+    # check output size
+    if not (0 <= inner_padding_factor <= 1.0):
+        raise FaceWarpException('Not (0 <= inner_padding_factor <= 1.0)')
+
+    if ((inner_padding_factor > 0 or outer_padding[0] > 0 or outer_padding[1] > 0)
+            and output_size is None):
+        output_size = tmp_crop_size * \
+                      (1 + inner_padding_factor * 2).astype(np.int32)
+        output_size += np.array(outer_padding)
+        print('              deduced from paddings, output_size = ', output_size)
+
+    if not (outer_padding[0] < output_size[0]
+            and outer_padding[1] < output_size[1]):
+        raise FaceWarpException('Not (outer_padding[0] < output_size[0]'
+                                'and outer_padding[1] < output_size[1])')
+
+    # 1) pad the inner region according inner_padding_factor
+    # print('---> STEP1: pad the inner region according inner_padding_factor')
+    if inner_padding_factor > 0:
+        size_diff = tmp_crop_size * inner_padding_factor * 2
+        tmp_5pts += size_diff / 2
+        tmp_crop_size += np.round(size_diff).astype(np.int32)
+
+    # print('              crop_size = ', tmp_crop_size)
+    # print('              reference_5pts = ', tmp_5pts)
+
+    # 2) resize the padded inner region
+    # print('---> STEP2: resize the padded inner region')
+    size_bf_outer_pad = np.array(output_size) - np.array(outer_padding) * 2
+    # print('              crop_size = ', tmp_crop_size)
+    # print('              size_bf_outer_pad = ', size_bf_outer_pad)
+
+    if size_bf_outer_pad[0] * tmp_crop_size[1] != size_bf_outer_pad[1] * tmp_crop_size[0]:
+        raise FaceWarpException('Must have (output_size - outer_padding)'
+                                '= some_scale * (crop_size * (1.0 + inner_padding_factor)')
+
+    scale_factor = size_bf_outer_pad[0].astype(np.float32) / tmp_crop_size[0]
+    # print('              resize scale_factor = ', scale_factor)
+    tmp_5pts = tmp_5pts * scale_factor
+    #    size_diff = tmp_crop_size * (scale_factor - min(scale_factor))
+    #    tmp_5pts = tmp_5pts + size_diff / 2
+    tmp_crop_size = size_bf_outer_pad
+    # print('              crop_size = ', tmp_crop_size)
+    # print('              reference_5pts = ', tmp_5pts)
+
+    # 3) add outer_padding to make output_size
+    reference_5point = tmp_5pts + np.array(outer_padding)
+    tmp_crop_size = output_size
+    # print('---> STEP3: add outer_padding to make output_size')
+    # print('              crop_size = ', tmp_crop_size)
+    # print('              reference_5pts = ', tmp_5pts)
+    #
+    # print('===> end get_reference_facial_points\n')
+
+    return reference_5point
+
+
+def get_affine_transform_matrix(src_pts, dst_pts):
+    tfm = np.float32([[1, 0, 0], [0, 1, 0]])
+    n_pts = src_pts.shape[0]
+    ones = np.ones((n_pts, 1), src_pts.dtype)
+    src_pts_ = np.hstack([src_pts, ones])
+    dst_pts_ = np.hstack([dst_pts, ones])
+
+    A, res, rank, s = np.linalg.lstsq(src_pts_, dst_pts_)
+
+    if rank == 3:
+        tfm = np.float32([
+            [A[0, 0], A[1, 0], A[2, 0]],
+            [A[0, 1], A[1, 1], A[2, 1]]
+        ])
+    elif rank == 2:
+        tfm = np.float32([
+            [A[0, 0], A[1, 0], 0],
+            [A[0, 1], A[1, 1], 0]
+        ])
+
+    return tfm
+
+
+def warp_and_crop_face(src_img,
+                       facial_pts,
+                       reference_pts=None,
+                       crop_size=(96, 112),
+                       align_type='smilarity'): #smilarity cv2_affine affine
+    if reference_pts is None:
+        if crop_size[0] == 96 and crop_size[1] == 112:
+            reference_pts = REFERENCE_FACIAL_POINTS
+        else:
+            default_square = False
+            inner_padding_factor = 0
+            outer_padding = (0, 0)
+            output_size = crop_size
+
+            reference_pts = get_reference_facial_points(output_size,
+                                                        inner_padding_factor,
+                                                        outer_padding,
+                                                        default_square)
+
+    ref_pts = np.float32(reference_pts)
+    ref_pts_shp = ref_pts.shape
+    if max(ref_pts_shp) < 3: #  or min(ref_pts_shp) != 2:
+        raise FaceWarpException(
+            'reference_pts.shape must be (K,2) or (2,K) and K>2')
+
+    if ref_pts_shp[0] == 2 or ref_pts_shp[0] == 3:
+        ref_pts = ref_pts.T
+
+    src_pts = np.float32(facial_pts)
+    src_pts_shp = src_pts.shape
+    if max(src_pts_shp) < 3: # or min(src_pts_shp) != 2:
+        raise FaceWarpException(
+            'facial_pts.shape must be (K,2) or (2,K) and K>2')
+
+    if src_pts_shp[0] == 2 or src_pts_shp[0] == 3:
+        src_pts = src_pts.T
+
+    if src_pts.shape != ref_pts.shape:
+        raise FaceWarpException(
+            'facial_pts and reference_pts must have the same shape')
+
+    if align_type is 'cv2_affine':
+        tfm = cv2.getAffineTransform(src_pts[0:3], ref_pts[0:3])
+        tfm_inv = cv2.getAffineTransform(ref_pts[0:3], src_pts[0:3])
+    elif align_type is 'cv2_rigid':
+        tfm, _ = cv2.estimateAffinePartial2D(src_pts[0:3], ref_pts[0:3])
+        tfm_inv, _ = cv2.estimateAffinePartial2D(ref_pts[0:3], src_pts[0:3])
+    elif align_type is 'affine':
+        tfm = get_affine_transform_matrix(src_pts, ref_pts)
+        tfm_inv = get_affine_transform_matrix(ref_pts, src_pts)
+    else:
+        params, scale = _umeyama(src_pts, ref_pts)
+        tfm = params[:2, :]
+
+        params, _ = _umeyama(ref_pts, src_pts, False, scale=1.0/scale)
+        tfm_inv = params[:2, :]
+
+    # M = cv2.getPerspectiveTransform(ref_pts[0:4], src_pts[0:4])
+    face_img = cv2.warpAffine(src_img, tfm, (crop_size[0], crop_size[1]), flags=3)
+    # face_img = cv2.warpPerspective(src_img, M, (crop_size[0], crop_size[1]), flags=cv2.INTER_LINEAR )
+
+    return face_img, tfm_inv

video-tetalking/third_part/GPEN/face_detect/data/FDDB/img_list.txt

@@ -0,0 +1,2845 @@
+2002/08/11/big/img_591
+2002/08/26/big/img_265
+2002/07/19/big/img_423
+2002/08/24/big/img_490
+2002/08/31/big/img_17676
+2002/07/31/big/img_228
+2002/07/24/big/img_402
+2002/08/04/big/img_769
+2002/07/19/big/img_581
+2002/08/13/big/img_723
+2002/08/12/big/img_821
+2003/01/17/big/img_610
+2002/08/13/big/img_1116
+2002/08/28/big/img_19238
+2002/08/21/big/img_660
+2002/08/14/big/img_607
+2002/08/05/big/img_3708
+2002/08/19/big/img_511
+2002/08/07/big/img_1316
+2002/07/25/big/img_1047
+2002/07/23/big/img_474
+2002/07/27/big/img_970
+2002/09/02/big/img_15752
+2002/09/01/big/img_16378
+2002/09/01/big/img_16189
+2002/08/26/big/img_276
+2002/07/24/big/img_518
+2002/08/14/big/img_1027
+2002/08/24/big/img_733
+2002/08/15/big/img_249
+2003/01/15/big/img_1371
+2002/08/07/big/img_1348
+2003/01/01/big/img_331
+2002/08/23/big/img_536
+2002/07/30/big/img_224
+2002/08/10/big/img_763
+2002/08/21/big/img_293
+2002/08/15/big/img_1211
+2002/08/15/big/img_1194
+2003/01/15/big/img_390
+2002/08/06/big/img_2893
+2002/08/17/big/img_691
+2002/08/07/big/img_1695
+2002/08/16/big/img_829
+2002/07/25/big/img_201
+2002/08/23/big/img_36
+2003/01/15/big/img_763
+2003/01/15/big/img_637
+2002/08/22/big/img_592
+2002/07/25/big/img_817
+2003/01/15/big/img_1219
+2002/08/05/big/img_3508
+2002/08/15/big/img_1108
+2002/07/19/big/img_488
+2003/01/16/big/img_704
+2003/01/13/big/img_1087
+2002/08/10/big/img_670
+2002/07/24/big/img_104
+2002/08/27/big/img_19823
+2002/09/01/big/img_16229
+2003/01/13/big/img_846
+2002/08/04/big/img_412
+2002/07/22/big/img_554
+2002/08/12/big/img_331
+2002/08/02/big/img_533
+2002/08/12/big/img_259
+2002/08/18/big/img_328
+2003/01/14/big/img_630
+2002/08/05/big/img_3541
+2002/08/06/big/img_2390
+2002/08/20/big/img_150
+2002/08/02/big/img_1231
+2002/08/16/big/img_710
+2002/08/19/big/img_591
+2002/07/22/big/img_725
+2002/07/24/big/img_820
+2003/01/13/big/img_568
+2002/08/22/big/img_853
+2002/08/09/big/img_648
+2002/08/23/big/img_528
+2003/01/14/big/img_888
+2002/08/30/big/img_18201
+2002/08/13/big/img_965
+2003/01/14/big/img_660
+2002/07/19/big/img_517
+2003/01/14/big/img_406
+2002/08/30/big/img_18433
+2002/08/07/big/img_1630
+2002/08/06/big/img_2717
+2002/08/21/big/img_470
+2002/07/23/big/img_633
+2002/08/20/big/img_915
+2002/08/16/big/img_893
+2002/07/29/big/img_644
+2002/08/15/big/img_529
+2002/08/16/big/img_668
+2002/08/07/big/img_1871
+2002/07/25/big/img_192
+2002/07/31/big/img_961
+2002/08/19/big/img_738
+2002/07/31/big/img_382
+2002/08/19/big/img_298
+2003/01/17/big/img_608
+2002/08/21/big/img_514
+2002/07/23/big/img_183
+2003/01/17/big/img_536
+2002/07/24/big/img_478
+2002/08/06/big/img_2997
+2002/09/02/big/img_15380
+2002/08/07/big/img_1153
+2002/07/31/big/img_967
+2002/07/31/big/img_711
+2002/08/26/big/img_664
+2003/01/01/big/img_326
+2002/08/24/big/img_775
+2002/08/08/big/img_961
+2002/08/16/big/img_77
+2002/08/12/big/img_296
+2002/07/22/big/img_905
+2003/01/13/big/img_284
+2002/08/13/big/img_887
+2002/08/24/big/img_849
+2002/07/30/big/img_345
+2002/08/18/big/img_419
+2002/08/01/big/img_1347
+2002/08/05/big/img_3670
+2002/07/21/big/img_479
+2002/08/08/big/img_913
+2002/09/02/big/img_15828
+2002/08/30/big/img_18194
+2002/08/08/big/img_471
+2002/08/22/big/img_734
+2002/08/09/big/img_586
+2002/08/09/big/img_454
+2002/07/29/big/img_47
+2002/07/19/big/img_381
+2002/07/29/big/img_733
+2002/08/20/big/img_327
+2002/07/21/big/img_96
+2002/08/06/big/img_2680
+2002/07/25/big/img_919
+2002/07/21/big/img_158
+2002/07/22/big/img_801
+2002/07/22/big/img_567
+2002/07/24/big/img_804
+2002/07/24/big/img_690
+2003/01/15/big/img_576
+2002/08/14/big/img_335
+2003/01/13/big/img_390
+2002/08/11/big/img_258
+2002/07/23/big/img_917
+2002/08/15/big/img_525
+2003/01/15/big/img_505
+2002/07/30/big/img_886
+2003/01/16/big/img_640
+2003/01/14/big/img_642
+2003/01/17/big/img_844
+2002/08/04/big/img_571
+2002/08/29/big/img_18702
+2003/01/15/big/img_240
+2002/07/29/big/img_553
+2002/08/10/big/img_354
+2002/08/18/big/img_17
+2003/01/15/big/img_782
+2002/07/27/big/img_382
+2002/08/14/big/img_970
+2003/01/16/big/img_70
+2003/01/16/big/img_625
+2002/08/18/big/img_341
+2002/08/26/big/img_188
+2002/08/09/big/img_405
+2002/08/02/big/img_37
+2002/08/13/big/img_748
+2002/07/22/big/img_399
+2002/07/25/big/img_844
+2002/08/12/big/img_340
+2003/01/13/big/img_815
+2002/08/26/big/img_5
+2002/08/10/big/img_158
+2002/08/18/big/img_95
+2002/07/29/big/img_1297
+2003/01/13/big/img_508
+2002/09/01/big/img_16680
+2003/01/16/big/img_338
+2002/08/13/big/img_517
+2002/07/22/big/img_626
+2002/08/06/big/img_3024
+2002/07/26/big/img_499
+2003/01/13/big/img_387
+2002/08/31/big/img_18025
+2002/08/13/big/img_520
+2003/01/16/big/img_576
+2002/07/26/big/img_121
+2002/08/25/big/img_703
+2002/08/26/big/img_615
+2002/08/17/big/img_434
+2002/08/02/big/img_677
+2002/08/18/big/img_276
+2002/08/05/big/img_3672
+2002/07/26/big/img_700
+2002/07/31/big/img_277
+2003/01/14/big/img_220
+2002/08/23/big/img_232
+2002/08/31/big/img_17422
+2002/07/22/big/img_508
+2002/08/13/big/img_681
+2003/01/15/big/img_638
+2002/08/30/big/img_18408
+2003/01/14/big/img_533
+2003/01/17/big/img_12
+2002/08/28/big/img_19388
+2002/08/08/big/img_133
+2002/07/26/big/img_885
+2002/08/19/big/img_387
+2002/08/27/big/img_19976
+2002/08/26/big/img_118
+2002/08/28/big/img_19146
+2002/08/05/big/img_3259
+2002/08/15/big/img_536
+2002/07/22/big/img_279
+2002/07/22/big/img_9
+2002/08/13/big/img_301
+2002/08/15/big/img_974
+2002/08/06/big/img_2355
+2002/08/01/big/img_1526
+2002/08/03/big/img_417
+2002/08/04/big/img_407
+2002/08/15/big/img_1029
+2002/07/29/big/img_700
+2002/08/01/big/img_1463
+2002/08/31/big/img_17365
+2002/07/28/big/img_223
+2002/07/19/big/img_827
+2002/07/27/big/img_531
+2002/07/19/big/img_845
+2002/08/20/big/img_382
+2002/07/31/big/img_268
+2002/08/27/big/img_19705
+2002/08/02/big/img_830
+2002/08/23/big/img_250
+2002/07/20/big/img_777
+2002/08/21/big/img_879
+2002/08/26/big/img_20146
+2002/08/23/big/img_789
+2002/08/06/big/img_2683
+2002/08/25/big/img_576
+2002/08/09/big/img_498
+2002/08/08/big/img_384
+2002/08/26/big/img_592
+2002/07/29/big/img_1470
+2002/08/21/big/img_452
+2002/08/30/big/img_18395
+2002/08/15/big/img_215
+2002/07/21/big/img_643
+2002/07/22/big/img_209
+2003/01/17/big/img_346
+2002/08/25/big/img_658
+2002/08/21/big/img_221
+2002/08/14/big/img_60
+2003/01/17/big/img_885
+2003/01/16/big/img_482
+2002/08/19/big/img_593
+2002/08/08/big/img_233
+2002/07/30/big/img_458
+2002/07/23/big/img_384
+2003/01/15/big/img_670
+2003/01/15/big/img_267
+2002/08/26/big/img_540
+2002/07/29/big/img_552
+2002/07/30/big/img_997
+2003/01/17/big/img_377
+2002/08/21/big/img_265
+2002/08/09/big/img_561
+2002/07/31/big/img_945
+2002/09/02/big/img_15252
+2002/08/11/big/img_276
+2002/07/22/big/img_491
+2002/07/26/big/img_517
+2002/08/14/big/img_726
+2002/08/08/big/img_46
+2002/08/28/big/img_19458
+2002/08/06/big/img_2935
+2002/07/29/big/img_1392
+2002/08/13/big/img_776
+2002/08/24/big/img_616
+2002/08/14/big/img_1065
+2002/07/29/big/img_889
+2002/08/18/big/img_188
+2002/08/07/big/img_1453
+2002/08/02/big/img_760
+2002/07/28/big/img_416
+2002/08/07/big/img_1393
+2002/08/26/big/img_292
+2002/08/26/big/img_301
+2003/01/13/big/img_195
+2002/07/26/big/img_532
+2002/08/20/big/img_550
+2002/08/05/big/img_3658
+2002/08/26/big/img_738
+2002/09/02/big/img_15750
+2003/01/17/big/img_451
+2002/07/23/big/img_339
+2002/08/16/big/img_637
+2002/08/14/big/img_748
+2002/08/06/big/img_2739
+2002/07/25/big/img_482
+2002/08/19/big/img_191
+2002/08/26/big/img_537
+2003/01/15/big/img_716
+2003/01/15/big/img_767
+2002/08/02/big/img_452
+2002/08/08/big/img_1011
+2002/08/10/big/img_144
+2003/01/14/big/img_122
+2002/07/24/big/img_586
+2002/07/24/big/img_762
+2002/08/20/big/img_369
+2002/07/30/big/img_146
+2002/08/23/big/img_396
+2003/01/15/big/img_200
+2002/08/15/big/img_1183
+2003/01/14/big/img_698
+2002/08/09/big/img_792
+2002/08/06/big/img_2347
+2002/07/31/big/img_911
+2002/08/26/big/img_722
+2002/08/23/big/img_621
+2002/08/05/big/img_3790
+2003/01/13/big/img_633
+2002/08/09/big/img_224
+2002/07/24/big/img_454
+2002/07/21/big/img_202
+2002/08/02/big/img_630
+2002/08/30/big/img_18315
+2002/07/19/big/img_491
+2002/09/01/big/img_16456
+2002/08/09/big/img_242
+2002/07/25/big/img_595
+2002/07/22/big/img_522
+2002/08/01/big/img_1593
+2002/07/29/big/img_336
+2002/08/15/big/img_448
+2002/08/28/big/img_19281
+2002/07/29/big/img_342
+2002/08/12/big/img_78
+2003/01/14/big/img_525
+2002/07/28/big/img_147
+2002/08/11/big/img_353
+2002/08/22/big/img_513
+2002/08/04/big/img_721
+2002/08/17/big/img_247
+2003/01/14/big/img_891
+2002/08/20/big/img_853
+2002/07/19/big/img_414
+2002/08/01/big/img_1530
+2003/01/14/big/img_924
+2002/08/22/big/img_468
+2002/08/18/big/img_354
+2002/08/30/big/img_18193
+2002/08/23/big/img_492
+2002/08/15/big/img_871
+2002/08/12/big/img_494
+2002/08/06/big/img_2470
+2002/07/23/big/img_923
+2002/08/26/big/img_155
+2002/08/08/big/img_669
+2002/07/23/big/img_404
+2002/08/28/big/img_19421
+2002/08/29/big/img_18993
+2002/08/25/big/img_416
+2003/01/17/big/img_434
+2002/07/29/big/img_1370
+2002/07/28/big/img_483
+2002/08/11/big/img_50
+2002/08/10/big/img_404
+2002/09/02/big/img_15057
+2003/01/14/big/img_911
+2002/09/01/big/img_16697
+2003/01/16/big/img_665
+2002/09/01/big/img_16708
+2002/08/22/big/img_612
+2002/08/28/big/img_19471
+2002/08/02/big/img_198
+2003/01/16/big/img_527
+2002/08/22/big/img_209
+2002/08/30/big/img_18205
+2003/01/14/big/img_114
+2003/01/14/big/img_1028
+2003/01/16/big/img_894
+2003/01/14/big/img_837
+2002/07/30/big/img_9
+2002/08/06/big/img_2821
+2002/08/04/big/img_85
+2003/01/13/big/img_884
+2002/07/22/big/img_570
+2002/08/07/big/img_1773
+2002/07/26/big/img_208
+2003/01/17/big/img_946
+2002/07/19/big/img_930
+2003/01/01/big/img_698
+2003/01/17/big/img_612
+2002/07/19/big/img_372
+2002/07/30/big/img_721
+2003/01/14/big/img_649
+2002/08/19/big/img_4
+2002/07/25/big/img_1024
+2003/01/15/big/img_601
+2002/08/30/big/img_18470
+2002/07/22/big/img_29
+2002/08/07/big/img_1686
+2002/07/20/big/img_294
+2002/08/14/big/img_800
+2002/08/19/big/img_353
+2002/08/19/big/img_350
+2002/08/05/big/img_3392
+2002/08/09/big/img_622
+2003/01/15/big/img_236
+2002/08/11/big/img_643
+2002/08/05/big/img_3458
+2002/08/12/big/img_413
+2002/08/22/big/img_415
+2002/08/13/big/img_635
+2002/08/07/big/img_1198
+2002/08/04/big/img_873
+2002/08/12/big/img_407
+2003/01/15/big/img_346
+2002/08/02/big/img_275
+2002/08/17/big/img_997
+2002/08/21/big/img_958
+2002/08/20/big/img_579
+2002/07/29/big/img_142
+2003/01/14/big/img_1115
+2002/08/16/big/img_365
+2002/07/29/big/img_1414
+2002/08/17/big/img_489
+2002/08/13/big/img_1010
+2002/07/31/big/img_276
+2002/07/25/big/img_1000
+2002/08/23/big/img_524
+2002/08/28/big/img_19147
+2003/01/13/big/img_433
+2002/08/20/big/img_205
+2003/01/01/big/img_458
+2002/07/29/big/img_1449
+2003/01/16/big/img_696
+2002/08/28/big/img_19296
+2002/08/29/big/img_18688
+2002/08/21/big/img_767
+2002/08/20/big/img_532
+2002/08/26/big/img_187
+2002/07/26/big/img_183
+2002/07/27/big/img_890
+2003/01/13/big/img_576
+2002/07/30/big/img_15
+2002/07/31/big/img_889
+2002/08/31/big/img_17759
+2003/01/14/big/img_1114
+2002/07/19/big/img_445
+2002/08/03/big/img_593
+2002/07/24/big/img_750
+2002/07/30/big/img_133
+2002/08/25/big/img_671
+2002/07/20/big/img_351
+2002/08/31/big/img_17276
+2002/08/05/big/img_3231
+2002/09/02/big/img_15882
+2002/08/14/big/img_115
+2002/08/02/big/img_1148
+2002/07/25/big/img_936
+2002/07/31/big/img_639
+2002/08/04/big/img_427
+2002/08/22/big/img_843
+2003/01/17/big/img_17
+2003/01/13/big/img_690
+2002/08/13/big/img_472
+2002/08/09/big/img_425
+2002/08/05/big/img_3450
+2003/01/17/big/img_439
+2002/08/13/big/img_539
+2002/07/28/big/img_35
+2002/08/16/big/img_241
+2002/08/06/big/img_2898
+2003/01/16/big/img_429
+2002/08/05/big/img_3817
+2002/08/27/big/img_19919
+2002/07/19/big/img_422
+2002/08/15/big/img_560
+2002/07/23/big/img_750
+2002/07/30/big/img_353
+2002/08/05/big/img_43
+2002/08/23/big/img_305
+2002/08/01/big/img_2137
+2002/08/30/big/img_18097
+2002/08/01/big/img_1389
+2002/08/02/big/img_308
+2003/01/14/big/img_652
+2002/08/01/big/img_1798
+2003/01/14/big/img_732
+2003/01/16/big/img_294
+2002/08/26/big/img_213
+2002/07/24/big/img_842
+2003/01/13/big/img_630
+2003/01/13/big/img_634
+2002/08/06/big/img_2285
+2002/08/01/big/img_2162
+2002/08/30/big/img_18134
+2002/08/02/big/img_1045
+2002/08/01/big/img_2143
+2002/07/25/big/img_135
+2002/07/20/big/img_645
+2002/08/05/big/img_3666
+2002/08/14/big/img_523
+2002/08/04/big/img_425
+2003/01/14/big/img_137
+2003/01/01/big/img_176
+2002/08/15/big/img_505
+2002/08/24/big/img_386
+2002/08/05/big/img_3187
+2002/08/15/big/img_419
+2003/01/13/big/img_520
+2002/08/04/big/img_444
+2002/08/26/big/img_483
+2002/08/05/big/img_3449
+2002/08/30/big/img_18409
+2002/08/28/big/img_19455
+2002/08/27/big/img_20090
+2002/07/23/big/img_625
+2002/08/24/big/img_205
+2002/08/08/big/img_938
+2003/01/13/big/img_527
+2002/08/07/big/img_1712
+2002/07/24/big/img_801
+2002/08/09/big/img_579
+2003/01/14/big/img_41
+2003/01/15/big/img_1130
+2002/07/21/big/img_672
+2002/08/07/big/img_1590
+2003/01/01/big/img_532
+2002/08/02/big/img_529
+2002/08/05/big/img_3591
+2002/08/23/big/img_5
+2003/01/14/big/img_882
+2002/08/28/big/img_19234
+2002/07/24/big/img_398
+2003/01/14/big/img_592
+2002/08/22/big/img_548
+2002/08/12/big/img_761
+2003/01/16/big/img_497
+2002/08/18/big/img_133
+2002/08/08/big/img_874
+2002/07/19/big/img_247
+2002/08/15/big/img_170
+2002/08/27/big/img_19679
+2002/08/20/big/img_246
+2002/08/24/big/img_358
+2002/07/29/big/img_599
+2002/08/01/big/img_1555
+2002/07/30/big/img_491
+2002/07/30/big/img_371
+2003/01/16/big/img_682
+2002/07/25/big/img_619
+2003/01/15/big/img_587
+2002/08/02/big/img_1212
+2002/08/01/big/img_2152
+2002/07/25/big/img_668
+2003/01/16/big/img_574
+2002/08/28/big/img_19464
+2002/08/11/big/img_536
+2002/07/24/big/img_201
+2002/08/05/big/img_3488
+2002/07/25/big/img_887
+2002/07/22/big/img_789
+2002/07/30/big/img_432
+2002/08/16/big/img_166
+2002/09/01/big/img_16333
+2002/07/26/big/img_1010
+2002/07/21/big/img_793
+2002/07/22/big/img_720
+2002/07/31/big/img_337
+2002/07/27/big/img_185
+2002/08/23/big/img_440
+2002/07/31/big/img_801
+2002/07/25/big/img_478
+2003/01/14/big/img_171
+2002/08/07/big/img_1054
+2002/09/02/big/img_15659
+2002/07/29/big/img_1348
+2002/08/09/big/img_337
+2002/08/26/big/img_684
+2002/07/31/big/img_537
+2002/08/15/big/img_808
+2003/01/13/big/img_740
+2002/08/07/big/img_1667
+2002/08/03/big/img_404
+2002/08/06/big/img_2520
+2002/07/19/big/img_230
+2002/07/19/big/img_356
+2003/01/16/big/img_627
+2002/08/04/big/img_474
+2002/07/29/big/img_833
+2002/07/25/big/img_176
+2002/08/01/big/img_1684
+2002/08/21/big/img_643
+2002/08/27/big/img_19673
+2002/08/02/big/img_838
+2002/08/06/big/img_2378
+2003/01/15/big/img_48
+2002/07/30/big/img_470
+2002/08/15/big/img_963
+2002/08/24/big/img_444
+2002/08/16/big/img_662
+2002/08/15/big/img_1209
+2002/07/24/big/img_25
+2002/08/06/big/img_2740
+2002/07/29/big/img_996
+2002/08/31/big/img_18074
+2002/08/04/big/img_343
+2003/01/17/big/img_509
+2003/01/13/big/img_726
+2002/08/07/big/img_1466
+2002/07/26/big/img_307
+2002/08/10/big/img_598
+2002/08/13/big/img_890
+2002/08/14/big/img_997
+2002/07/19/big/img_392
+2002/08/02/big/img_475
+2002/08/29/big/img_19038
+2002/07/29/big/img_538
+2002/07/29/big/img_502
+2002/08/02/big/img_364
+2002/08/31/big/img_17353
+2002/08/08/big/img_539
+2002/08/01/big/img_1449
+2002/07/22/big/img_363
+2002/08/02/big/img_90
+2002/09/01/big/img_16867
+2002/08/05/big/img_3371
+2002/07/30/big/img_342
+2002/08/07/big/img_1363
+2002/08/22/big/img_790
+2003/01/15/big/img_404
+2002/08/05/big/img_3447
+2002/09/01/big/img_16167
+2003/01/13/big/img_840
+2002/08/22/big/img_1001
+2002/08/09/big/img_431
+2002/07/27/big/img_618
+2002/07/31/big/img_741
+2002/07/30/big/img_964
+2002/07/25/big/img_86
+2002/07/29/big/img_275
+2002/08/21/big/img_921
+2002/07/26/big/img_892
+2002/08/21/big/img_663
+2003/01/13/big/img_567
+2003/01/14/big/img_719
+2002/07/28/big/img_251
+2003/01/15/big/img_1123
+2002/07/29/big/img_260
+2002/08/24/big/img_337
+2002/08/01/big/img_1914
+2002/08/13/big/img_373
+2003/01/15/big/img_589
+2002/08/13/big/img_906
+2002/07/26/big/img_270
+2002/08/26/big/img_313
+2002/08/25/big/img_694
+2003/01/01/big/img_327
+2002/07/23/big/img_261
+2002/08/26/big/img_642
+2002/07/29/big/img_918
+2002/07/23/big/img_455
+2002/07/24/big/img_612
+2002/07/23/big/img_534
+2002/07/19/big/img_534
+2002/07/19/big/img_726
+2002/08/01/big/img_2146
+2002/08/02/big/img_543
+2003/01/16/big/img_777
+2002/07/30/big/img_484
+2002/08/13/big/img_1161
+2002/07/21/big/img_390
+2002/08/06/big/img_2288
+2002/08/21/big/img_677
+2002/08/13/big/img_747
+2002/08/15/big/img_1248
+2002/07/31/big/img_416
+2002/09/02/big/img_15259
+2002/08/16/big/img_781
+2002/08/24/big/img_754
+2002/07/24/big/img_803
+2002/08/20/big/img_609
+2002/08/28/big/img_19571
+2002/09/01/big/img_16140
+2002/08/26/big/img_769
+2002/07/20/big/img_588
+2002/08/02/big/img_898
+2002/07/21/big/img_466
+2002/08/14/big/img_1046
+2002/07/25/big/img_212
+2002/08/26/big/img_353
+2002/08/19/big/img_810
+2002/08/31/big/img_17824
+2002/08/12/big/img_631
+2002/07/19/big/img_828
+2002/07/24/big/img_130
+2002/08/25/big/img_580
+2002/07/31/big/img_699
+2002/07/23/big/img_808
+2002/07/31/big/img_377
+2003/01/16/big/img_570
+2002/09/01/big/img_16254
+2002/07/21/big/img_471
+2002/08/01/big/img_1548
+2002/08/18/big/img_252
+2002/08/19/big/img_576
+2002/08/20/big/img_464
+2002/07/27/big/img_735
+2002/08/21/big/img_589
+2003/01/15/big/img_1192
+2002/08/09/big/img_302
+2002/07/31/big/img_594
+2002/08/23/big/img_19
+2002/08/29/big/img_18819
+2002/08/19/big/img_293
+2002/07/30/big/img_331
+2002/08/23/big/img_607
+2002/07/30/big/img_363
+2002/08/16/big/img_766
+2003/01/13/big/img_481
+2002/08/06/big/img_2515
+2002/09/02/big/img_15913
+2002/09/02/big/img_15827
+2002/09/02/big/img_15053
+2002/08/07/big/img_1576
+2002/07/23/big/img_268
+2002/08/21/big/img_152
+2003/01/15/big/img_578
+2002/07/21/big/img_589
+2002/07/20/big/img_548
+2002/08/27/big/img_19693
+2002/08/31/big/img_17252
+2002/07/31/big/img_138
+2002/07/23/big/img_372
+2002/08/16/big/img_695
+2002/07/27/big/img_287
+2002/08/15/big/img_315
+2002/08/10/big/img_361
+2002/07/29/big/img_899
+2002/08/13/big/img_771
+2002/08/21/big/img_92
+2003/01/15/big/img_425
+2003/01/16/big/img_450
+2002/09/01/big/img_16942
+2002/08/02/big/img_51
+2002/09/02/big/img_15379
+2002/08/24/big/img_147
+2002/08/30/big/img_18122
+2002/07/26/big/img_950
+2002/08/07/big/img_1400
+2002/08/17/big/img_468
+2002/08/15/big/img_470
+2002/07/30/big/img_318
+2002/07/22/big/img_644
+2002/08/27/big/img_19732
+2002/07/23/big/img_601
+2002/08/26/big/img_398
+2002/08/21/big/img_428
+2002/08/06/big/img_2119
+2002/08/29/big/img_19103
+2003/01/14/big/img_933
+2002/08/11/big/img_674
+2002/08/28/big/img_19420
+2002/08/03/big/img_418
+2002/08/17/big/img_312
+2002/07/25/big/img_1044
+2003/01/17/big/img_671
+2002/08/30/big/img_18297
+2002/07/25/big/img_755
+2002/07/23/big/img_471
+2002/08/21/big/img_39
+2002/07/26/big/img_699
+2003/01/14/big/img_33
+2002/07/31/big/img_411
+2002/08/16/big/img_645
+2003/01/17/big/img_116
+2002/09/02/big/img_15903
+2002/08/20/big/img_120
+2002/08/22/big/img_176
+2002/07/29/big/img_1316
+2002/08/27/big/img_19914
+2002/07/22/big/img_719
+2002/08/28/big/img_19239
+2003/01/13/big/img_385
+2002/08/08/big/img_525
+2002/07/19/big/img_782
+2002/08/13/big/img_843
+2002/07/30/big/img_107
+2002/08/11/big/img_752
+2002/07/29/big/img_383
+2002/08/26/big/img_249
+2002/08/29/big/img_18860
+2002/07/30/big/img_70
+2002/07/26/big/img_194
+2002/08/15/big/img_530
+2002/08/08/big/img_816
+2002/07/31/big/img_286
+2003/01/13/big/img_294
+2002/07/31/big/img_251
+2002/07/24/big/img_13
+2002/08/31/big/img_17938
+2002/07/22/big/img_642
+2003/01/14/big/img_728
+2002/08/18/big/img_47
+2002/08/22/big/img_306
+2002/08/20/big/img_348
+2002/08/15/big/img_764
+2002/08/08/big/img_163
+2002/07/23/big/img_531
+2002/07/23/big/img_467
+2003/01/16/big/img_743
+2003/01/13/big/img_535
+2002/08/02/big/img_523
+2002/08/22/big/img_120
+2002/08/11/big/img_496
+2002/08/29/big/img_19075
+2002/08/08/big/img_465
+2002/08/09/big/img_790
+2002/08/19/big/img_588
+2002/08/23/big/img_407
+2003/01/17/big/img_435
+2002/08/24/big/img_398
+2002/08/27/big/img_19899
+2003/01/15/big/img_335
+2002/08/13/big/img_493
+2002/09/02/big/img_15460
+2002/07/31/big/img_470
+2002/08/05/big/img_3550
+2002/07/28/big/img_123
+2002/08/01/big/img_1498
+2002/08/04/big/img_504
+2003/01/17/big/img_427
+2002/08/27/big/img_19708
+2002/07/27/big/img_861
+2002/07/25/big/img_685
+2002/07/31/big/img_207
+2003/01/14/big/img_745
+2002/08/31/big/img_17756
+2002/08/24/big/img_288
+2002/08/18/big/img_181
+2002/08/10/big/img_520
+2002/08/25/big/img_705
+2002/08/23/big/img_226
+2002/08/04/big/img_727
+2002/07/24/big/img_625
+2002/08/28/big/img_19157
+2002/08/23/big/img_586
+2002/07/31/big/img_232
+2003/01/13/big/img_240
+2003/01/14/big/img_321
+2003/01/15/big/img_533
+2002/07/23/big/img_480
+2002/07/24/big/img_371
+2002/08/21/big/img_702
+2002/08/31/big/img_17075
+2002/09/02/big/img_15278
+2002/07/29/big/img_246
+2003/01/15/big/img_829
+2003/01/15/big/img_1213
+2003/01/16/big/img_441
+2002/08/14/big/img_921
+2002/07/23/big/img_425
+2002/08/15/big/img_296
+2002/07/19/big/img_135
+2002/07/26/big/img_402
+2003/01/17/big/img_88
+2002/08/20/big/img_872
+2002/08/13/big/img_1110
+2003/01/16/big/img_1040
+2002/07/23/big/img_9
+2002/08/13/big/img_700
+2002/08/16/big/img_371
+2002/08/27/big/img_19966
+2003/01/17/big/img_391
+2002/08/18/big/img_426
+2002/08/01/big/img_1618
+2002/07/21/big/img_754
+2003/01/14/big/img_1101
+2003/01/16/big/img_1022
+2002/07/22/big/img_275
+2002/08/24/big/img_86
+2002/08/17/big/img_582
+2003/01/15/big/img_765
+2003/01/17/big/img_449
+2002/07/28/big/img_265
+2003/01/13/big/img_552
+2002/07/28/big/img_115
+2003/01/16/big/img_56
+2002/08/02/big/img_1232
+2003/01/17/big/img_925
+2002/07/22/big/img_445
+2002/07/25/big/img_957
+2002/07/20/big/img_589
+2002/08/31/big/img_17107
+2002/07/29/big/img_483
+2002/08/14/big/img_1063
+2002/08/07/big/img_1545
+2002/08/14/big/img_680
+2002/09/01/big/img_16694
+2002/08/14/big/img_257
+2002/08/11/big/img_726
+2002/07/26/big/img_681
+2002/07/25/big/img_481
+2003/01/14/big/img_737
+2002/08/28/big/img_19480
+2003/01/16/big/img_362
+2002/08/27/big/img_19865
+2003/01/01/big/img_547
+2002/09/02/big/img_15074
+2002/08/01/big/img_1453
+2002/08/22/big/img_594
+2002/08/28/big/img_19263
+2002/08/13/big/img_478
+2002/07/29/big/img_1358
+2003/01/14/big/img_1022
+2002/08/16/big/img_450
+2002/08/02/big/img_159
+2002/07/26/big/img_781
+2003/01/13/big/img_601
+2002/08/20/big/img_407
+2002/08/15/big/img_468
+2002/08/31/big/img_17902
+2002/08/16/big/img_81
+2002/07/25/big/img_987
+2002/07/25/big/img_500
+2002/08/02/big/img_31
+2002/08/18/big/img_538
+2002/08/08/big/img_54
+2002/07/23/big/img_686
+2002/07/24/big/img_836
+2003/01/17/big/img_734
+2002/08/16/big/img_1055
+2003/01/16/big/img_521
+2002/07/25/big/img_612
+2002/08/22/big/img_778
+2002/08/03/big/img_251
+2002/08/12/big/img_436
+2002/08/23/big/img_705
+2002/07/28/big/img_243
+2002/07/25/big/img_1029
+2002/08/20/big/img_287
+2002/08/29/big/img_18739
+2002/08/05/big/img_3272
+2002/07/27/big/img_214
+2003/01/14/big/img_5
+2002/08/01/big/img_1380
+2002/08/29/big/img_19097
+2002/07/30/big/img_486
+2002/08/29/big/img_18707
+2002/08/10/big/img_559
+2002/08/15/big/img_365
+2002/08/09/big/img_525
+2002/08/10/big/img_689
+2002/07/25/big/img_502
+2002/08/03/big/img_667
+2002/08/10/big/img_855
+2002/08/10/big/img_706
+2002/08/18/big/img_603
+2003/01/16/big/img_1055
+2002/08/31/big/img_17890
+2002/08/15/big/img_761
+2003/01/15/big/img_489
+2002/08/26/big/img_351
+2002/08/01/big/img_1772
+2002/08/31/big/img_17729
+2002/07/25/big/img_609
+2003/01/13/big/img_539
+2002/07/27/big/img_686
+2002/07/31/big/img_311
+2002/08/22/big/img_799
+2003/01/16/big/img_936
+2002/08/31/big/img_17813
+2002/08/04/big/img_862
+2002/08/09/big/img_332
+2002/07/20/big/img_148
+2002/08/12/big/img_426
+2002/07/24/big/img_69
+2002/07/27/big/img_685
+2002/08/02/big/img_480
+2002/08/26/big/img_154
+2002/07/24/big/img_598
+2002/08/01/big/img_1881
+2002/08/20/big/img_667
+2003/01/14/big/img_495
+2002/07/21/big/img_744
+2002/07/30/big/img_150
+2002/07/23/big/img_924
+2002/08/08/big/img_272
+2002/07/23/big/img_310
+2002/07/25/big/img_1011
+2002/09/02/big/img_15725
+2002/07/19/big/img_814
+2002/08/20/big/img_936
+2002/07/25/big/img_85
+2002/08/24/big/img_662
+2002/08/09/big/img_495
+2003/01/15/big/img_196
+2002/08/16/big/img_707
+2002/08/28/big/img_19370
+2002/08/06/big/img_2366
+2002/08/06/big/img_3012
+2002/08/01/big/img_1452
+2002/07/31/big/img_742
+2002/07/27/big/img_914
+2003/01/13/big/img_290
+2002/07/31/big/img_288
+2002/08/02/big/img_171
+2002/08/22/big/img_191
+2002/07/27/big/img_1066
+2002/08/12/big/img_383
+2003/01/17/big/img_1018
+2002/08/01/big/img_1785
+2002/08/11/big/img_390
+2002/08/27/big/img_20037
+2002/08/12/big/img_38
+2003/01/15/big/img_103
+2002/08/26/big/img_31
+2002/08/18/big/img_660
+2002/07/22/big/img_694
+2002/08/15/big/img_24
+2002/07/27/big/img_1077
+2002/08/01/big/img_1943
+2002/07/22/big/img_292
+2002/09/01/big/img_16857
+2002/07/22/big/img_892
+2003/01/14/big/img_46
+2002/08/09/big/img_469
+2002/08/09/big/img_414
+2003/01/16/big/img_40
+2002/08/28/big/img_19231
+2002/07/27/big/img_978
+2002/07/23/big/img_475
+2002/07/25/big/img_92
+2002/08/09/big/img_799
+2002/07/25/big/img_491
+2002/08/03/big/img_654
+2003/01/15/big/img_687
+2002/08/11/big/img_478
+2002/08/07/big/img_1664
+2002/08/20/big/img_362
+2002/08/01/big/img_1298
+2003/01/13/big/img_500
+2002/08/06/big/img_2896
+2002/08/30/big/img_18529
+2002/08/16/big/img_1020
+2002/07/29/big/img_892
+2002/08/29/big/img_18726
+2002/07/21/big/img_453
+2002/08/17/big/img_437
+2002/07/19/big/img_665
+2002/07/22/big/img_440
+2002/07/19/big/img_582
+2002/07/21/big/img_233
+2003/01/01/big/img_82
+2002/07/25/big/img_341
+2002/07/29/big/img_864
+2002/08/02/big/img_276
+2002/08/29/big/img_18654
+2002/07/27/big/img_1024
+2002/08/19/big/img_373
+2003/01/15/big/img_241
+2002/07/25/big/img_84
+2002/08/13/big/img_834
+2002/08/10/big/img_511
+2002/08/01/big/img_1627
+2002/08/08/big/img_607
+2002/08/06/big/img_2083
+2002/08/01/big/img_1486
+2002/08/08/big/img_700
+2002/08/01/big/img_1954
+2002/08/21/big/img_54
+2002/07/30/big/img_847
+2002/08/28/big/img_19169
+2002/07/21/big/img_549
+2002/08/03/big/img_693
+2002/07/31/big/img_1002
+2003/01/14/big/img_1035
+2003/01/16/big/img_622
+2002/07/30/big/img_1201
+2002/08/10/big/img_444
+2002/07/31/big/img_374
+2002/08/21/big/img_301
+2002/08/13/big/img_1095
+2003/01/13/big/img_288
+2002/07/25/big/img_232
+2003/01/13/big/img_967
+2002/08/26/big/img_360
+2002/08/05/big/img_67
+2002/08/29/big/img_18969
+2002/07/28/big/img_16
+2002/08/16/big/img_515
+2002/07/20/big/img_708
+2002/08/18/big/img_178
+2003/01/15/big/img_509
+2002/07/25/big/img_430
+2002/08/21/big/img_738
+2002/08/16/big/img_886
+2002/09/02/big/img_15605
+2002/09/01/big/img_16242
+2002/08/24/big/img_711
+2002/07/25/big/img_90
+2002/08/09/big/img_491
+2002/07/30/big/img_534
+2003/01/13/big/img_474
+2002/08/25/big/img_510
+2002/08/15/big/img_555
+2002/08/02/big/img_775
+2002/07/23/big/img_975
+2002/08/19/big/img_229
+2003/01/17/big/img_860
+2003/01/02/big/img_10
+2002/07/23/big/img_542
+2002/08/06/big/img_2535
+2002/07/22/big/img_37
+2002/08/06/big/img_2342
+2002/08/25/big/img_515
+2002/08/25/big/img_336
+2002/08/18/big/img_837
+2002/08/21/big/img_616
+2003/01/17/big/img_24
+2002/07/26/big/img_936
+2002/08/14/big/img_896
+2002/07/29/big/img_465
+2002/07/31/big/img_543
+2002/08/01/big/img_1411
+2002/08/02/big/img_423
+2002/08/21/big/img_44
+2002/07/31/big/img_11
+2003/01/15/big/img_628
+2003/01/15/big/img_605
+2002/07/30/big/img_571
+2002/07/23/big/img_428
+2002/08/15/big/img_942
+2002/07/26/big/img_531
+2003/01/16/big/img_59
+2002/08/02/big/img_410
+2002/07/31/big/img_230
+2002/08/19/big/img_806
+2003/01/14/big/img_462
+2002/08/16/big/img_370
+2002/08/13/big/img_380
+2002/08/16/big/img_932
+2002/07/19/big/img_393
+2002/08/20/big/img_764
+2002/08/15/big/img_616
+2002/07/26/big/img_267
+2002/07/27/big/img_1069
+2002/08/14/big/img_1041
+2003/01/13/big/img_594
+2002/09/01/big/img_16845
+2002/08/09/big/img_229
+2003/01/16/big/img_639
+2002/08/19/big/img_398
+2002/08/18/big/img_978
+2002/08/24/big/img_296
+2002/07/29/big/img_415
+2002/07/30/big/img_923
+2002/08/18/big/img_575
+2002/08/22/big/img_182
+2002/07/25/big/img_806
+2002/07/22/big/img_49
+2002/07/29/big/img_989
+2003/01/17/big/img_789
+2003/01/15/big/img_503
+2002/09/01/big/img_16062
+2003/01/17/big/img_794
+2002/08/15/big/img_564
+2003/01/15/big/img_222
+2002/08/01/big/img_1656
+2003/01/13/big/img_432
+2002/07/19/big/img_426
+2002/08/17/big/img_244
+2002/08/13/big/img_805
+2002/09/02/big/img_15067
+2002/08/11/big/img_58
+2002/08/22/big/img_636
+2002/07/22/big/img_416
+2002/08/13/big/img_836
+2002/08/26/big/img_363
+2002/07/30/big/img_917
+2003/01/14/big/img_206
+2002/08/12/big/img_311
+2002/08/31/big/img_17623
+2002/07/29/big/img_661
+2003/01/13/big/img_417
+2002/08/02/big/img_463
+2002/08/02/big/img_669
+2002/08/26/big/img_670
+2002/08/02/big/img_375
+2002/07/19/big/img_209
+2002/08/08/big/img_115
+2002/08/21/big/img_399
+2002/08/20/big/img_911
+2002/08/07/big/img_1212
+2002/08/20/big/img_578
+2002/08/22/big/img_554
+2002/08/21/big/img_484
+2002/07/25/big/img_450
+2002/08/03/big/img_542
+2002/08/15/big/img_561
+2002/07/23/big/img_360
+2002/08/30/big/img_18137
+2002/07/25/big/img_250
+2002/08/03/big/img_647
+2002/08/20/big/img_375
+2002/08/14/big/img_387
+2002/09/01/big/img_16990
+2002/08/28/big/img_19341
+2003/01/15/big/img_239
+2002/08/20/big/img_528
+2002/08/12/big/img_130
+2002/09/02/big/img_15108
+2003/01/15/big/img_372
+2002/08/16/big/img_678
+2002/08/04/big/img_623
+2002/07/23/big/img_477
+2002/08/28/big/img_19590
+2003/01/17/big/img_978
+2002/09/01/big/img_16692
+2002/07/20/big/img_109
+2002/08/06/big/img_2660
+2003/01/14/big/img_464
+2002/08/09/big/img_618
+2002/07/22/big/img_722
+2002/08/25/big/img_419
+2002/08/03/big/img_314
+2002/08/25/big/img_40
+2002/07/27/big/img_430
+2002/08/10/big/img_569
+2002/08/23/big/img_398
+2002/07/23/big/img_893
+2002/08/16/big/img_261
+2002/08/06/big/img_2668
+2002/07/22/big/img_835
+2002/09/02/big/img_15093
+2003/01/16/big/img_65
+2002/08/21/big/img_448
+2003/01/14/big/img_351
+2003/01/17/big/img_133
+2002/07/28/big/img_493
+2003/01/15/big/img_640
+2002/09/01/big/img_16880
+2002/08/15/big/img_350
+2002/08/20/big/img_624
+2002/08/25/big/img_604
+2002/08/06/big/img_2200
+2002/08/23/big/img_290
+2002/08/13/big/img_1152
+2003/01/14/big/img_251
+2002/08/02/big/img_538
+2002/08/22/big/img_613
+2003/01/13/big/img_351
+2002/08/18/big/img_368
+2002/07/23/big/img_392
+2002/07/25/big/img_198
+2002/07/25/big/img_418
+2002/08/26/big/img_614
+2002/07/23/big/img_405
+2003/01/14/big/img_445
+2002/07/25/big/img_326
+2002/08/10/big/img_734
+2003/01/14/big/img_530
+2002/08/08/big/img_561
+2002/08/29/big/img_18990
+2002/08/10/big/img_576
+2002/07/29/big/img_1494
+2002/07/19/big/img_198
+2002/08/10/big/img_562
+2002/07/22/big/img_901
+2003/01/14/big/img_37
+2002/09/02/big/img_15629
+2003/01/14/big/img_58
+2002/08/01/big/img_1364
+2002/07/27/big/img_636
+2003/01/13/big/img_241
+2002/09/01/big/img_16988
+2003/01/13/big/img_560
+2002/08/09/big/img_533
+2002/07/31/big/img_249
+2003/01/17/big/img_1007
+2002/07/21/big/img_64
+2003/01/13/big/img_537
+2003/01/15/big/img_606
+2002/08/18/big/img_651
+2002/08/24/big/img_405
+2002/07/26/big/img_837
+2002/08/09/big/img_562
+2002/08/01/big/img_1983
+2002/08/03/big/img_514
+2002/07/29/big/img_314
+2002/08/12/big/img_493
+2003/01/14/big/img_121
+2003/01/14/big/img_479
+2002/08/04/big/img_410
+2002/07/22/big/img_607
+2003/01/17/big/img_417
+2002/07/20/big/img_547
+2002/08/13/big/img_396
+2002/08/31/big/img_17538
+2002/08/13/big/img_187
+2002/08/12/big/img_328
+2003/01/14/big/img_569
+2002/07/27/big/img_1081
+2002/08/14/big/img_504
+2002/08/23/big/img_785
+2002/07/26/big/img_339
+2002/08/07/big/img_1156
+2002/08/07/big/img_1456
+2002/08/23/big/img_378
+2002/08/27/big/img_19719
+2002/07/31/big/img_39
+2002/07/31/big/img_883
+2003/01/14/big/img_676
+2002/07/29/big/img_214
+2002/07/26/big/img_669
+2002/07/25/big/img_202
+2002/08/08/big/img_259
+2003/01/17/big/img_943
+2003/01/15/big/img_512
+2002/08/05/big/img_3295
+2002/08/27/big/img_19685
+2002/08/08/big/img_277
+2002/08/30/big/img_18154
+2002/07/22/big/img_663
+2002/08/29/big/img_18914
+2002/07/31/big/img_908
+2002/08/27/big/img_19926
+2003/01/13/big/img_791
+2003/01/15/big/img_827
+2002/08/18/big/img_878
+2002/08/14/big/img_670
+2002/07/20/big/img_182
+2002/08/15/big/img_291
+2002/08/06/big/img_2600
+2002/07/23/big/img_587
+2002/08/14/big/img_577
+2003/01/15/big/img_585
+2002/07/30/big/img_310
+2002/08/03/big/img_658
+2002/08/10/big/img_157
+2002/08/19/big/img_811
+2002/07/29/big/img_1318
+2002/08/04/big/img_104
+2002/07/30/big/img_332
+2002/07/24/big/img_789
+2002/07/29/big/img_516
+2002/07/23/big/img_843
+2002/08/01/big/img_1528
+2002/08/13/big/img_798
+2002/08/07/big/img_1729
+2002/08/28/big/img_19448
+2003/01/16/big/img_95
+2002/08/12/big/img_473
+2002/07/27/big/img_269
+2003/01/16/big/img_621
+2002/07/29/big/img_772
+2002/07/24/big/img_171
+2002/07/19/big/img_429
+2002/08/07/big/img_1933
+2002/08/27/big/img_19629
+2002/08/05/big/img_3688
+2002/08/07/big/img_1691
+2002/07/23/big/img_600
+2002/07/29/big/img_666
+2002/08/25/big/img_566
+2002/08/06/big/img_2659
+2002/08/29/big/img_18929
+2002/08/16/big/img_407
+2002/08/18/big/img_774
+2002/08/19/big/img_249
+2002/08/06/big/img_2427
+2002/08/29/big/img_18899
+2002/08/01/big/img_1818
+2002/07/31/big/img_108
+2002/07/29/big/img_500
+2002/08/11/big/img_115
+2002/07/19/big/img_521
+2002/08/02/big/img_1163
+2002/07/22/big/img_62
+2002/08/13/big/img_466
+2002/08/21/big/img_956
+2002/08/23/big/img_602
+2002/08/20/big/img_858
+2002/07/25/big/img_690
+2002/07/19/big/img_130
+2002/08/04/big/img_874
+2002/07/26/big/img_489
+2002/07/22/big/img_548
+2002/08/10/big/img_191
+2002/07/25/big/img_1051
+2002/08/18/big/img_473
+2002/08/12/big/img_755
+2002/08/18/big/img_413
+2002/08/08/big/img_1044
+2002/08/17/big/img_680
+2002/08/26/big/img_235
+2002/08/20/big/img_330
+2002/08/22/big/img_344
+2002/08/09/big/img_593
+2002/07/31/big/img_1006
+2002/08/14/big/img_337
+2002/08/16/big/img_728
+2002/07/24/big/img_834
+2002/08/04/big/img_552
+2002/09/02/big/img_15213
+2002/07/25/big/img_725
+2002/08/30/big/img_18290
+2003/01/01/big/img_475
+2002/07/27/big/img_1083
+2002/08/29/big/img_18955
+2002/08/31/big/img_17232
+2002/08/08/big/img_480
+2002/08/01/big/img_1311
+2002/07/30/big/img_745
+2002/08/03/big/img_649
+2002/08/12/big/img_193
+2002/07/29/big/img_228
+2002/07/25/big/img_836
+2002/08/20/big/img_400
+2002/07/30/big/img_507
+2002/09/02/big/img_15072
+2002/07/26/big/img_658
+2002/07/28/big/img_503
+2002/08/05/big/img_3814
+2002/08/24/big/img_745
+2003/01/13/big/img_817
+2002/08/08/big/img_579
+2002/07/22/big/img_251
+2003/01/13/big/img_689
+2002/07/25/big/img_407
+2002/08/13/big/img_1050
+2002/08/14/big/img_733
+2002/07/24/big/img_82
+2003/01/17/big/img_288
+2003/01/15/big/img_475
+2002/08/14/big/img_620
+2002/08/21/big/img_167
+2002/07/19/big/img_300
+2002/07/26/big/img_219
+2002/08/01/big/img_1468
+2002/07/23/big/img_260
+2002/08/09/big/img_555
+2002/07/19/big/img_160
+2002/08/02/big/img_1060
+2003/01/14/big/img_149
+2002/08/15/big/img_346
+2002/08/24/big/img_597
+2002/08/22/big/img_502
+2002/08/30/big/img_18228
+2002/07/21/big/img_766
+2003/01/15/big/img_841
+2002/07/24/big/img_516
+2002/08/02/big/img_265
+2002/08/15/big/img_1243
+2003/01/15/big/img_223
+2002/08/04/big/img_236
+2002/07/22/big/img_309
+2002/07/20/big/img_656
+2002/07/31/big/img_412
+2002/09/01/big/img_16462
+2003/01/16/big/img_431
+2002/07/22/big/img_793
+2002/08/15/big/img_877
+2002/07/26/big/img_282
+2002/07/25/big/img_529
+2002/08/24/big/img_613
+2003/01/17/big/img_700
+2002/08/06/big/img_2526
+2002/08/24/big/img_394
+2002/08/21/big/img_521
+2002/08/25/big/img_560
+2002/07/29/big/img_966
+2002/07/25/big/img_448
+2003/01/13/big/img_782
+2002/08/21/big/img_296
+2002/09/01/big/img_16755
+2002/08/05/big/img_3552
+2002/09/02/big/img_15823
+2003/01/14/big/img_193
+2002/07/21/big/img_159
+2002/08/02/big/img_564
+2002/08/16/big/img_300
+2002/07/19/big/img_269
+2002/08/13/big/img_676
+2002/07/28/big/img_57
+2002/08/05/big/img_3318
+2002/07/31/big/img_218
+2002/08/21/big/img_898
+2002/07/29/big/img_109
+2002/07/19/big/img_854
+2002/08/23/big/img_311
+2002/08/14/big/img_318
+2002/07/25/big/img_523
+2002/07/21/big/img_678
+2003/01/17/big/img_690
+2002/08/28/big/img_19503
+2002/08/18/big/img_251
+2002/08/22/big/img_672
+2002/08/20/big/img_663
+2002/08/02/big/img_148
+2002/09/02/big/img_15580
+2002/07/25/big/img_778
+2002/08/14/big/img_565
+2002/08/12/big/img_374
+2002/08/13/big/img_1018
+2002/08/20/big/img_474
+2002/08/25/big/img_33
+2002/08/02/big/img_1190
+2002/08/08/big/img_864
+2002/08/14/big/img_1071
+2002/08/30/big/img_18103
+2002/08/18/big/img_533
+2003/01/16/big/img_650
+2002/07/25/big/img_108
+2002/07/26/big/img_81
+2002/07/27/big/img_543
+2002/07/29/big/img_521
+2003/01/13/big/img_434
+2002/08/26/big/img_674
+2002/08/06/big/img_2932
+2002/08/07/big/img_1262
+2003/01/15/big/img_201
+2003/01/16/big/img_673
+2002/09/02/big/img_15988
+2002/07/29/big/img_1306
+2003/01/14/big/img_1072
+2002/08/30/big/img_18232
+2002/08/05/big/img_3711
+2002/07/23/big/img_775
+2002/08/01/big/img_16
+2003/01/16/big/img_630
+2002/08/22/big/img_695
+2002/08/14/big/img_51
+2002/08/14/big/img_782
+2002/08/24/big/img_742
+2003/01/14/big/img_512
+2003/01/15/big/img_1183
+2003/01/15/big/img_714
+2002/08/01/big/img_2078
+2002/07/31/big/img_682
+2002/09/02/big/img_15687
+2002/07/26/big/img_518
+2002/08/27/big/img_19676
+2002/09/02/big/img_15969
+2002/08/02/big/img_931
+2002/08/25/big/img_508
+2002/08/29/big/img_18616
+2002/07/22/big/img_839
+2002/07/28/big/img_313
+2003/01/14/big/img_155
+2002/08/02/big/img_1105
+2002/08/09/big/img_53
+2002/08/16/big/img_469
+2002/08/15/big/img_502
+2002/08/20/big/img_575
+2002/07/25/big/img_138
+2003/01/16/big/img_579
+2002/07/19/big/img_352
+2003/01/14/big/img_762
+2003/01/01/big/img_588
+2002/08/02/big/img_981
+2002/08/21/big/img_447
+2002/09/01/big/img_16151
+2003/01/14/big/img_769
+2002/08/23/big/img_461
+2002/08/17/big/img_240
+2002/09/02/big/img_15220
+2002/07/19/big/img_408
+2002/09/02/big/img_15496
+2002/07/29/big/img_758
+2002/08/28/big/img_19392
+2002/08/06/big/img_2723
+2002/08/31/big/img_17752
+2002/08/23/big/img_469
+2002/08/13/big/img_515
+2002/09/02/big/img_15551
+2002/08/03/big/img_462
+2002/07/24/big/img_613
+2002/07/22/big/img_61
+2002/08/08/big/img_171
+2002/08/21/big/img_177
+2003/01/14/big/img_105
+2002/08/02/big/img_1017
+2002/08/22/big/img_106
+2002/07/27/big/img_542
+2002/07/21/big/img_665
+2002/07/23/big/img_595
+2002/08/04/big/img_657
+2002/08/29/big/img_19002
+2003/01/15/big/img_550
+2002/08/14/big/img_662
+2002/07/20/big/img_425
+2002/08/30/big/img_18528
+2002/07/26/big/img_611
+2002/07/22/big/img_849
+2002/08/07/big/img_1655
+2002/08/21/big/img_638
+2003/01/17/big/img_732
+2003/01/01/big/img_496
+2002/08/18/big/img_713
+2002/08/08/big/img_109
+2002/07/27/big/img_1008
+2002/07/20/big/img_559
+2002/08/16/big/img_699
+2002/08/31/big/img_17702
+2002/07/31/big/img_1013
+2002/08/01/big/img_2027
+2002/08/02/big/img_1001
+2002/08/03/big/img_210
+2002/08/01/big/img_2087
+2003/01/14/big/img_199
+2002/07/29/big/img_48
+2002/07/19/big/img_727
+2002/08/09/big/img_249
+2002/08/04/big/img_632
+2002/08/22/big/img_620
+2003/01/01/big/img_457
+2002/08/05/big/img_3223
+2002/07/27/big/img_240
+2002/07/25/big/img_797
+2002/08/13/big/img_430
+2002/07/25/big/img_615
+2002/08/12/big/img_28
+2002/07/30/big/img_220
+2002/07/24/big/img_89
+2002/08/21/big/img_357
+2002/08/09/big/img_590
+2003/01/13/big/img_525
+2002/08/17/big/img_818
+2003/01/02/big/img_7
+2002/07/26/big/img_636
+2003/01/13/big/img_1122
+2002/07/23/big/img_810
+2002/08/20/big/img_888
+2002/07/27/big/img_3
+2002/08/15/big/img_451
+2002/09/02/big/img_15787
+2002/07/31/big/img_281
+2002/08/05/big/img_3274
+2002/08/07/big/img_1254
+2002/07/31/big/img_27
+2002/08/01/big/img_1366
+2002/07/30/big/img_182
+2002/08/27/big/img_19690
+2002/07/29/big/img_68
+2002/08/23/big/img_754
+2002/07/30/big/img_540
+2002/08/27/big/img_20063
+2002/08/14/big/img_471
+2002/08/02/big/img_615
+2002/07/30/big/img_186
+2002/08/25/big/img_150
+2002/07/27/big/img_626
+2002/07/20/big/img_225
+2003/01/15/big/img_1252
+2002/07/19/big/img_367
+2003/01/15/big/img_582
+2002/08/09/big/img_572
+2002/08/08/big/img_428
+2003/01/15/big/img_639
+2002/08/28/big/img_19245
+2002/07/24/big/img_321
+2002/08/02/big/img_662
+2002/08/08/big/img_1033
+2003/01/17/big/img_867
+2002/07/22/big/img_652
+2003/01/14/big/img_224
+2002/08/18/big/img_49
+2002/07/26/big/img_46
+2002/08/31/big/img_18021
+2002/07/25/big/img_151
+2002/08/23/big/img_540
+2002/08/25/big/img_693
+2002/07/23/big/img_340
+2002/07/28/big/img_117
+2002/09/02/big/img_15768
+2002/08/26/big/img_562
+2002/07/24/big/img_480
+2003/01/15/big/img_341
+2002/08/10/big/img_783
+2002/08/20/big/img_132
+2003/01/14/big/img_370
+2002/07/20/big/img_720
+2002/08/03/big/img_144
+2002/08/20/big/img_538
+2002/08/01/big/img_1745
+2002/08/11/big/img_683
+2002/08/03/big/img_328
+2002/08/10/big/img_793
+2002/08/14/big/img_689
+2002/08/02/big/img_162
+2003/01/17/big/img_411
+2002/07/31/big/img_361
+2002/08/15/big/img_289
+2002/08/08/big/img_254
+2002/08/15/big/img_996
+2002/08/20/big/img_785
+2002/07/24/big/img_511
+2002/08/06/big/img_2614
+2002/08/29/big/img_18733
+2002/08/17/big/img_78
+2002/07/30/big/img_378
+2002/08/31/big/img_17947
+2002/08/26/big/img_88
+2002/07/30/big/img_558
+2002/08/02/big/img_67
+2003/01/14/big/img_325
+2002/07/29/big/img_1357
+2002/07/19/big/img_391
+2002/07/30/big/img_307
+2003/01/13/big/img_219
+2002/07/24/big/img_807
+2002/08/23/big/img_543
+2002/08/29/big/img_18620
+2002/07/22/big/img_769
+2002/08/26/big/img_503
+2002/07/30/big/img_78
+2002/08/14/big/img_1036
+2002/08/09/big/img_58
+2002/07/24/big/img_616
+2002/08/02/big/img_464
+2002/07/26/big/img_576
+2002/07/22/big/img_273
+2003/01/16/big/img_470
+2002/07/29/big/img_329
+2002/07/30/big/img_1086
+2002/07/31/big/img_353
+2002/09/02/big/img_15275
+2003/01/17/big/img_555
+2002/08/26/big/img_212
+2002/08/01/big/img_1692
+2003/01/15/big/img_600
+2002/07/29/big/img_825
+2002/08/08/big/img_68
+2002/08/10/big/img_719
+2002/07/31/big/img_636
+2002/07/29/big/img_325
+2002/07/21/big/img_515
+2002/07/22/big/img_705
+2003/01/13/big/img_818
+2002/08/09/big/img_486
+2002/08/22/big/img_141
+2002/07/22/big/img_303
+2002/08/09/big/img_393
+2002/07/29/big/img_963
+2002/08/02/big/img_1215
+2002/08/19/big/img_674
+2002/08/12/big/img_690
+2002/08/21/big/img_637
+2002/08/21/big/img_841
+2002/08/24/big/img_71
+2002/07/25/big/img_596
+2002/07/24/big/img_864
+2002/08/18/big/img_293
+2003/01/14/big/img_657
+2002/08/15/big/img_411
+2002/08/16/big/img_348
+2002/08/05/big/img_3157
+2002/07/20/big/img_663
+2003/01/13/big/img_654
+2003/01/16/big/img_433
+2002/08/30/big/img_18200
+2002/08/12/big/img_226
+2003/01/16/big/img_491
+2002/08/08/big/img_666
+2002/07/19/big/img_576
+2003/01/15/big/img_776
+2003/01/16/big/img_899
+2002/07/19/big/img_397
+2002/08/14/big/img_44
+2003/01/15/big/img_762
+2002/08/02/big/img_982
+2002/09/02/big/img_15234
+2002/08/17/big/img_556
+2002/08/21/big/img_410
+2002/08/21/big/img_386
+2002/07/19/big/img_690
+2002/08/05/big/img_3052
+2002/08/14/big/img_219
+2002/08/16/big/img_273
+2003/01/15/big/img_752
+2002/08/08/big/img_184
+2002/07/31/big/img_743
+2002/08/23/big/img_338
+2003/01/14/big/img_1055
+2002/08/05/big/img_3405
+2003/01/15/big/img_17
+2002/08/03/big/img_141
+2002/08/14/big/img_549
+2002/07/27/big/img_1034
+2002/07/31/big/img_932
+2002/08/30/big/img_18487
+2002/09/02/big/img_15814
+2002/08/01/big/img_2086
+2002/09/01/big/img_16535
+2002/07/22/big/img_500
+2003/01/13/big/img_400
+2002/08/25/big/img_607
+2002/08/30/big/img_18384
+2003/01/14/big/img_951
+2002/08/13/big/img_1150
+2002/08/08/big/img_1022
+2002/08/10/big/img_428
+2002/08/28/big/img_19242
+2002/08/05/big/img_3098
+2002/07/23/big/img_400
+2002/08/26/big/img_365
+2002/07/20/big/img_318
+2002/08/13/big/img_740
+2003/01/16/big/img_37
+2002/08/26/big/img_274
+2002/08/02/big/img_205
+2002/08/21/big/img_695
+2002/08/06/big/img_2289
+2002/08/20/big/img_794
+2002/08/18/big/img_438
+2002/08/07/big/img_1380
+2002/08/02/big/img_737
+2002/08/07/big/img_1651
+2002/08/15/big/img_1238
+2002/08/01/big/img_1681
+2002/08/06/big/img_3017
+2002/07/23/big/img_706
+2002/07/31/big/img_392
+2002/08/09/big/img_539
+2002/07/29/big/img_835
+2002/08/26/big/img_723
+2002/08/28/big/img_19235
+2003/01/16/big/img_353
+2002/08/10/big/img_150
+2002/08/29/big/img_19025
+2002/08/21/big/img_310
+2002/08/10/big/img_823
+2002/07/26/big/img_981
+2002/08/11/big/img_288
+2002/08/19/big/img_534
+2002/08/21/big/img_300
+2002/07/31/big/img_49
+2002/07/30/big/img_469
+2002/08/28/big/img_19197
+2002/08/25/big/img_205
+2002/08/10/big/img_390
+2002/08/23/big/img_291
+2002/08/26/big/img_230
+2002/08/18/big/img_76
+2002/07/23/big/img_409
+2002/08/14/big/img_1053
+2003/01/14/big/img_291
+2002/08/10/big/img_503
+2002/08/27/big/img_19928
+2002/08/03/big/img_563
+2002/08/17/big/img_250
+2002/08/06/big/img_2381
+2002/08/17/big/img_948
+2002/08/06/big/img_2710
+2002/07/22/big/img_696
+2002/07/31/big/img_670
+2002/08/12/big/img_594
+2002/07/29/big/img_624
+2003/01/17/big/img_934
+2002/08/03/big/img_584
+2002/08/22/big/img_1003
+2002/08/05/big/img_3396
+2003/01/13/big/img_570
+2002/08/02/big/img_219
+2002/09/02/big/img_15774
+2002/08/16/big/img_818
+2002/08/23/big/img_402
+2003/01/14/big/img_552
+2002/07/29/big/img_71
+2002/08/05/big/img_3592
+2002/08/16/big/img_80
+2002/07/27/big/img_672
+2003/01/13/big/img_470
+2003/01/16/big/img_702
+2002/09/01/big/img_16130
+2002/08/08/big/img_240
+2002/09/01/big/img_16338
+2002/07/26/big/img_312
+2003/01/14/big/img_538
+2002/07/20/big/img_695
+2002/08/30/big/img_18098
+2002/08/25/big/img_259
+2002/08/16/big/img_1042
+2002/08/09/big/img_837
+2002/08/31/big/img_17760
+2002/07/31/big/img_14
+2002/08/09/big/img_361
+2003/01/16/big/img_107
+2002/08/14/big/img_124
+2002/07/19/big/img_463
+2003/01/15/big/img_275
+2002/07/25/big/img_1151
+2002/07/29/big/img_1501
+2002/08/27/big/img_19889
+2002/08/29/big/img_18603
+2003/01/17/big/img_601
+2002/08/25/big/img_355
+2002/08/08/big/img_297
+2002/08/20/big/img_290
+2002/07/31/big/img_195
+2003/01/01/big/img_336
+2002/08/18/big/img_369
+2002/07/25/big/img_621
+2002/08/11/big/img_508
+2003/01/14/big/img_458
+2003/01/15/big/img_795
+2002/08/12/big/img_498
+2002/08/01/big/img_1734
+2002/08/02/big/img_246
+2002/08/16/big/img_565
+2002/08/11/big/img_475
+2002/08/22/big/img_408
+2002/07/28/big/img_78
+2002/07/21/big/img_81
+2003/01/14/big/img_697
+2002/08/14/big/img_661
+2002/08/15/big/img_507
+2002/08/19/big/img_55
+2002/07/22/big/img_152
+2003/01/14/big/img_470
+2002/08/03/big/img_379
+2002/08/22/big/img_506
+2003/01/16/big/img_966
+2002/08/18/big/img_698
+2002/08/24/big/img_528
+2002/08/23/big/img_10
+2002/08/01/big/img_1655
+2002/08/22/big/img_953
+2002/07/19/big/img_630
+2002/07/22/big/img_889
+2002/08/16/big/img_351
+2003/01/16/big/img_83
+2002/07/19/big/img_805
+2002/08/14/big/img_704
+2002/07/19/big/img_389
+2002/08/31/big/img_17765
+2002/07/29/big/img_606
+2003/01/17/big/img_939
+2002/09/02/big/img_15081
+2002/08/21/big/img_181
+2002/07/29/big/img_1321
+2002/07/21/big/img_497
+2002/07/20/big/img_539
+2002/08/24/big/img_119
+2002/08/01/big/img_1281
+2002/07/26/big/img_207
+2002/07/26/big/img_432
+2002/07/27/big/img_1006
+2002/08/05/big/img_3087
+2002/08/14/big/img_252
+2002/08/14/big/img_798
+2002/07/24/big/img_538
+2002/09/02/big/img_15507
+2002/08/08/big/img_901
+2003/01/14/big/img_557
+2002/08/07/big/img_1819
+2002/08/04/big/img_470
+2002/08/01/big/img_1504
+2002/08/16/big/img_1070
+2002/08/16/big/img_372
+2002/08/23/big/img_416
+2002/08/30/big/img_18208
+2002/08/01/big/img_2043
+2002/07/22/big/img_385
+2002/08/22/big/img_466
+2002/08/21/big/img_869
+2002/08/28/big/img_19429
+2002/08/02/big/img_770
+2002/07/23/big/img_433
+2003/01/14/big/img_13
+2002/07/27/big/img_953
+2002/09/02/big/img_15728
+2002/08/01/big/img_1361
+2002/08/29/big/img_18897
+2002/08/26/big/img_534
+2002/08/11/big/img_121
+2002/08/26/big/img_20130
+2002/07/31/big/img_363
+2002/08/13/big/img_978
+2002/07/25/big/img_835
+2002/08/02/big/img_906
+2003/01/14/big/img_548
+2002/07/30/big/img_80
+2002/07/26/big/img_982
+2003/01/16/big/img_99
+2002/08/19/big/img_362
+2002/08/24/big/img_376
+2002/08/07/big/img_1264
+2002/07/27/big/img_938
+2003/01/17/big/img_535
+2002/07/26/big/img_457
+2002/08/08/big/img_848
+2003/01/15/big/img_859
+2003/01/15/big/img_622
+2002/07/30/big/img_403
+2002/07/29/big/img_217
+2002/07/26/big/img_891
+2002/07/24/big/img_70
+2002/08/25/big/img_619
+2002/08/05/big/img_3375
+2002/08/01/big/img_2160
+2002/08/06/big/img_2227
+2003/01/14/big/img_117
+2002/08/14/big/img_227
+2002/08/13/big/img_565
+2002/08/19/big/img_625
+2002/08/03/big/img_812
+2002/07/24/big/img_41
+2002/08/16/big/img_235
+2002/07/29/big/img_759
+2002/07/21/big/img_433
+2002/07/29/big/img_190
+2003/01/16/big/img_435
+2003/01/13/big/img_708
+2002/07/30/big/img_57
+2002/08/22/big/img_162
+2003/01/01/big/img_558
+2003/01/15/big/img_604
+2002/08/16/big/img_935
+2002/08/20/big/img_394
+2002/07/28/big/img_465
+2002/09/02/big/img_15534
+2002/08/16/big/img_87
+2002/07/22/big/img_469
+2002/08/12/big/img_245
+2003/01/13/big/img_236
+2002/08/06/big/img_2736
+2002/08/03/big/img_348
+2003/01/14/big/img_218
+2002/07/26/big/img_232
+2003/01/15/big/img_244
+2002/07/25/big/img_1121
+2002/08/01/big/img_1484
+2002/07/26/big/img_541
+2002/08/07/big/img_1244
+2002/07/31/big/img_3
+2002/08/30/big/img_18437
+2002/08/29/big/img_19094
+2002/08/01/big/img_1355
+2002/08/19/big/img_338
+2002/07/19/big/img_255
+2002/07/21/big/img_76
+2002/08/25/big/img_199
+2002/08/12/big/img_740
+2002/07/30/big/img_852
+2002/08/15/big/img_599
+2002/08/23/big/img_254
+2002/08/19/big/img_125
+2002/07/24/big/img_2
+2002/08/04/big/img_145
+2002/08/05/big/img_3137
+2002/07/28/big/img_463
+2003/01/14/big/img_801
+2002/07/23/big/img_366
+2002/08/26/big/img_600
+2002/08/26/big/img_649
+2002/09/02/big/img_15849
+2002/07/26/big/img_248
+2003/01/13/big/img_200
+2002/08/07/big/img_1794
+2002/08/31/big/img_17270
+2002/08/23/big/img_608
+2003/01/13/big/img_837
+2002/08/23/big/img_581
+2002/08/20/big/img_754
+2002/08/18/big/img_183
+2002/08/20/big/img_328
+2002/07/22/big/img_494
+2002/07/29/big/img_399
+2002/08/28/big/img_19284
+2002/08/08/big/img_566
+2002/07/25/big/img_376
+2002/07/23/big/img_138
+2002/07/25/big/img_435
+2002/08/17/big/img_685
+2002/07/19/big/img_90
+2002/07/20/big/img_716
+2002/08/31/big/img_17458
+2002/08/26/big/img_461
+2002/07/25/big/img_355
+2002/08/06/big/img_2152
+2002/07/27/big/img_932
+2002/07/23/big/img_232
+2002/08/08/big/img_1020
+2002/07/31/big/img_366
+2002/08/06/big/img_2667
+2002/08/21/big/img_465
+2002/08/15/big/img_305
+2002/08/02/big/img_247
+2002/07/28/big/img_46
+2002/08/27/big/img_19922
+2002/08/23/big/img_643
+2003/01/13/big/img_624
+2002/08/23/big/img_625
+2002/08/05/big/img_3787
+2003/01/13/big/img_627
+2002/09/01/big/img_16381
+2002/08/05/big/img_3668
+2002/07/21/big/img_535
+2002/08/27/big/img_19680
+2002/07/22/big/img_413
+2002/07/29/big/img_481
+2003/01/15/big/img_496
+2002/07/23/big/img_701
+2002/08/29/big/img_18670
+2002/07/28/big/img_319
+2003/01/14/big/img_517
+2002/07/26/big/img_256
+2003/01/16/big/img_593
+2002/07/30/big/img_956
+2002/07/30/big/img_667
+2002/07/25/big/img_100
+2002/08/11/big/img_570
+2002/07/26/big/img_745
+2002/08/04/big/img_834
+2002/08/25/big/img_521
+2002/08/01/big/img_2148
+2002/09/02/big/img_15183
+2002/08/22/big/img_514
+2002/08/23/big/img_477
+2002/07/23/big/img_336
+2002/07/26/big/img_481
+2002/08/20/big/img_409
+2002/07/23/big/img_918
+2002/08/09/big/img_474
+2002/08/02/big/img_929
+2002/08/31/big/img_17932
+2002/08/19/big/img_161
+2002/08/09/big/img_667
+2002/07/31/big/img_805
+2002/09/02/big/img_15678
+2002/08/31/big/img_17509
+2002/08/29/big/img_18998
+2002/07/23/big/img_301
+2002/08/07/big/img_1612
+2002/08/06/big/img_2472
+2002/07/23/big/img_466
+2002/08/27/big/img_19634
+2003/01/16/big/img_16
+2002/08/14/big/img_193
+2002/08/21/big/img_340
+2002/08/27/big/img_19799
+2002/08/01/big/img_1345
+2002/08/07/big/img_1448
+2002/08/11/big/img_324
+2003/01/16/big/img_754
+2002/08/13/big/img_418
+2003/01/16/big/img_544
+2002/08/19/big/img_135
+2002/08/10/big/img_455
+2002/08/10/big/img_693
+2002/08/31/big/img_17967
+2002/08/28/big/img_19229
+2002/08/04/big/img_811
+2002/09/01/big/img_16225
+2003/01/16/big/img_428
+2002/09/02/big/img_15295
+2002/07/26/big/img_108
+2002/07/21/big/img_477
+2002/08/07/big/img_1354
+2002/08/23/big/img_246
+2002/08/16/big/img_652
+2002/07/27/big/img_553
+2002/07/31/big/img_346
+2002/08/04/big/img_537
+2002/08/08/big/img_498
+2002/08/29/big/img_18956
+2003/01/13/big/img_922
+2002/08/31/big/img_17425
+2002/07/26/big/img_438
+2002/08/19/big/img_185
+2003/01/16/big/img_33
+2002/08/10/big/img_252
+2002/07/29/big/img_598
+2002/08/27/big/img_19820
+2002/08/06/big/img_2664
+2002/08/20/big/img_705
+2003/01/14/big/img_816
+2002/08/03/big/img_552
+2002/07/25/big/img_561
+2002/07/25/big/img_934
+2002/08/01/big/img_1893
+2003/01/14/big/img_746
+2003/01/16/big/img_519
+2002/08/03/big/img_681
+2002/07/24/big/img_808
+2002/08/14/big/img_803
+2002/08/25/big/img_155
+2002/07/30/big/img_1107
+2002/08/29/big/img_18882
+2003/01/15/big/img_598
+2002/08/19/big/img_122
+2002/07/30/big/img_428
+2002/07/24/big/img_684
+2002/08/22/big/img_192
+2002/08/22/big/img_543
+2002/08/07/big/img_1318
+2002/08/18/big/img_25
+2002/07/26/big/img_583
+2002/07/20/big/img_464
+2002/08/19/big/img_664
+2002/08/24/big/img_861
+2002/09/01/big/img_16136
+2002/08/22/big/img_400
+2002/08/12/big/img_445
+2003/01/14/big/img_174
+2002/08/27/big/img_19677
+2002/08/31/big/img_17214
+2002/08/30/big/img_18175
+2003/01/17/big/img_402
+2002/08/06/big/img_2396
+2002/08/18/big/img_448
+2002/08/21/big/img_165
+2002/08/31/big/img_17609
+2003/01/01/big/img_151
+2002/08/26/big/img_372
+2002/09/02/big/img_15994
+2002/07/26/big/img_660
+2002/09/02/big/img_15197
+2002/07/29/big/img_258
+2002/08/30/big/img_18525
+2003/01/13/big/img_368
+2002/07/29/big/img_1538
+2002/07/21/big/img_787
+2002/08/18/big/img_152
+2002/08/06/big/img_2379
+2003/01/17/big/img_864
+2002/08/27/big/img_19998
+2002/08/01/big/img_1634
+2002/07/25/big/img_414
+2002/08/22/big/img_627
+2002/08/07/big/img_1669
+2002/08/16/big/img_1052
+2002/08/31/big/img_17796
+2002/08/18/big/img_199
+2002/09/02/big/img_15147
+2002/08/09/big/img_460
+2002/08/14/big/img_581
+2002/08/30/big/img_18286
+2002/07/26/big/img_337
+2002/08/18/big/img_589
+2003/01/14/big/img_866
+2002/07/20/big/img_624
+2002/08/01/big/img_1801
+2002/07/24/big/img_683
+2002/08/09/big/img_725
+2003/01/14/big/img_34
+2002/07/30/big/img_144
+2002/07/30/big/img_706
+2002/08/08/big/img_394
+2002/08/19/big/img_619
+2002/08/06/big/img_2703
+2002/08/29/big/img_19034
+2002/07/24/big/img_67
+2002/08/27/big/img_19841
+2002/08/19/big/img_427
+2003/01/14/big/img_333
+2002/09/01/big/img_16406
+2002/07/19/big/img_882
+2002/08/17/big/img_238
+2003/01/14/big/img_739
+2002/07/22/big/img_151
+2002/08/21/big/img_743
+2002/07/25/big/img_1048
+2002/07/30/big/img_395
+2003/01/13/big/img_584
+2002/08/13/big/img_742
+2002/08/13/big/img_1168
+2003/01/14/big/img_147
+2002/07/26/big/img_803
+2002/08/05/big/img_3298
+2002/08/07/big/img_1451
+2002/08/16/big/img_424
+2002/07/29/big/img_1069
+2002/09/01/big/img_16735
+2002/07/21/big/img_637
+2003/01/14/big/img_585
+2002/08/02/big/img_358
+2003/01/13/big/img_358
+2002/08/14/big/img_198
+2002/08/17/big/img_935
+2002/08/04/big/img_42
+2002/08/30/big/img_18245
+2002/07/25/big/img_158
+2002/08/22/big/img_744
+2002/08/06/big/img_2291
+2002/08/05/big/img_3044
+2002/07/30/big/img_272
+2002/08/23/big/img_641
+2002/07/24/big/img_797
+2002/07/30/big/img_392
+2003/01/14/big/img_447
+2002/07/31/big/img_898
+2002/08/06/big/img_2812
+2002/08/13/big/img_564
+2002/07/22/big/img_43
+2002/07/26/big/img_634
+2002/07/19/big/img_843
+2002/08/26/big/img_58
+2002/07/21/big/img_375
+2002/08/25/big/img_729
+2002/07/19/big/img_561
+2003/01/15/big/img_884
+2002/07/25/big/img_891
+2002/08/09/big/img_558
+2002/08/26/big/img_587
+2002/08/13/big/img_1146
+2002/09/02/big/img_15153
+2002/07/26/big/img_316
+2002/08/01/big/img_1940
+2002/08/26/big/img_90
+2003/01/13/big/img_347
+2002/07/25/big/img_520
+2002/08/29/big/img_18718
+2002/08/28/big/img_19219
+2002/08/13/big/img_375
+2002/07/20/big/img_719
+2002/08/31/big/img_17431
+2002/07/28/big/img_192
+2002/08/26/big/img_259
+2002/08/18/big/img_484
+2002/07/29/big/img_580
+2002/07/26/big/img_84
+2002/08/02/big/img_302
+2002/08/31/big/img_17007
+2003/01/15/big/img_543
+2002/09/01/big/img_16488
+2002/08/22/big/img_798
+2002/07/30/big/img_383
+2002/08/04/big/img_668
+2002/08/13/big/img_156
+2002/08/07/big/img_1353
+2002/07/25/big/img_281
+2003/01/14/big/img_587
+2003/01/15/big/img_524
+2002/08/19/big/img_726
+2002/08/21/big/img_709
+2002/08/26/big/img_465
+2002/07/31/big/img_658
+2002/08/28/big/img_19148
+2002/07/23/big/img_423
+2002/08/16/big/img_758
+2002/08/22/big/img_523
+2002/08/16/big/img_591
+2002/08/23/big/img_845
+2002/07/26/big/img_678
+2002/08/09/big/img_806
+2002/08/06/big/img_2369
+2002/07/29/big/img_457
+2002/07/19/big/img_278
+2002/08/30/big/img_18107
+2002/07/26/big/img_444
+2002/08/20/big/img_278
+2002/08/26/big/img_92
+2002/08/26/big/img_257
+2002/07/25/big/img_266
+2002/08/05/big/img_3829
+2002/07/26/big/img_757
+2002/07/29/big/img_1536
+2002/08/09/big/img_472
+2003/01/17/big/img_480
+2002/08/28/big/img_19355
+2002/07/26/big/img_97
+2002/08/06/big/img_2503
+2002/07/19/big/img_254
+2002/08/01/big/img_1470
+2002/08/21/big/img_42
+2002/08/20/big/img_217
+2002/08/06/big/img_2459
+2002/07/19/big/img_552
+2002/08/13/big/img_717
+2002/08/12/big/img_586
+2002/08/20/big/img_411
+2003/01/13/big/img_768
+2002/08/07/big/img_1747
+2002/08/15/big/img_385
+2002/08/01/big/img_1648
+2002/08/15/big/img_311
+2002/08/21/big/img_95
+2002/08/09/big/img_108
+2002/08/21/big/img_398
+2002/08/17/big/img_340
+2002/08/14/big/img_474
+2002/08/13/big/img_294
+2002/08/24/big/img_840
+2002/08/09/big/img_808
+2002/08/23/big/img_491
+2002/07/28/big/img_33
+2003/01/13/big/img_664
+2002/08/02/big/img_261
+2002/08/09/big/img_591
+2002/07/26/big/img_309
+2003/01/14/big/img_372
+2002/08/19/big/img_581
+2002/08/19/big/img_168
+2002/08/26/big/img_422
+2002/07/24/big/img_106
+2002/08/01/big/img_1936
+2002/08/05/big/img_3764
+2002/08/21/big/img_266
+2002/08/31/big/img_17968
+2002/08/01/big/img_1941
+2002/08/15/big/img_550
+2002/08/14/big/img_13
+2002/07/30/big/img_171
+2003/01/13/big/img_490
+2002/07/25/big/img_427
+2002/07/19/big/img_770
+2002/08/12/big/img_759
+2003/01/15/big/img_1360
+2002/08/05/big/img_3692
+2003/01/16/big/img_30
+2002/07/25/big/img_1026
+2002/07/22/big/img_288
+2002/08/29/big/img_18801
+2002/07/24/big/img_793
+2002/08/13/big/img_178
+2002/08/06/big/img_2322
+2003/01/14/big/img_560
+2002/08/18/big/img_408
+2003/01/16/big/img_915
+2003/01/16/big/img_679
+2002/08/07/big/img_1552
+2002/08/29/big/img_19050
+2002/08/01/big/img_2172
+2002/07/31/big/img_30
+2002/07/30/big/img_1019
+2002/07/30/big/img_587
+2003/01/13/big/img_773
+2002/07/30/big/img_410
+2002/07/28/big/img_65
+2002/08/05/big/img_3138
+2002/07/23/big/img_541
+2002/08/22/big/img_963
+2002/07/27/big/img_657
+2002/07/30/big/img_1051
+2003/01/16/big/img_150
+2002/07/31/big/img_519
+2002/08/01/big/img_1961
+2002/08/05/big/img_3752
+2002/07/23/big/img_631
+2003/01/14/big/img_237
+2002/07/28/big/img_21
+2002/07/22/big/img_813
+2002/08/05/big/img_3563
+2003/01/17/big/img_620
+2002/07/19/big/img_523
+2002/07/30/big/img_904
+2002/08/29/big/img_18642
+2002/08/11/big/img_492
+2002/08/01/big/img_2130
+2002/07/25/big/img_618
+2002/08/17/big/img_305
+2003/01/16/big/img_520
+2002/07/26/big/img_495
+2002/08/17/big/img_164
+2002/08/03/big/img_440
+2002/07/24/big/img_441
+2002/08/06/big/img_2146
+2002/08/11/big/img_558
+2002/08/02/big/img_545
+2002/08/31/big/img_18090
+2003/01/01/big/img_136
+2002/07/25/big/img_1099
+2003/01/13/big/img_728
+2003/01/16/big/img_197
+2002/07/26/big/img_651
+2002/08/11/big/img_676
+2003/01/15/big/img_10
+2002/08/21/big/img_250
+2002/08/14/big/img_325
+2002/08/04/big/img_390
+2002/07/24/big/img_554
+2003/01/16/big/img_333
+2002/07/31/big/img_922
+2002/09/02/big/img_15586
+2003/01/16/big/img_184
+2002/07/22/big/img_766
+2002/07/21/big/img_608
+2002/08/07/big/img_1578
+2002/08/17/big/img_961
+2002/07/27/big/img_324
+2002/08/05/big/img_3765
+2002/08/23/big/img_462
+2003/01/16/big/img_382
+2002/08/27/big/img_19838
+2002/08/01/big/img_1505
+2002/08/21/big/img_662
+2002/08/14/big/img_605
+2002/08/19/big/img_816
+2002/07/29/big/img_136
+2002/08/20/big/img_719
+2002/08/06/big/img_2826
+2002/08/10/big/img_630
+2003/01/17/big/img_973
+2002/08/14/big/img_116
+2002/08/02/big/img_666
+2002/08/21/big/img_710
+2002/08/05/big/img_55
+2002/07/31/big/img_229
+2002/08/01/big/img_1549
+2002/07/23/big/img_432
+2002/07/21/big/img_430
+2002/08/21/big/img_549
+2002/08/08/big/img_985
+2002/07/20/big/img_610
+2002/07/23/big/img_978
+2002/08/23/big/img_219
+2002/07/25/big/img_175
+2003/01/15/big/img_230
+2002/08/23/big/img_385
+2002/07/31/big/img_879
+2002/08/12/big/img_495
+2002/08/22/big/img_499
+2002/08/30/big/img_18322
+2002/08/15/big/img_795
+2002/08/13/big/img_835
+2003/01/17/big/img_930
+2002/07/30/big/img_873
+2002/08/11/big/img_257
+2002/07/31/big/img_593
+2002/08/21/big/img_916
+2003/01/13/big/img_814
+2002/07/25/big/img_722
+2002/08/16/big/img_379
+2002/07/31/big/img_497
+2002/07/22/big/img_602
+2002/08/21/big/img_642
+2002/08/21/big/img_614
+2002/08/23/big/img_482
+2002/07/29/big/img_603
+2002/08/13/big/img_705
+2002/07/23/big/img_833
+2003/01/14/big/img_511
+2002/07/24/big/img_376
+2002/08/17/big/img_1030
+2002/08/05/big/img_3576
+2002/08/16/big/img_540
+2002/07/22/big/img_630
+2002/08/10/big/img_180
+2002/08/14/big/img_905
+2002/08/29/big/img_18777
+2002/08/22/big/img_693
+2003/01/16/big/img_933
+2002/08/20/big/img_555
+2002/08/15/big/img_549
+2003/01/14/big/img_830
+2003/01/16/big/img_64
+2002/08/27/big/img_19670
+2002/08/22/big/img_729
+2002/07/27/big/img_981
+2002/08/09/big/img_458
+2003/01/17/big/img_884
+2002/07/25/big/img_639
+2002/08/31/big/img_18008
+2002/08/22/big/img_249
+2002/08/17/big/img_971
+2002/08/04/big/img_308
+2002/07/28/big/img_362
+2002/08/12/big/img_142
+2002/08/26/big/img_61
+2002/08/14/big/img_422
+2002/07/19/big/img_607
+2003/01/15/big/img_717
+2002/08/01/big/img_1475
+2002/08/29/big/img_19061
+2003/01/01/big/img_346
+2002/07/20/big/img_315
+2003/01/15/big/img_756
+2002/08/15/big/img_879
+2002/08/08/big/img_615
+2003/01/13/big/img_431
+2002/08/05/big/img_3233
+2002/08/24/big/img_526
+2003/01/13/big/img_717
+2002/09/01/big/img_16408
+2002/07/22/big/img_217
+2002/07/31/big/img_960
+2002/08/21/big/img_610
+2002/08/05/big/img_3753
+2002/08/03/big/img_151
+2002/08/21/big/img_267
+2002/08/01/big/img_2175
+2002/08/04/big/img_556
+2002/08/21/big/img_527
+2002/09/02/big/img_15800
+2002/07/27/big/img_156
+2002/07/20/big/img_590
+2002/08/15/big/img_700
+2002/08/08/big/img_444
+2002/07/25/big/img_94
+2002/07/24/big/img_778
+2002/08/14/big/img_694
+2002/07/20/big/img_666
+2002/08/02/big/img_200
+2002/08/02/big/img_578
+2003/01/17/big/img_332
+2002/09/01/big/img_16352
+2002/08/27/big/img_19668
+2002/07/23/big/img_823
+2002/08/13/big/img_431
+2003/01/16/big/img_463
+2002/08/27/big/img_19711
+2002/08/23/big/img_154
+2002/07/31/big/img_360
+2002/08/23/big/img_555
+2002/08/10/big/img_561
+2003/01/14/big/img_550
+2002/08/07/big/img_1370
+2002/07/30/big/img_1184
+2002/08/01/big/img_1445
+2002/08/23/big/img_22
+2002/07/30/big/img_606
+2003/01/17/big/img_271
+2002/08/31/big/img_17316
+2002/08/16/big/img_973
+2002/07/26/big/img_77
+2002/07/20/big/img_788
+2002/08/06/big/img_2426
+2002/08/07/big/img_1498
+2002/08/16/big/img_358
+2002/08/06/big/img_2851
+2002/08/12/big/img_359
+2002/08/01/big/img_1521
+2002/08/02/big/img_709
+2002/08/20/big/img_935
+2002/08/12/big/img_188
+2002/08/24/big/img_411
+2002/08/22/big/img_680
+2002/08/06/big/img_2480
+2002/07/20/big/img_627
+2002/07/30/big/img_214
+2002/07/25/big/img_354
+2002/08/02/big/img_636
+2003/01/15/big/img_661
+2002/08/07/big/img_1327
+2002/08/01/big/img_2108
+2002/08/31/big/img_17919
+2002/08/29/big/img_18768
+2002/08/05/big/img_3840
+2002/07/26/big/img_242
+2003/01/14/big/img_451
+2002/08/20/big/img_923
+2002/08/27/big/img_19908
+2002/08/16/big/img_282
+2002/08/19/big/img_440
+2003/01/01/big/img_230
+2002/08/08/big/img_212
+2002/07/20/big/img_443
+2002/08/25/big/img_635
+2003/01/13/big/img_1169
+2002/07/26/big/img_998
+2002/08/15/big/img_995
+2002/08/06/big/img_3002
+2002/07/29/big/img_460
+2003/01/14/big/img_925
+2002/07/23/big/img_539
+2002/08/16/big/img_694
+2003/01/13/big/img_459
+2002/07/23/big/img_249
+2002/08/20/big/img_539
+2002/08/04/big/img_186
+2002/08/26/big/img_264
+2002/07/22/big/img_704
+2002/08/25/big/img_277
+2002/08/22/big/img_988
+2002/07/29/big/img_504
+2002/08/05/big/img_3600
+2002/08/30/big/img_18380
+2003/01/14/big/img_937
+2002/08/21/big/img_254
+2002/08/10/big/img_130
+2002/08/20/big/img_339
+2003/01/14/big/img_428
+2002/08/20/big/img_889
+2002/08/31/big/img_17637
+2002/07/26/big/img_644
+2002/09/01/big/img_16776
+2002/08/06/big/img_2239
+2002/08/06/big/img_2646
+2003/01/13/big/img_491
+2002/08/10/big/img_579
+2002/08/21/big/img_713
+2002/08/22/big/img_482
+2002/07/22/big/img_167
+2002/07/24/big/img_539
+2002/08/14/big/img_721
+2002/07/25/big/img_389
+2002/09/01/big/img_16591
+2002/08/13/big/img_543
+2003/01/14/big/img_432
+2002/08/09/big/img_287
+2002/07/26/big/img_126
+2002/08/23/big/img_412
+2002/08/15/big/img_1034
+2002/08/28/big/img_19485
+2002/07/31/big/img_236
+2002/07/30/big/img_523
+2002/07/19/big/img_141
+2003/01/17/big/img_957
+2002/08/04/big/img_81
+2002/07/25/big/img_206
+2002/08/15/big/img_716
+2002/08/13/big/img_403
+2002/08/15/big/img_685
+2002/07/26/big/img_884
+2002/07/19/big/img_499
+2002/07/23/big/img_772
+2002/07/27/big/img_752
+2003/01/14/big/img_493
+2002/08/25/big/img_664
+2002/07/31/big/img_334
+2002/08/26/big/img_678
+2002/09/01/big/img_16541
+2003/01/14/big/img_347
+2002/07/23/big/img_187
+2002/07/30/big/img_1163
+2002/08/05/big/img_35
+2002/08/22/big/img_944
+2002/08/07/big/img_1239
+2002/07/29/big/img_1215
+2002/08/03/big/img_312
+2002/08/05/big/img_3523
+2002/07/29/big/img_218
+2002/08/13/big/img_672
+2002/08/16/big/img_205
+2002/08/17/big/img_594
+2002/07/29/big/img_1411
+2002/07/30/big/img_942
+2003/01/16/big/img_312
+2002/08/08/big/img_312
+2002/07/25/big/img_15
+2002/08/09/big/img_839
+2002/08/01/big/img_2069
+2002/08/31/big/img_17512
+2002/08/01/big/img_3
+2002/07/31/big/img_320
+2003/01/15/big/img_1265
+2002/08/14/big/img_563
+2002/07/31/big/img_167
+2002/08/20/big/img_374
+2002/08/13/big/img_406
+2002/08/08/big/img_625
+2002/08/02/big/img_314
+2002/08/27/big/img_19964
+2002/09/01/big/img_16670
+2002/07/31/big/img_599
+2002/08/29/big/img_18906
+2002/07/24/big/img_373
+2002/07/26/big/img_513
+2002/09/02/big/img_15497
+2002/08/19/big/img_117
+2003/01/01/big/img_158
+2002/08/24/big/img_178
+2003/01/13/big/img_935
+2002/08/13/big/img_609
+2002/08/30/big/img_18341
+2002/08/25/big/img_674
+2003/01/13/big/img_209
+2002/08/13/big/img_258
+2002/08/05/big/img_3543
+2002/08/07/big/img_1970
+2002/08/06/big/img_3004
+2003/01/17/big/img_487
+2002/08/24/big/img_873
+2002/08/29/big/img_18730
+2002/08/09/big/img_375
+2003/01/16/big/img_751
+2002/08/02/big/img_603
+2002/08/19/big/img_325
+2002/09/01/big/img_16420
+2002/08/05/big/img_3633
+2002/08/21/big/img_516
+2002/07/19/big/img_501
+2002/07/26/big/img_688
+2002/07/24/big/img_256
+2002/07/25/big/img_438
+2002/07/31/big/img_1017
+2002/08/22/big/img_512
+2002/07/21/big/img_543
+2002/08/08/big/img_223
+2002/08/19/big/img_189
+2002/08/12/big/img_630
+2002/07/30/big/img_958
+2002/07/28/big/img_208
+2002/08/31/big/img_17691
+2002/07/22/big/img_542
+2002/07/19/big/img_741
+2002/07/19/big/img_158
+2002/08/15/big/img_399
+2002/08/01/big/img_2159
+2002/08/14/big/img_455
+2002/08/17/big/img_1011
+2002/08/26/big/img_744
+2002/08/12/big/img_624
+2003/01/17/big/img_821
+2002/08/16/big/img_980
+2002/07/28/big/img_281
+2002/07/25/big/img_171
+2002/08/03/big/img_116
+2002/07/22/big/img_467
+2002/07/31/big/img_750
+2002/07/26/big/img_435
+2002/07/19/big/img_822
+2002/08/13/big/img_626
+2002/08/11/big/img_344
+2002/08/02/big/img_473
+2002/09/01/big/img_16817
+2002/08/01/big/img_1275
+2002/08/28/big/img_19270
+2002/07/23/big/img_607
+2002/08/09/big/img_316
+2002/07/29/big/img_626
+2002/07/24/big/img_824
+2002/07/22/big/img_342
+2002/08/08/big/img_794
+2002/08/07/big/img_1209
+2002/07/19/big/img_18
+2002/08/25/big/img_634
+2002/07/24/big/img_730
+2003/01/17/big/img_356
+2002/07/23/big/img_305
+2002/07/30/big/img_453
+2003/01/13/big/img_972
+2002/08/06/big/img_2610
+2002/08/29/big/img_18920
+2002/07/31/big/img_123
+2002/07/26/big/img_979
+2002/08/24/big/img_635
+2002/08/05/big/img_3704
+2002/08/07/big/img_1358
+2002/07/22/big/img_306
+2002/08/13/big/img_619
+2002/08/02/big/img_366

video-tetalking/third_part/GPEN/face_detect/data/__init__.py

@@ -0,0 +1,3 @@
+from .wider_face import WiderFaceDetection, detection_collate
+from .data_augment import *
+from .config import *

video-tetalking/third_part/GPEN/face_detect/data/config.py

@@ -0,0 +1,42 @@
+# config.py
+
+cfg_mnet = {
+    'name': 'mobilenet0.25',
+    'min_sizes': [[16, 32], [64, 128], [256, 512]],
+    'steps': [8, 16, 32],
+    'variance': [0.1, 0.2],
+    'clip': False,
+    'loc_weight': 2.0,
+    'gpu_train': True,
+    'batch_size': 32,
+    'ngpu': 1,
+    'epoch': 250,
+    'decay1': 190,
+    'decay2': 220,
+    'image_size': 640,
+    'pretrain': False,
+    'return_layers': {'stage1': 1, 'stage2': 2, 'stage3': 3},
+    'in_channel': 32,
+    'out_channel': 64
+}
+
+cfg_re50 = {
+    'name': 'Resnet50',
+    'min_sizes': [[16, 32], [64, 128], [256, 512]],
+    'steps': [8, 16, 32],
+    'variance': [0.1, 0.2],
+    'clip': False,
+    'loc_weight': 2.0,
+    'gpu_train': True,
+    'batch_size': 24,
+    'ngpu': 4,
+    'epoch': 100,
+    'decay1': 70,
+    'decay2': 90,
+    'image_size': 840,
+    'pretrain': False,
+    'return_layers': {'layer2': 1, 'layer3': 2, 'layer4': 3},
+    'in_channel': 256,
+    'out_channel': 256
+}
+

video-tetalking/third_part/GPEN/face_detect/data/data_augment.py

@@ -0,0 +1,237 @@
+import cv2
+import numpy as np
+import random
+from face_detect.utils.box_utils import matrix_iof
+
+
+def _crop(image, boxes, labels, landm, img_dim):
+    height, width, _ = image.shape
+    pad_image_flag = True
+
+    for _ in range(250):
+        """
+        if random.uniform(0, 1) <= 0.2:
+            scale = 1.0
+        else:
+            scale = random.uniform(0.3, 1.0)
+        """
+        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
+        scale = random.choice(PRE_SCALES)
+        short_side = min(width, height)
+        w = int(scale * short_side)
+        h = w
+
+        if width == w:
+            l = 0
+        else:
+            l = random.randrange(width - w)
+        if height == h:
+            t = 0
+        else:
+            t = random.randrange(height - h)
+        roi = np.array((l, t, l + w, t + h))
+
+        value = matrix_iof(boxes, roi[np.newaxis])
+        flag = (value >= 1)
+        if not flag.any():
+            continue
+
+        centers = (boxes[:, :2] + boxes[:, 2:]) / 2
+        mask_a = np.logical_and(roi[:2] < centers, centers < roi[2:]).all(axis=1)
+        boxes_t = boxes[mask_a].copy()
+        labels_t = labels[mask_a].copy()
+        landms_t = landm[mask_a].copy()
+        landms_t = landms_t.reshape([-1, 5, 2])
+
+        if boxes_t.shape[0] == 0:
+            continue
+
+        image_t = image[roi[1]:roi[3], roi[0]:roi[2]]
+
+        boxes_t[:, :2] = np.maximum(boxes_t[:, :2], roi[:2])
+        boxes_t[:, :2] -= roi[:2]
+        boxes_t[:, 2:] = np.minimum(boxes_t[:, 2:], roi[2:])
+        boxes_t[:, 2:] -= roi[:2]
+
+        # landm
+        landms_t[:, :, :2] = landms_t[:, :, :2] - roi[:2]
+        landms_t[:, :, :2] = np.maximum(landms_t[:, :, :2], np.array([0, 0]))
+        landms_t[:, :, :2] = np.minimum(landms_t[:, :, :2], roi[2:] - roi[:2])
+        landms_t = landms_t.reshape([-1, 10])
+
+
+	# make sure that the cropped image contains at least one face > 16 pixel at training image scale
+        b_w_t = (boxes_t[:, 2] - boxes_t[:, 0] + 1) / w * img_dim
+        b_h_t = (boxes_t[:, 3] - boxes_t[:, 1] + 1) / h * img_dim
+        mask_b = np.minimum(b_w_t, b_h_t) > 0.0
+        boxes_t = boxes_t[mask_b]
+        labels_t = labels_t[mask_b]
+        landms_t = landms_t[mask_b]
+
+        if boxes_t.shape[0] == 0:
+            continue
+
+        pad_image_flag = False
+
+        return image_t, boxes_t, labels_t, landms_t, pad_image_flag
+    return image, boxes, labels, landm, pad_image_flag
+
+
+def _distort(image):
+
+    def _convert(image, alpha=1, beta=0):
+        tmp = image.astype(float) * alpha + beta
+        tmp[tmp < 0] = 0
+        tmp[tmp > 255] = 255
+        image[:] = tmp
+
+    image = image.copy()
+
+    if random.randrange(2):
+
+        #brightness distortion
+        if random.randrange(2):
+            _convert(image, beta=random.uniform(-32, 32))
+
+        #contrast distortion
+        if random.randrange(2):
+            _convert(image, alpha=random.uniform(0.5, 1.5))
+
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+
+        #saturation distortion
+        if random.randrange(2):
+            _convert(image[:, :, 1], alpha=random.uniform(0.5, 1.5))
+
+        #hue distortion
+        if random.randrange(2):
+            tmp = image[:, :, 0].astype(int) + random.randint(-18, 18)
+            tmp %= 180
+            image[:, :, 0] = tmp
+
+        image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
+
+    else:
+
+        #brightness distortion
+        if random.randrange(2):
+            _convert(image, beta=random.uniform(-32, 32))
+
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+
+        #saturation distortion
+        if random.randrange(2):
+            _convert(image[:, :, 1], alpha=random.uniform(0.5, 1.5))
+
+        #hue distortion
+        if random.randrange(2):
+            tmp = image[:, :, 0].astype(int) + random.randint(-18, 18)
+            tmp %= 180
+            image[:, :, 0] = tmp
+
+        image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
+
+        #contrast distortion
+        if random.randrange(2):
+            _convert(image, alpha=random.uniform(0.5, 1.5))
+
+    return image
+
+
+def _expand(image, boxes, fill, p):
+    if random.randrange(2):
+        return image, boxes
+
+    height, width, depth = image.shape
+
+    scale = random.uniform(1, p)
+    w = int(scale * width)
+    h = int(scale * height)
+
+    left = random.randint(0, w - width)
+    top = random.randint(0, h - height)
+
+    boxes_t = boxes.copy()
+    boxes_t[:, :2] += (left, top)
+    boxes_t[:, 2:] += (left, top)
+    expand_image = np.empty(
+        (h, w, depth),
+        dtype=image.dtype)
+    expand_image[:, :] = fill
+    expand_image[top:top + height, left:left + width] = image
+    image = expand_image
+
+    return image, boxes_t
+
+
+def _mirror(image, boxes, landms):
+    _, width, _ = image.shape
+    if random.randrange(2):
+        image = image[:, ::-1]
+        boxes = boxes.copy()
+        boxes[:, 0::2] = width - boxes[:, 2::-2]
+
+        # landm
+        landms = landms.copy()
+        landms = landms.reshape([-1, 5, 2])
+        landms[:, :, 0] = width - landms[:, :, 0]
+        tmp = landms[:, 1, :].copy()
+        landms[:, 1, :] = landms[:, 0, :]
+        landms[:, 0, :] = tmp
+        tmp1 = landms[:, 4, :].copy()
+        landms[:, 4, :] = landms[:, 3, :]
+        landms[:, 3, :] = tmp1
+        landms = landms.reshape([-1, 10])
+
+    return image, boxes, landms
+
+
+def _pad_to_square(image, rgb_mean, pad_image_flag):
+    if not pad_image_flag:
+        return image
+    height, width, _ = image.shape
+    long_side = max(width, height)
+    image_t = np.empty((long_side, long_side, 3), dtype=image.dtype)
+    image_t[:, :] = rgb_mean
+    image_t[0:0 + height, 0:0 + width] = image
+    return image_t
+
+
+def _resize_subtract_mean(image, insize, rgb_mean):
+    interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST, cv2.INTER_LANCZOS4]
+    interp_method = interp_methods[random.randrange(5)]
+    image = cv2.resize(image, (insize, insize), interpolation=interp_method)
+    image = image.astype(np.float32)
+    image -= rgb_mean
+    return image.transpose(2, 0, 1)
+
+
+class preproc(object):
+
+    def __init__(self, img_dim, rgb_means):
+        self.img_dim = img_dim
+        self.rgb_means = rgb_means
+
+    def __call__(self, image, targets):
+        assert targets.shape[0] > 0, "this image does not have gt"
+
+        boxes = targets[:, :4].copy()
+        labels = targets[:, -1].copy()
+        landm = targets[:, 4:-1].copy()
+
+        image_t, boxes_t, labels_t, landm_t, pad_image_flag = _crop(image, boxes, labels, landm, self.img_dim)
+        image_t = _distort(image_t)
+        image_t = _pad_to_square(image_t,self.rgb_means, pad_image_flag)
+        image_t, boxes_t, landm_t = _mirror(image_t, boxes_t, landm_t)
+        height, width, _ = image_t.shape
+        image_t = _resize_subtract_mean(image_t, self.img_dim, self.rgb_means)
+        boxes_t[:, 0::2] /= width
+        boxes_t[:, 1::2] /= height
+
+        landm_t[:, 0::2] /= width
+        landm_t[:, 1::2] /= height
+
+        labels_t = np.expand_dims(labels_t, 1)
+        targets_t = np.hstack((boxes_t, landm_t, labels_t))
+
+        return image_t, targets_t

video-tetalking/third_part/GPEN/face_detect/data/wider_face.py

@@ -0,0 +1,101 @@
+import os
+import os.path
+import sys
+import torch
+import torch.utils.data as data
+import cv2
+import numpy as np
+
+class WiderFaceDetection(data.Dataset):
+    def __init__(self, txt_path, preproc=None):
+        self.preproc = preproc
+        self.imgs_path = []
+        self.words = []
+        f = open(txt_path,'r')
+        lines = f.readlines()
+        isFirst = True
+        labels = []
+        for line in lines:
+            line = line.rstrip()
+            if line.startswith('#'):
+                if isFirst is True:
+                    isFirst = False
+                else:
+                    labels_copy = labels.copy()
+                    self.words.append(labels_copy)
+                    labels.clear()
+                path = line[2:]
+                path = txt_path.replace('label.txt','images/') + path
+                self.imgs_path.append(path)
+            else:
+                line = line.split(' ')
+                label = [float(x) for x in line]
+                labels.append(label)
+
+        self.words.append(labels)
+
+    def __len__(self):
+        return len(self.imgs_path)
+
+    def __getitem__(self, index):
+        img = cv2.imread(self.imgs_path[index])
+        height, width, _ = img.shape
+
+        labels = self.words[index]
+        annotations = np.zeros((0, 15))
+        if len(labels) == 0:
+            return annotations
+        for idx, label in enumerate(labels):
+            annotation = np.zeros((1, 15))
+            # bbox
+            annotation[0, 0] = label[0]  # x1
+            annotation[0, 1] = label[1]  # y1
+            annotation[0, 2] = label[0] + label[2]  # x2
+            annotation[0, 3] = label[1] + label[3]  # y2
+
+            # landmarks
+            annotation[0, 4] = label[4]    # l0_x
+            annotation[0, 5] = label[5]    # l0_y
+            annotation[0, 6] = label[7]    # l1_x
+            annotation[0, 7] = label[8]    # l1_y
+            annotation[0, 8] = label[10]   # l2_x
+            annotation[0, 9] = label[11]   # l2_y
+            annotation[0, 10] = label[13]  # l3_x
+            annotation[0, 11] = label[14]  # l3_y
+            annotation[0, 12] = label[16]  # l4_x
+            annotation[0, 13] = label[17]  # l4_y
+            if (annotation[0, 4]<0):
+                annotation[0, 14] = -1
+            else:
+                annotation[0, 14] = 1
+
+            annotations = np.append(annotations, annotation, axis=0)
+        target = np.array(annotations)
+        if self.preproc is not None:
+            img, target = self.preproc(img, target)
+
+        return torch.from_numpy(img), target
+
+def detection_collate(batch):
+    """Custom collate fn for dealing with batches of images that have a different
+    number of associated object annotations (bounding boxes).
+
+    Arguments:
+        batch: (tuple) A tuple of tensor images and lists of annotations
+
+    Return:
+        A tuple containing:
+            1) (tensor) batch of images stacked on their 0 dim
+            2) (list of tensors) annotations for a given image are stacked on 0 dim
+    """
+    targets = []
+    imgs = []
+    for _, sample in enumerate(batch):
+        for _, tup in enumerate(sample):
+            if torch.is_tensor(tup):
+                imgs.append(tup)
+            elif isinstance(tup, type(np.empty(0))):
+                annos = torch.from_numpy(tup).float()
+                targets.append(annos)
+
+    return (torch.stack(imgs, 0), targets)

video-tetalking/third_part/GPEN/face_detect/facemodels/net.py

@@ -0,0 +1,137 @@
+import time
+import torch
+import torch.nn as nn
+import torchvision.models._utils as _utils
+import torchvision.models as models
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+def conv_bn(inp, oup, stride = 1, leaky = 0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True)
+    )
+
+def conv_bn_no_relu(inp, oup, stride):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        nn.BatchNorm2d(oup),
+    )
+
+def conv_bn1X1(inp, oup, stride, leaky=0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 1, stride, padding=0, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True)
+    )
+
+def conv_dw(inp, oup, stride, leaky=0.1):
+    return nn.Sequential(
+        nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
+        nn.BatchNorm2d(inp),
+        nn.LeakyReLU(negative_slope= leaky,inplace=True),
+
+        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope= leaky,inplace=True),
+    )
+
+class SSH(nn.Module):
+    def __init__(self, in_channel, out_channel):
+        super(SSH, self).__init__()
+        assert out_channel % 4 == 0
+        leaky = 0
+        if (out_channel <= 64):
+            leaky = 0.1
+        self.conv3X3 = conv_bn_no_relu(in_channel, out_channel//2, stride=1)
+
+        self.conv5X5_1 = conv_bn(in_channel, out_channel//4, stride=1, leaky = leaky)
+        self.conv5X5_2 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)
+
+        self.conv7X7_2 = conv_bn(out_channel//4, out_channel//4, stride=1, leaky = leaky)
+        self.conv7x7_3 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)
+
+    def forward(self, input):
+        conv3X3 = self.conv3X3(input)
+
+        conv5X5_1 = self.conv5X5_1(input)
+        conv5X5 = self.conv5X5_2(conv5X5_1)
+
+        conv7X7_2 = self.conv7X7_2(conv5X5_1)
+        conv7X7 = self.conv7x7_3(conv7X7_2)
+
+        out = torch.cat([conv3X3, conv5X5, conv7X7], dim=1)
+        out = F.relu(out)
+        return out
+
+class FPN(nn.Module):
+    def __init__(self,in_channels_list,out_channels):
+        super(FPN,self).__init__()
+        leaky = 0
+        if (out_channels <= 64):
+            leaky = 0.1
+        self.output1 = conv_bn1X1(in_channels_list[0], out_channels, stride = 1, leaky = leaky)
+        self.output2 = conv_bn1X1(in_channels_list[1], out_channels, stride = 1, leaky = leaky)
+        self.output3 = conv_bn1X1(in_channels_list[2], out_channels, stride = 1, leaky = leaky)
+
+        self.merge1 = conv_bn(out_channels, out_channels, leaky = leaky)
+        self.merge2 = conv_bn(out_channels, out_channels, leaky = leaky)
+
+    def forward(self, input):
+        # names = list(input.keys())
+        input = list(input.values())
+
+        output1 = self.output1(input[0])
+        output2 = self.output2(input[1])
+        output3 = self.output3(input[2])
+
+        up3 = F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode="nearest")
+        output2 = output2 + up3
+        output2 = self.merge2(output2)
+
+        up2 = F.interpolate(output2, size=[output1.size(2), output1.size(3)], mode="nearest")
+        output1 = output1 + up2
+        output1 = self.merge1(output1)
+
+        out = [output1, output2, output3]
+        return out
+
+
+
+class MobileNetV1(nn.Module):
+    def __init__(self):
+        super(MobileNetV1, self).__init__()
+        self.stage1 = nn.Sequential(
+            conv_bn(3, 8, 2, leaky = 0.1),    # 3
+            conv_dw(8, 16, 1),   # 7
+            conv_dw(16, 32, 2),  # 11
+            conv_dw(32, 32, 1),  # 19
+            conv_dw(32, 64, 2),  # 27
+            conv_dw(64, 64, 1),  # 43
+        )
+        self.stage2 = nn.Sequential(
+            conv_dw(64, 128, 2),  # 43 + 16 = 59
+            conv_dw(128, 128, 1), # 59 + 32 = 91
+            conv_dw(128, 128, 1), # 91 + 32 = 123
+            conv_dw(128, 128, 1), # 123 + 32 = 155
+            conv_dw(128, 128, 1), # 155 + 32 = 187
+            conv_dw(128, 128, 1), # 187 + 32 = 219
+        )
+        self.stage3 = nn.Sequential(
+            conv_dw(128, 256, 2), # 219 +3 2 = 241
+            conv_dw(256, 256, 1), # 241 + 64 = 301
+        )
+        self.avg = nn.AdaptiveAvgPool2d((1,1))
+        self.fc = nn.Linear(256, 1000)
+
+    def forward(self, x):
+        x = self.stage1(x)
+        x = self.stage2(x)
+        x = self.stage3(x)
+        x = self.avg(x)
+        # x = self.model(x)
+        x = x.view(-1, 256)
+        x = self.fc(x)
+        return x
+

video-tetalking/third_part/GPEN/face_detect/facemodels/retinaface.py

@@ -0,0 +1,127 @@
+import torch
+import torch.nn as nn
+import torchvision.models.detection.backbone_utils as backbone_utils
+import torchvision.models._utils as _utils
+import torch.nn.functional as F
+from collections import OrderedDict
+
+from face_detect.facemodels.net import MobileNetV1 as MobileNetV1
+from face_detect.facemodels.net import FPN as FPN
+from face_detect.facemodels.net import SSH as SSH
+
+
+
+class ClassHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(ClassHead,self).__init__()
+        self.num_anchors = num_anchors
+        self.conv1x1 = nn.Conv2d(inchannels,self.num_anchors*2,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+        
+        return out.view(out.shape[0], -1, 2)
+
+class BboxHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(BboxHead,self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels,num_anchors*4,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+
+        return out.view(out.shape[0], -1, 4)
+
+class LandmarkHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(LandmarkHead,self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels,num_anchors*10,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+
+        return out.view(out.shape[0], -1, 10)
+
+class RetinaFace(nn.Module):
+    def __init__(self, cfg = None, phase = 'train'):
+        """
+        :param cfg:  Network related settings.
+        :param phase: train or test.
+        """
+        super(RetinaFace,self).__init__()
+        self.phase = phase
+        backbone = None
+        if cfg['name'] == 'mobilenet0.25':
+            backbone = MobileNetV1()
+            if cfg['pretrain']:
+                checkpoint = torch.load("./weights/mobilenetV1X0.25_pretrain.tar", map_location=torch.device('cpu'))
+                from collections import OrderedDict
+                new_state_dict = OrderedDict()
+                for k, v in checkpoint['state_dict'].items():
+                    name = k[7:]  # remove module.
+                    new_state_dict[name] = v
+                # load params
+                backbone.load_state_dict(new_state_dict)
+        elif cfg['name'] == 'Resnet50':
+            import torchvision.models as models
+            backbone = models.resnet50(pretrained=cfg['pretrain'])
+
+        self.body = _utils.IntermediateLayerGetter(backbone, cfg['return_layers'])
+        in_channels_stage2 = cfg['in_channel']
+        in_channels_list = [
+            in_channels_stage2 * 2,
+            in_channels_stage2 * 4,
+            in_channels_stage2 * 8,
+        ]
+        out_channels = cfg['out_channel']
+        self.fpn = FPN(in_channels_list,out_channels)
+        self.ssh1 = SSH(out_channels, out_channels)
+        self.ssh2 = SSH(out_channels, out_channels)
+        self.ssh3 = SSH(out_channels, out_channels)
+
+        self.ClassHead = self._make_class_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.BboxHead = self._make_bbox_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.LandmarkHead = self._make_landmark_head(fpn_num=3, inchannels=cfg['out_channel'])
+
+    def _make_class_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        classhead = nn.ModuleList()
+        for i in range(fpn_num):
+            classhead.append(ClassHead(inchannels,anchor_num))
+        return classhead
+    
+    def _make_bbox_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        bboxhead = nn.ModuleList()
+        for i in range(fpn_num):
+            bboxhead.append(BboxHead(inchannels,anchor_num))
+        return bboxhead
+
+    def _make_landmark_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        landmarkhead = nn.ModuleList()
+        for i in range(fpn_num):
+            landmarkhead.append(LandmarkHead(inchannels,anchor_num))
+        return landmarkhead
+
+    def forward(self,inputs):
+        out = self.body(inputs)
+
+        # FPN
+        fpn = self.fpn(out)
+
+        # SSH
+        feature1 = self.ssh1(fpn[0])
+        feature2 = self.ssh2(fpn[1])
+        feature3 = self.ssh3(fpn[2])
+        features = [feature1, feature2, feature3]
+
+        bbox_regressions = torch.cat([self.BboxHead[i](feature) for i, feature in enumerate(features)], dim=1)
+        classifications = torch.cat([self.ClassHead[i](feature) for i, feature in enumerate(features)],dim=1)
+        ldm_regressions = torch.cat([self.LandmarkHead[i](feature) for i, feature in enumerate(features)], dim=1)
+
+        if self.phase == 'train':
+            output = (bbox_regressions, classifications, ldm_regressions)
+        else:
+            output = (bbox_regressions, F.softmax(classifications, dim=-1), ldm_regressions)
+        return output

video-tetalking/third_part/GPEN/face_detect/layers/__init__.py

@@ -0,0 +1,2 @@
+from .functions import *
+from .modules import *

video-tetalking/third_part/GPEN/face_detect/layers/functions/prior_box.py

@@ -0,0 +1,34 @@
+import torch
+from itertools import product as product
+import numpy as np
+from math import ceil
+
+
+class PriorBox(object):
+    def __init__(self, cfg, image_size=None, phase='train'):
+        super(PriorBox, self).__init__()
+        self.min_sizes = cfg['min_sizes']
+        self.steps = cfg['steps']
+        self.clip = cfg['clip']
+        self.image_size = image_size
+        self.feature_maps = [[ceil(self.image_size[0]/step), ceil(self.image_size[1]/step)] for step in self.steps]
+        self.name = "s"
+
+    def forward(self):
+        anchors = []
+        for k, f in enumerate(self.feature_maps):
+            min_sizes = self.min_sizes[k]
+            for i, j in product(range(f[0]), range(f[1])):
+                for min_size in min_sizes:
+                    s_kx = min_size / self.image_size[1]
+                    s_ky = min_size / self.image_size[0]
+                    dense_cx = [x * self.steps[k] / self.image_size[1] for x in [j + 0.5]]
+                    dense_cy = [y * self.steps[k] / self.image_size[0] for y in [i + 0.5]]
+                    for cy, cx in product(dense_cy, dense_cx):
+                        anchors += [cx, cy, s_kx, s_ky]
+
+        # back to torch land
+        output = torch.Tensor(anchors).view(-1, 4)
+        if self.clip:
+            output.clamp_(max=1, min=0)
+        return output

video-tetalking/third_part/GPEN/face_detect/layers/modules/__init__.py

@@ -0,0 +1,3 @@
+from .multibox_loss import MultiBoxLoss
+
+__all__ = ['MultiBoxLoss']

video-tetalking/third_part/GPEN/face_detect/layers/modules/multibox_loss.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+from face_detect.utils.box_utils import match, log_sum_exp
+from face_detect.data import cfg_mnet
+GPU = cfg_mnet['gpu_train']
+
+class MultiBoxLoss(nn.Module):
+    """SSD Weighted Loss Function
+    Compute Targets:
+        1) Produce Confidence Target Indices by matching  ground truth boxes
+           with (default) 'priorboxes' that have jaccard index > threshold parameter
+           (default threshold: 0.5).
+        2) Produce localization target by 'encoding' variance into offsets of ground
+           truth boxes and their matched  'priorboxes'.
+        3) Hard negative mining to filter the excessive number of negative examples
+           that comes with using a large number of default bounding boxes.
+           (default negative:positive ratio 3:1)
+    Objective Loss:
+        L(x,c,l,g) = (Lconf(x, c) + αLloc(x,l,g)) / N
+        Where, Lconf is the CrossEntropy Loss and Lloc is the SmoothL1 Loss
+        weighted by α which is set to 1 by cross val.
+        Args:
+            c: class confidences,
+            l: predicted boxes,
+            g: ground truth boxes
+            N: number of matched default boxes
+        See: https://arxiv.org/pdf/1512.02325.pdf for more details.
+    """
+
+    def __init__(self, num_classes, overlap_thresh, prior_for_matching, bkg_label, neg_mining, neg_pos, neg_overlap, encode_target):
+        super(MultiBoxLoss, self).__init__()
+        self.num_classes = num_classes
+        self.threshold = overlap_thresh
+        self.background_label = bkg_label
+        self.encode_target = encode_target
+        self.use_prior_for_matching = prior_for_matching
+        self.do_neg_mining = neg_mining
+        self.negpos_ratio = neg_pos
+        self.neg_overlap = neg_overlap
+        self.variance = [0.1, 0.2]
+
+    def forward(self, predictions, priors, targets):
+        """Multibox Loss
+        Args:
+            predictions (tuple): A tuple containing loc preds, conf preds,
+            and prior boxes from SSD net.
+                conf shape: torch.size(batch_size,num_priors,num_classes)
+                loc shape: torch.size(batch_size,num_priors,4)
+                priors shape: torch.size(num_priors,4)
+
+            ground_truth (tensor): Ground truth boxes and labels for a batch,
+                shape: [batch_size,num_objs,5] (last idx is the label).
+        """
+
+        loc_data, conf_data, landm_data = predictions
+        priors = priors
+        num = loc_data.size(0)
+        num_priors = (priors.size(0))
+
+        # match priors (default boxes) and ground truth boxes
+        loc_t = torch.Tensor(num, num_priors, 4)
+        landm_t = torch.Tensor(num, num_priors, 10)
+        conf_t = torch.LongTensor(num, num_priors)
+        for idx in range(num):
+            truths = targets[idx][:, :4].data
+            labels = targets[idx][:, -1].data
+            landms = targets[idx][:, 4:14].data
+            defaults = priors.data
+            match(self.threshold, truths, defaults, self.variance, labels, landms, loc_t, conf_t, landm_t, idx)
+        if GPU:
+            loc_t = loc_t.cuda()
+            conf_t = conf_t.cuda()
+            landm_t = landm_t.cuda()
+
+        zeros = torch.tensor(0).cuda()
+        # landm Loss (Smooth L1)
+        # Shape: [batch,num_priors,10]
+        pos1 = conf_t > zeros
+        num_pos_landm = pos1.long().sum(1, keepdim=True)
+        N1 = max(num_pos_landm.data.sum().float(), 1)
+        pos_idx1 = pos1.unsqueeze(pos1.dim()).expand_as(landm_data)
+        landm_p = landm_data[pos_idx1].view(-1, 10)
+        landm_t = landm_t[pos_idx1].view(-1, 10)
+        loss_landm = F.smooth_l1_loss(landm_p, landm_t, reduction='sum')
+
+
+        pos = conf_t != zeros
+        conf_t[pos] = 1
+
+        # Localization Loss (Smooth L1)
+        # Shape: [batch,num_priors,4]
+        pos_idx = pos.unsqueeze(pos.dim()).expand_as(loc_data)
+        loc_p = loc_data[pos_idx].view(-1, 4)
+        loc_t = loc_t[pos_idx].view(-1, 4)
+        loss_l = F.smooth_l1_loss(loc_p, loc_t, reduction='sum')
+
+        # Compute max conf across batch for hard negative mining
+        batch_conf = conf_data.view(-1, self.num_classes)
+        loss_c = log_sum_exp(batch_conf) - batch_conf.gather(1, conf_t.view(-1, 1))
+
+        # Hard Negative Mining
+        loss_c[pos.view(-1, 1)] = 0 # filter out pos boxes for now
+        loss_c = loss_c.view(num, -1)
+        _, loss_idx = loss_c.sort(1, descending=True)
+        _, idx_rank = loss_idx.sort(1)
+        num_pos = pos.long().sum(1, keepdim=True)
+        num_neg = torch.clamp(self.negpos_ratio*num_pos, max=pos.size(1)-1)
+        neg = idx_rank < num_neg.expand_as(idx_rank)
+
+        # Confidence Loss Including Positive and Negative Examples
+        pos_idx = pos.unsqueeze(2).expand_as(conf_data)
+        neg_idx = neg.unsqueeze(2).expand_as(conf_data)
+        conf_p = conf_data[(pos_idx+neg_idx).gt(0)].view(-1,self.num_classes)
+        targets_weighted = conf_t[(pos+neg).gt(0)]
+        loss_c = F.cross_entropy(conf_p, targets_weighted, reduction='sum')
+
+        # Sum of losses: L(x,c,l,g) = (Lconf(x, c) + αLloc(x,l,g)) / N
+        N = max(num_pos.data.sum().float(), 1)
+        loss_l /= N
+        loss_c /= N
+        loss_landm /= N1
+
+        return loss_l, loss_c, loss_landm

video-tetalking/third_part/GPEN/face_detect/retinaface_detection.py

@@ -0,0 +1,193 @@
+'''
+@paper: GAN Prior Embedded Network for Blind Face Restoration in the Wild (CVPR2021)
+@author: yangxy (yangtao9009@gmail.com)
+'''
+import os
+import torch
+import torch.backends.cudnn as cudnn
+import numpy as np
+from face_detect.data import cfg_re50
+from face_detect.layers.functions.prior_box import PriorBox
+from face_detect.utils.nms.py_cpu_nms import py_cpu_nms
+import cv2
+from face_detect.facemodels.retinaface import RetinaFace
+from face_detect.utils.box_utils import decode, decode_landm
+import time
+import torch.nn.functional as F
+
+
+class RetinaFaceDetection(object):
+    def __init__(self, base_dir, device='cuda', network='RetinaFace-R50'):
+        torch.set_grad_enabled(False)
+        cudnn.benchmark = True
+        self.pretrained_path = os.path.join(base_dir, network+'.pth')
+        self.device = device #torch.cuda.current_device()
+        self.cfg = cfg_re50
+        self.net = RetinaFace(cfg=self.cfg, phase='test')
+        self.load_model()
+        self.net = self.net.to(device)
+
+        self.mean = torch.tensor([[[[104]], [[117]], [[123]]]]).to(device)
+
+    def check_keys(self, pretrained_state_dict):
+        ckpt_keys = set(pretrained_state_dict.keys())
+        model_keys = set(self.net.state_dict().keys())
+        used_pretrained_keys = model_keys & ckpt_keys
+        unused_pretrained_keys = ckpt_keys - model_keys
+        missing_keys = model_keys - ckpt_keys
+        assert len(used_pretrained_keys) > 0, 'load NONE from pretrained checkpoint'
+        return True
+
+    def remove_prefix(self, state_dict, prefix):
+        ''' Old style model is stored with all names of parameters sharing common prefix 'module.' '''
+        f = lambda x: x.split(prefix, 1)[-1] if x.startswith(prefix) else x
+        return {f(key): value for key, value in state_dict.items()}
+
+    def load_model(self, load_to_cpu=False):
+        #if load_to_cpu:
+        #    pretrained_dict = torch.load(self.pretrained_path, map_location=lambda storage, loc: storage)
+        #else:
+        #    pretrained_dict = torch.load(self.pretrained_path, map_location=lambda storage, loc: storage.cuda())
+        pretrained_dict = torch.load(self.pretrained_path, map_location=torch.device('cpu'))
+        if "state_dict" in pretrained_dict.keys():
+            pretrained_dict = self.remove_prefix(pretrained_dict['state_dict'], 'module.')
+        else:
+            pretrained_dict = self.remove_prefix(pretrained_dict, 'module.')
+        self.check_keys(pretrained_dict)
+        self.net.load_state_dict(pretrained_dict, strict=False)
+        self.net.eval()
+    
+    def detect(self, img_raw, resize=1, confidence_threshold=0.9, nms_threshold=0.4, top_k=5000, keep_top_k=750, save_image=False):
+        img = np.float32(img_raw)
+
+        im_height, im_width = img.shape[:2]
+        ss = 1.0
+        # tricky
+        if max(im_height, im_width) > 1500:
+            ss = 1000.0/max(im_height, im_width)
+            img = cv2.resize(img, (0,0), fx=ss, fy=ss)
+            im_height, im_width = img.shape[:2]
+
+        scale = torch.Tensor([img.shape[1], img.shape[0], img.shape[1], img.shape[0]])
+        img -= (104, 117, 123)
+        img = img.transpose(2, 0, 1)
+        img = torch.from_numpy(img).unsqueeze(0)
+        img = img.to(self.device)
+        scale = scale.to(self.device)
+        
+        with torch.no_grad():
+            loc, conf, landms = self.net(img)  # forward pass
+
+        priorbox = PriorBox(self.cfg, image_size=(im_height, im_width))
+        priors = priorbox.forward()
+        priors = priors.to(self.device)
+        prior_data = priors.data
+        boxes = decode(loc.data.squeeze(0), prior_data, self.cfg['variance'])
+        boxes = boxes * scale / resize
+        boxes = boxes.cpu().numpy()
+        scores = conf.squeeze(0).data.cpu().numpy()[:, 1]
+        landms = decode_landm(landms.data.squeeze(0), prior_data, self.cfg['variance'])
+        scale1 = torch.Tensor([img.shape[3], img.shape[2], img.shape[3], img.shape[2],
+                               img.shape[3], img.shape[2], img.shape[3], img.shape[2],
+                               img.shape[3], img.shape[2]])
+        scale1 = scale1.to(self.device)
+        landms = landms * scale1 / resize
+        landms = landms.cpu().numpy()
+
+        # ignore low scores
+        inds = np.where(scores > confidence_threshold)[0]
+        boxes = boxes[inds]
+        landms = landms[inds]
+        scores = scores[inds]
+
+        # keep top-K before NMS
+        order = scores.argsort()[::-1][:top_k]
+        boxes = boxes[order]
+        landms = landms[order]
+        scores = scores[order]
+
+        # do NMS
+        dets = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)
+        keep = py_cpu_nms(dets, nms_threshold)
+        # keep = nms(dets, nms_threshold,force_cpu=args.cpu)
+        dets = dets[keep, :]
+        landms = landms[keep]
+
+        # keep top-K faster NMS
+        dets = dets[:keep_top_k, :]
+        landms = landms[:keep_top_k, :]
+
+        # sort faces(delete)
+        '''
+        fscores = [det[4] for det in dets]
+        sorted_idx = sorted(range(len(fscores)), key=lambda k:fscores[k], reverse=False) # sort index
+        tmp = [landms[idx] for idx in sorted_idx]
+        landms = np.asarray(tmp)
+        '''
+        
+        landms = landms.reshape((-1, 5, 2))
+        landms = landms.transpose((0, 2, 1))
+        landms = landms.reshape(-1, 10, )
+        return dets/ss, landms/ss
+
+    def detect_tensor(self, img, resize=1, confidence_threshold=0.9, nms_threshold=0.4, top_k=5000, keep_top_k=750, save_image=False):
+        im_height, im_width = img.shape[-2:]
+        ss = 1000/max(im_height, im_width)
+        img = F.interpolate(img, scale_factor=ss)
+        im_height, im_width = img.shape[-2:]
+        scale = torch.Tensor([im_width, im_height, im_width, im_height]).to(self.device)
+        img -= self.mean
+
+        loc, conf, landms = self.net(img)  # forward pass
+
+        priorbox = PriorBox(self.cfg, image_size=(im_height, im_width))
+        priors = priorbox.forward()
+        priors = priors.to(self.device)
+        prior_data = priors.data
+        boxes = decode(loc.data.squeeze(0), prior_data, self.cfg['variance'])
+        boxes = boxes * scale / resize
+        boxes = boxes.cpu().numpy()
+        scores = conf.squeeze(0).data.cpu().numpy()[:, 1]
+        landms = decode_landm(landms.data.squeeze(0), prior_data, self.cfg['variance'])
+        scale1 = torch.Tensor([img.shape[3], img.shape[2], img.shape[3], img.shape[2],
+                               img.shape[3], img.shape[2], img.shape[3], img.shape[2],
+                               img.shape[3], img.shape[2]])
+        scale1 = scale1.to(self.device)
+        landms = landms * scale1 / resize
+        landms = landms.cpu().numpy()
+
+        # ignore low scores
+        inds = np.where(scores > confidence_threshold)[0]
+        boxes = boxes[inds]
+        landms = landms[inds]
+        scores = scores[inds]
+
+        # keep top-K before NMS
+        order = scores.argsort()[::-1][:top_k]
+        boxes = boxes[order]
+        landms = landms[order]
+        scores = scores[order]
+
+        # do NMS
+        dets = np.hstack((boxes, scores[:, np.newaxis])).astype(np.float32, copy=False)
+        keep = py_cpu_nms(dets, nms_threshold)
+        # keep = nms(dets, nms_threshold,force_cpu=args.cpu)
+        dets = dets[keep, :]
+        landms = landms[keep]
+
+        # keep top-K faster NMS
+        dets = dets[:keep_top_k, :]
+        landms = landms[:keep_top_k, :]
+
+        # sort faces(delete)
+        '''
+        fscores = [det[4] for det in dets]
+        sorted_idx = sorted(range(len(fscores)), key=lambda k:fscores[k], reverse=False) # sort index
+        tmp = [landms[idx] for idx in sorted_idx]
+        landms = np.asarray(tmp)
+        '''
+        
+        landms = landms.reshape((-1, 5, 2))
+        landms = landms.transpose((0, 2, 1))
+        landms = landms.reshape(-1, 10, )
+        return dets/ss, landms/ss

video-tetalking/third_part/GPEN/face_detect/utils/box_utils.py

@@ -0,0 +1,330 @@
+import torch
+import numpy as np
+
+
+def point_form(boxes):
+    """ Convert prior_boxes to (xmin, ymin, xmax, ymax)
+    representation for comparison to point form ground truth data.
+    Args:
+        boxes: (tensor) center-size default boxes from priorbox layers.
+    Return:
+        boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
+    """
+    return torch.cat((boxes[:, :2] - boxes[:, 2:]/2,     # xmin, ymin
+                     boxes[:, :2] + boxes[:, 2:]/2), 1)  # xmax, ymax
+
+
+def center_size(boxes):
+    """ Convert prior_boxes to (cx, cy, w, h)
+    representation for comparison to center-size form ground truth data.
+    Args:
+        boxes: (tensor) point_form boxes
+    Return:
+        boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
+    """
+    return torch.cat((boxes[:, 2:] + boxes[:, :2])/2,  # cx, cy
+                     boxes[:, 2:] - boxes[:, :2], 1)  # w, h
+
+
+def intersect(box_a, box_b):
+    """ We resize both tensors to [A,B,2] without new malloc:
+    [A,2] -> [A,1,2] -> [A,B,2]
+    [B,2] -> [1,B,2] -> [A,B,2]
+    Then we compute the area of intersect between box_a and box_b.
+    Args:
+      box_a: (tensor) bounding boxes, Shape: [A,4].
+      box_b: (tensor) bounding boxes, Shape: [B,4].
+    Return:
+      (tensor) intersection area, Shape: [A,B].
+    """
+    A = box_a.size(0)
+    B = box_b.size(0)
+    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
+                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
+    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
+                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
+    inter = torch.clamp((max_xy - min_xy), min=0)
+    return inter[:, :, 0] * inter[:, :, 1]
+
+
+def jaccard(box_a, box_b):
+    """Compute the jaccard overlap of two sets of boxes.  The jaccard overlap
+    is simply the intersection over union of two boxes.  Here we operate on
+    ground truth boxes and default boxes.
+    E.g.:
+        A ∩ B / A ∪ B = A ∩ B / (area(A) + area(B) - A ∩ B)
+    Args:
+        box_a: (tensor) Ground truth bounding boxes, Shape: [num_objects,4]
+        box_b: (tensor) Prior boxes from priorbox layers, Shape: [num_priors,4]
+    Return:
+        jaccard overlap: (tensor) Shape: [box_a.size(0), box_b.size(0)]
+    """
+    inter = intersect(box_a, box_b)
+    area_a = ((box_a[:, 2]-box_a[:, 0]) *
+              (box_a[:, 3]-box_a[:, 1])).unsqueeze(1).expand_as(inter)  # [A,B]
+    area_b = ((box_b[:, 2]-box_b[:, 0]) *
+              (box_b[:, 3]-box_b[:, 1])).unsqueeze(0).expand_as(inter)  # [A,B]
+    union = area_a + area_b - inter
+    return inter / union  # [A,B]
+
+
+def matrix_iou(a, b):
+    """
+    return iou of a and b, numpy version for data augenmentation
+    """
+    lt = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+    rb = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+
+    area_i = np.prod(rb - lt, axis=2) * (lt < rb).all(axis=2)
+    area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+    area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
+    return area_i / (area_a[:, np.newaxis] + area_b - area_i)
+
+
+def matrix_iof(a, b):
+    """
+    return iof of a and b, numpy version for data augenmentation
+    """
+    lt = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+    rb = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+
+    area_i = np.prod(rb - lt, axis=2) * (lt < rb).all(axis=2)
+    area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+    return area_i / np.maximum(area_a[:, np.newaxis], 1)
+
+
+def match(threshold, truths, priors, variances, labels, landms, loc_t, conf_t, landm_t, idx):
+    """Match each prior box with the ground truth box of the highest jaccard
+    overlap, encode the bounding boxes, then return the matched indices
+    corresponding to both confidence and location preds.
+    Args:
+        threshold: (float) The overlap threshold used when mathing boxes.
+        truths: (tensor) Ground truth boxes, Shape: [num_obj, 4].
+        priors: (tensor) Prior boxes from priorbox layers, Shape: [n_priors,4].
+        variances: (tensor) Variances corresponding to each prior coord,
+            Shape: [num_priors, 4].
+        labels: (tensor) All the class labels for the image, Shape: [num_obj].
+        landms: (tensor) Ground truth landms, Shape [num_obj, 10].
+        loc_t: (tensor) Tensor to be filled w/ endcoded location targets.
+        conf_t: (tensor) Tensor to be filled w/ matched indices for conf preds.
+        landm_t: (tensor) Tensor to be filled w/ endcoded landm targets.
+        idx: (int) current batch index
+    Return:
+        The matched indices corresponding to 1)location 2)confidence 3)landm preds.
+    """
+    # jaccard index
+    overlaps = jaccard(
+        truths,
+        point_form(priors)
+    )
+    # (Bipartite Matching)
+    # [1,num_objects] best prior for each ground truth
+    best_prior_overlap, best_prior_idx = overlaps.max(1, keepdim=True)
+
+    # ignore hard gt
+    valid_gt_idx = best_prior_overlap[:, 0] >= 0.2
+    best_prior_idx_filter = best_prior_idx[valid_gt_idx, :]
+    if best_prior_idx_filter.shape[0] <= 0:
+        loc_t[idx] = 0
+        conf_t[idx] = 0
+        return
+
+    # [1,num_priors] best ground truth for each prior
+    best_truth_overlap, best_truth_idx = overlaps.max(0, keepdim=True)
+    best_truth_idx.squeeze_(0)
+    best_truth_overlap.squeeze_(0)
+    best_prior_idx.squeeze_(1)
+    best_prior_idx_filter.squeeze_(1)
+    best_prior_overlap.squeeze_(1)
+    best_truth_overlap.index_fill_(0, best_prior_idx_filter, 2)  # ensure best prior
+    # TODO refactor: index  best_prior_idx with long tensor
+    # ensure every gt matches with its prior of max overlap
+    for j in range(best_prior_idx.size(0)):     # 判别此anchor是预测哪一个boxes
+        best_truth_idx[best_prior_idx[j]] = j
+    matches = truths[best_truth_idx]            # Shape: [num_priors,4] 此处为每一个anchor对应的bbox取出来
+    conf = labels[best_truth_idx]               # Shape: [num_priors]      此处为每一个anchor对应的label取出来
+    conf[best_truth_overlap < threshold] = 0    # label as background   overlap<0.35的全部作为负样本
+    loc = encode(matches, priors, variances)
+
+    matches_landm = landms[best_truth_idx]
+    landm = encode_landm(matches_landm, priors, variances)
+    loc_t[idx] = loc    # [num_priors,4] encoded offsets to learn
+    conf_t[idx] = conf  # [num_priors] top class label for each prior
+    landm_t[idx] = landm
+
+
+def encode(matched, priors, variances):
+    """Encode the variances from the priorbox layers into the ground truth boxes
+    we have matched (based on jaccard overlap) with the prior boxes.
+    Args:
+        matched: (tensor) Coords of ground truth for each prior in point-form
+            Shape: [num_priors, 4].
+        priors: (tensor) Prior boxes in center-offset form
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        encoded boxes (tensor), Shape: [num_priors, 4]
+    """
+
+    # dist b/t match center and prior's center
+    g_cxcy = (matched[:, :2] + matched[:, 2:])/2 - priors[:, :2]
+    # encode variance
+    g_cxcy /= (variances[0] * priors[:, 2:])
+    # match wh / prior wh
+    g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
+    g_wh = torch.log(g_wh) / variances[1]
+    # return target for smooth_l1_loss
+    return torch.cat([g_cxcy, g_wh], 1)  # [num_priors,4]
+
+def encode_landm(matched, priors, variances):
+    """Encode the variances from the priorbox layers into the ground truth boxes
+    we have matched (based on jaccard overlap) with the prior boxes.
+    Args:
+        matched: (tensor) Coords of ground truth for each prior in point-form
+            Shape: [num_priors, 10].
+        priors: (tensor) Prior boxes in center-offset form
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        encoded landm (tensor), Shape: [num_priors, 10]
+    """
+
+    # dist b/t match center and prior's center
+    matched = torch.reshape(matched, (matched.size(0), 5, 2))
+    priors_cx = priors[:, 0].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_cy = priors[:, 1].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_w = priors[:, 2].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors_h = priors[:, 3].unsqueeze(1).expand(matched.size(0), 5).unsqueeze(2)
+    priors = torch.cat([priors_cx, priors_cy, priors_w, priors_h], dim=2)
+    g_cxcy = matched[:, :, :2] - priors[:, :, :2]
+    # encode variance
+    g_cxcy /= (variances[0] * priors[:, :, 2:])
+    # g_cxcy /= priors[:, :, 2:]
+    g_cxcy = g_cxcy.reshape(g_cxcy.size(0), -1)
+    # return target for smooth_l1_loss
+    return g_cxcy
+
+
+# Adapted from https://github.com/Hakuyume/chainer-ssd
+def decode(loc, priors, variances):
+    """Decode locations from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        loc (tensor): location predictions for loc layers,
+            Shape: [num_priors,4]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded bounding box predictions
+    """
+
+    boxes = torch.cat((
+        priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:],
+        priors[:, 2:] * torch.exp(loc[:, 2:] * variances[1])), 1)
+    boxes[:, :2] -= boxes[:, 2:] / 2
+    boxes[:, 2:] += boxes[:, :2]
+    return boxes
+
+def decode_landm(pre, priors, variances):
+    """Decode landm from predictions using priors to undo
+    the encoding we did for offset regression at train time.
+    Args:
+        pre (tensor): landm predictions for loc layers,
+            Shape: [num_priors,10]
+        priors (tensor): Prior boxes in center-offset form.
+            Shape: [num_priors,4].
+        variances: (list[float]) Variances of priorboxes
+    Return:
+        decoded landm predictions
+    """
+    landms = torch.cat((priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:],
+                        priors[:, :2] + pre[:, 2:4] * variances[0] * priors[:, 2:],
+                        priors[:, :2] + pre[:, 4:6] * variances[0] * priors[:, 2:],
+                        priors[:, :2] + pre[:, 6:8] * variances[0] * priors[:, 2:],
+                        priors[:, :2] + pre[:, 8:10] * variances[0] * priors[:, 2:],
+                        ), dim=1)
+    return landms
+
+
+def log_sum_exp(x):
+    """Utility function for computing log_sum_exp while determining
+    This will be used to determine unaveraged confidence loss across
+    all examples in a batch.
+    Args:
+        x (Variable(tensor)): conf_preds from conf layers
+    """
+    x_max = x.data.max()
+    return torch.log(torch.sum(torch.exp(x-x_max), 1, keepdim=True)) + x_max
+
+
+# Original author: Francisco Massa:
+# https://github.com/fmassa/object-detection.torch
+# Ported to PyTorch by Max deGroot (02/01/2017)
+def nms(boxes, scores, overlap=0.5, top_k=200):
+    """Apply non-maximum suppression at test time to avoid detecting too many
+    overlapping bounding boxes for a given object.
+    Args:
+        boxes: (tensor) The location preds for the img, Shape: [num_priors,4].
+        scores: (tensor) The class predscores for the img, Shape:[num_priors].
+        overlap: (float) The overlap thresh for suppressing unnecessary boxes.
+        top_k: (int) The Maximum number of box preds to consider.
+    Return:
+        The indices of the kept boxes with respect to num_priors.
+    """
+
+    keep = torch.Tensor(scores.size(0)).fill_(0).long()
+    if boxes.numel() == 0:
+        return keep
+    x1 = boxes[:, 0]
+    y1 = boxes[:, 1]
+    x2 = boxes[:, 2]
+    y2 = boxes[:, 3]
+    area = torch.mul(x2 - x1, y2 - y1)
+    v, idx = scores.sort(0)  # sort in ascending order
+    # I = I[v >= 0.01]
+    idx = idx[-top_k:]  # indices of the top-k largest vals
+    xx1 = boxes.new()
+    yy1 = boxes.new()
+    xx2 = boxes.new()
+    yy2 = boxes.new()
+    w = boxes.new()
+    h = boxes.new()
+
+    # keep = torch.Tensor()
+    count = 0
+    while idx.numel() > 0:
+        i = idx[-1]  # index of current largest val
+        # keep.append(i)
+        keep[count] = i
+        count += 1
+        if idx.size(0) == 1:
+            break
+        idx = idx[:-1]  # remove kept element from view
+        # load bboxes of next highest vals
+        torch.index_select(x1, 0, idx, out=xx1)
+        torch.index_select(y1, 0, idx, out=yy1)
+        torch.index_select(x2, 0, idx, out=xx2)
+        torch.index_select(y2, 0, idx, out=yy2)
+        # store element-wise max with next highest score
+        xx1 = torch.clamp(xx1, min=x1[i])
+        yy1 = torch.clamp(yy1, min=y1[i])
+        xx2 = torch.clamp(xx2, max=x2[i])
+        yy2 = torch.clamp(yy2, max=y2[i])
+        w.resize_as_(xx2)
+        h.resize_as_(yy2)
+        w = xx2 - xx1
+        h = yy2 - yy1
+        # check sizes of xx1 and xx2.. after each iteration
+        w = torch.clamp(w, min=0.0)
+        h = torch.clamp(h, min=0.0)
+        inter = w*h
+        # IoU = i / (area(a) + area(b) - i)
+        rem_areas = torch.index_select(area, 0, idx)  # load remaining areas)
+        union = (rem_areas - inter) + area[i]
+        IoU = inter/union  # store result in iou
+        # keep only elements with an IoU <= overlap
+        idx = idx[IoU.le(overlap)]
+    return keep, count
+
+