upload chkpt

chkpt, src, doc

checkpoints/BFM_Fitting/01_MorphableModel.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/37b1f0742db356a3b1568a8365a06f5b0fe0ab687ac1c3068c803666cbd4d8e2

checkpoints/BFM_Fitting/BFM09_model_info.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/db8d00544f0b0182f1b8430a3bb87662b3ff674eb33c84e6f52dbe2971adb81b

checkpoints/BFM_Fitting/BFM_exp_idx.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/1146e4e9c3bef303a497383aa7974c014fe945c7

checkpoints/BFM_Fitting/BFM_front_idx.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/b9d7b0953dd1dc5b1e28144610485409ac321f9b

checkpoints/BFM_Fitting/Exp_Pca.bin

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/e7f31380e6cbdaf2aeec698db220bac4f221946e4d551d88c092d47ec49b1726

checkpoints/BFM_Fitting/facemodel_info.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/3e516ec7297fa3248098f49ecea10579f4831c0a

checkpoints/BFM_Fitting/select_vertex_id.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/5b8b220093d93b133acc94ffed159f31a74854cd

checkpoints/BFM_Fitting/similarity_Lm3D_all.mat

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/a0e23588302bc71fc899eef53ff06df5f4df4c1d

checkpoints/BFM_Fitting/std_exp.txt

@@ -0,0 +1 @@
+../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/767b8de4ea1ca78b6f22b98ff2dee4fa345500bb

checkpoints/auido2exp_00300-model.pth

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/b7608f0e6b477e50e03ca569ac5b04a841b9217f89d502862fc78fda4e46dec4

checkpoints/auido2pose_00140-model.pth

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/4fba6701852dc57efbed25b1e4276e4ff752941860d69fc4429f08a02326ebce

checkpoints/epoch_20.pth

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/6d17a6b23457b521801baae583cb6a58f7238fe6721fc3d65d76407460e9149b

checkpoints/facevid2vid_00189-model.pth.tar

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/fbad01d46f0510276dc4521322dde6824a873a4222cd0740c85762e7067ea71d

checkpoints/hub/checkpoints/2DFAN4-cd938726ad.zip

@@ -0,0 +1 @@
+../../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/cd938726adb1f15f361263cce2db9cb820c42585fa8796ec72ce19107f369a46

checkpoints/hub/checkpoints/s3fd-619a316812.pth

@@ -0,0 +1 @@
+../../../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/619a31681264d3f7f7fc7a16a42cbbe8b23f31a256f75a366e5a1bcd59b33543

checkpoints/mapping_00229-model.pth.tar

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker-V002rc/blobs/62a1e06006cc963220f6477438518ed86e9788226c62ae382ddc42fbcefb83f1

checkpoints/shape_predictor_68_face_landmarks.dat

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/fbdc2cb80eb9aa7a758672cbfdda32ba6300efe9b6e6c7a299ff7e736b11b92f

checkpoints/wav2lip.pth

@@ -0,0 +1 @@
+../../../../root/.cache/huggingface/hub/models--vinthony--SadTalker/blobs/b78b681b68ad9fe6c6fb1debc6ff43ad05834a8af8a62ffc4167b7b34ef63c37

src/audio2exp_models/audio2exp.py

@@ -0,0 +1,41 @@
+from tqdm import tqdm
+import torch
+from torch import nn
+
+
+class Audio2Exp(nn.Module):
+    def __init__(self, netG, cfg, device, prepare_training_loss=False):
+        super(Audio2Exp, self).__init__()
+        self.cfg = cfg
+        self.device = device
+        self.netG = netG.to(device)
+
+    def test(self, batch):
+
+        mel_input = batch['indiv_mels']                         # bs T 1 80 16
+        bs = mel_input.shape[0]
+        T = mel_input.shape[1]
+
+        exp_coeff_pred = []
+
+        for i in tqdm(range(0, T, 10),'audio2exp:'): # every 10 frames
+            
+            current_mel_input = mel_input[:,i:i+10]
+
+            #ref = batch['ref'][:, :, :64].repeat((1,current_mel_input.shape[1],1))           #bs T 64
+            ref = batch['ref'][:, :, :64][:, i:i+10]
+            ratio = batch['ratio_gt'][:, i:i+10]                               #bs T
+
+            audiox = current_mel_input.view(-1, 1, 80, 16)                  # bs*T 1 80 16
+
+            curr_exp_coeff_pred  = self.netG(audiox, ref, ratio)         # bs T 64 
+
+            exp_coeff_pred += [curr_exp_coeff_pred]
+
+        # BS x T x 64
+        results_dict = {
+            'exp_coeff_pred': torch.cat(exp_coeff_pred, axis=1)
+            }
+        return results_dict
+
+

src/audio2exp_models/networks.py

@@ -0,0 +1,74 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+class Conv2d(nn.Module):
+    def __init__(self, cin, cout, kernel_size, stride, padding, residual=False, use_act = True, *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
+        self.use_act = use_act
+
+    def forward(self, x):
+        out = self.conv_block(x)
+        if self.residual:
+            out += x
+        
+        if self.use_act:
+            return self.act(out)
+        else:
+            return out
+
+class SimpleWrapperV2(nn.Module):
+    def __init__(self) -> None:
+        super().__init__()
+        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, 512, kernel_size=1, stride=1, padding=0),
+            )
+
+        #### load the pre-trained audio_encoder 
+        #self.audio_encoder = self.audio_encoder.to(device)  
+        '''
+        wav2lip_state_dict = torch.load('/apdcephfs_cq2/share_1290939/wenxuazhang/checkpoints/wav2lip.pth')['state_dict']
+        state_dict = self.audio_encoder.state_dict()
+
+        for k,v in wav2lip_state_dict.items():
+            if 'audio_encoder' in k:
+                print('init:', k)
+                state_dict[k.replace('module.audio_encoder.', '')] = v
+        self.audio_encoder.load_state_dict(state_dict)
+        '''
+
+        self.mapping1 = nn.Linear(512+64+1, 64)
+        #self.mapping2 = nn.Linear(30, 64)
+        #nn.init.constant_(self.mapping1.weight, 0.)
+        nn.init.constant_(self.mapping1.bias, 0.)
+
+    def forward(self, x, ref, ratio):
+        x = self.audio_encoder(x).view(x.size(0), -1)
+        ref_reshape = ref.reshape(x.size(0), -1)
+        ratio = ratio.reshape(x.size(0), -1)
+        
+        y = self.mapping1(torch.cat([x, ref_reshape, ratio], dim=1)) 
+        out = y.reshape(ref.shape[0], ref.shape[1], -1) #+ ref # resudial
+        return out

src/audio2pose_models/audio2pose.py

@@ -0,0 +1,94 @@
+import torch
+from torch import nn
+from src.audio2pose_models.cvae import CVAE
+from src.audio2pose_models.discriminator import PoseSequenceDiscriminator
+from src.audio2pose_models.audio_encoder import AudioEncoder
+
+class Audio2Pose(nn.Module):
+    def __init__(self, cfg, wav2lip_checkpoint, device='cuda'):
+        super().__init__()
+        self.cfg = cfg
+        self.seq_len = cfg.MODEL.CVAE.SEQ_LEN
+        self.latent_dim = cfg.MODEL.CVAE.LATENT_SIZE
+        self.device = device
+
+        self.audio_encoder = AudioEncoder(wav2lip_checkpoint, device)
+        self.audio_encoder.eval()
+        for param in self.audio_encoder.parameters():
+            param.requires_grad = False
+
+        self.netG = CVAE(cfg)
+        self.netD_motion = PoseSequenceDiscriminator(cfg)
+        
+        
+    def forward(self, x):
+
+        batch = {}
+        coeff_gt = x['gt'].cuda().squeeze(0)           #bs frame_len+1 73
+        batch['pose_motion_gt'] = coeff_gt[:, 1:, 64:70] - coeff_gt[:, :1, 64:70] #bs frame_len 6
+        batch['ref'] = coeff_gt[:, 0, 64:70]  #bs  6
+        batch['class'] = x['class'].squeeze(0).cuda() # bs
+        indiv_mels= x['indiv_mels'].cuda().squeeze(0) # bs seq_len+1 80 16
+
+        # forward
+        audio_emb_list = []
+        audio_emb = self.audio_encoder(indiv_mels[:, 1:, :, :].unsqueeze(2)) #bs seq_len 512
+        batch['audio_emb'] = audio_emb
+        batch = self.netG(batch)
+
+        pose_motion_pred = batch['pose_motion_pred']           # bs frame_len 6
+        pose_gt = coeff_gt[:, 1:, 64:70].clone()               # bs frame_len 6
+        pose_pred = coeff_gt[:, :1, 64:70] + pose_motion_pred  # bs frame_len 6
+
+        batch['pose_pred'] = pose_pred
+        batch['pose_gt'] = pose_gt
+
+        return batch
+
+    def test(self, x):
+
+        batch = {}
+        ref = x['ref']                            #bs 1 70
+        batch['ref'] = x['ref'][:,0,-6:]  
+        batch['class'] = x['class']  
+        bs = ref.shape[0]
+        
+        indiv_mels= x['indiv_mels']               # bs T 1 80 16
+        indiv_mels_use = indiv_mels[:, 1:]        # we regard the ref as the first frame
+        num_frames = x['num_frames']
+        num_frames = int(num_frames) - 1
+
+        #  
+        div = num_frames//self.seq_len
+        re = num_frames%self.seq_len
+        audio_emb_list = []
+        pose_motion_pred_list = [torch.zeros(batch['ref'].unsqueeze(1).shape, dtype=batch['ref'].dtype, 
+                                                device=batch['ref'].device)]
+
+        for i in range(div):
+            z = torch.randn(bs, self.latent_dim).to(ref.device)
+            batch['z'] = z
+            audio_emb = self.audio_encoder(indiv_mels_use[:, i*self.seq_len:(i+1)*self.seq_len,:,:,:]) #bs seq_len 512
+            batch['audio_emb'] = audio_emb
+            batch = self.netG.test(batch)
+            pose_motion_pred_list.append(batch['pose_motion_pred'])  #list of bs seq_len 6
+        
+        if re != 0:
+            z = torch.randn(bs, self.latent_dim).to(ref.device)
+            batch['z'] = z
+            audio_emb = self.audio_encoder(indiv_mels_use[:, -1*self.seq_len:,:,:,:]) #bs seq_len  512
+            if audio_emb.shape[1] != self.seq_len:
+                pad_dim = self.seq_len-audio_emb.shape[1]
+                pad_audio_emb = audio_emb[:, :1].repeat(1, pad_dim, 1) 
+                audio_emb = torch.cat([pad_audio_emb, audio_emb], 1) 
+            batch['audio_emb'] = audio_emb
+            batch = self.netG.test(batch)
+            pose_motion_pred_list.append(batch['pose_motion_pred'][:,-1*re:,:])   
+        
+        pose_motion_pred = torch.cat(pose_motion_pred_list, dim = 1)
+        batch['pose_motion_pred'] = pose_motion_pred
+
+        pose_pred = ref[:, :1, -6:] + pose_motion_pred  # bs T 6
+
+        batch['pose_pred'] = pose_pred
+        return batch

src/audio2pose_models/audio_encoder.py

@@ -0,0 +1,64 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+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 AudioEncoder(nn.Module):
+    def __init__(self, wav2lip_checkpoint, device):
+        super(AudioEncoder, self).__init__()
+
+        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, 512, kernel_size=1, stride=1, padding=0),)
+
+        #### load the pre-trained audio_encoder, we do not need to load wav2lip model here.
+        # wav2lip_state_dict = torch.load(wav2lip_checkpoint, map_location=torch.device(device))['state_dict']
+        # state_dict = self.audio_encoder.state_dict()
+
+        # for k,v in wav2lip_state_dict.items():
+        #     if 'audio_encoder' in k:
+        #         state_dict[k.replace('module.audio_encoder.', '')] = v
+        # self.audio_encoder.load_state_dict(state_dict)
+
+
+    def forward(self, audio_sequences):
+        # audio_sequences = (B, T, 1, 80, 16)
+        B = audio_sequences.size(0)
+
+        audio_sequences = torch.cat([audio_sequences[:, i] for i in range(audio_sequences.size(1))], dim=0)
+
+        audio_embedding = self.audio_encoder(audio_sequences) # B, 512, 1, 1
+        dim = audio_embedding.shape[1]
+        audio_embedding = audio_embedding.reshape((B, -1, dim, 1, 1))
+
+        return audio_embedding.squeeze(-1).squeeze(-1) #B seq_len+1 512 

src/audio2pose_models/cvae.py

@@ -0,0 +1,149 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+from src.audio2pose_models.res_unet import ResUnet
+
+def class2onehot(idx, class_num):
+
+    assert torch.max(idx).item() < class_num
+    onehot = torch.zeros(idx.size(0), class_num).to(idx.device)
+    onehot.scatter_(1, idx, 1)
+    return onehot
+
+class CVAE(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        encoder_layer_sizes = cfg.MODEL.CVAE.ENCODER_LAYER_SIZES
+        decoder_layer_sizes = cfg.MODEL.CVAE.DECODER_LAYER_SIZES
+        latent_size = cfg.MODEL.CVAE.LATENT_SIZE
+        num_classes = cfg.DATASET.NUM_CLASSES
+        audio_emb_in_size = cfg.MODEL.CVAE.AUDIO_EMB_IN_SIZE
+        audio_emb_out_size = cfg.MODEL.CVAE.AUDIO_EMB_OUT_SIZE
+        seq_len = cfg.MODEL.CVAE.SEQ_LEN
+
+        self.latent_size = latent_size
+
+        self.encoder = ENCODER(encoder_layer_sizes, latent_size, num_classes,
+                                audio_emb_in_size, audio_emb_out_size, seq_len)
+        self.decoder = DECODER(decoder_layer_sizes, latent_size, num_classes,
+                                audio_emb_in_size, audio_emb_out_size, seq_len)
+    def reparameterize(self, mu, logvar):
+        std = torch.exp(0.5 * logvar)
+        eps = torch.randn_like(std)
+        return mu + eps * std
+
+    def forward(self, batch):
+        batch = self.encoder(batch)
+        mu = batch['mu']
+        logvar = batch['logvar']
+        z = self.reparameterize(mu, logvar)
+        batch['z'] = z
+        return self.decoder(batch)
+
+    def test(self, batch):
+        '''
+        class_id = batch['class']
+        z = torch.randn([class_id.size(0), self.latent_size]).to(class_id.device)
+        batch['z'] = z
+        '''
+        return self.decoder(batch)
+
+class ENCODER(nn.Module):
+    def __init__(self, layer_sizes, latent_size, num_classes, 
+                audio_emb_in_size, audio_emb_out_size, seq_len):
+        super().__init__()
+
+        self.resunet = ResUnet()
+        self.num_classes = num_classes
+        self.seq_len = seq_len
+
+        self.MLP = nn.Sequential()
+        layer_sizes[0] += latent_size + seq_len*audio_emb_out_size + 6
+        for i, (in_size, out_size) in enumerate(zip(layer_sizes[:-1], layer_sizes[1:])):
+            self.MLP.add_module(
+                name="L{:d}".format(i), module=nn.Linear(in_size, out_size))
+            self.MLP.add_module(name="A{:d}".format(i), module=nn.ReLU())
+
+        self.linear_means = nn.Linear(layer_sizes[-1], latent_size)
+        self.linear_logvar = nn.Linear(layer_sizes[-1], latent_size)
+        self.linear_audio = nn.Linear(audio_emb_in_size, audio_emb_out_size)
+
+        self.classbias = nn.Parameter(torch.randn(self.num_classes, latent_size))
+
+    def forward(self, batch):
+        class_id = batch['class']
+        pose_motion_gt = batch['pose_motion_gt']                             #bs seq_len 6
+        ref = batch['ref']                             #bs 6
+        bs = pose_motion_gt.shape[0]
+        audio_in = batch['audio_emb']                          # bs seq_len audio_emb_in_size
+
+        #pose encode
+        pose_emb = self.resunet(pose_motion_gt.unsqueeze(1))          #bs 1 seq_len 6 
+        pose_emb = pose_emb.reshape(bs, -1)                    #bs seq_len*6
+
+        #audio mapping
+        print(audio_in.shape)
+        audio_out = self.linear_audio(audio_in)                # bs seq_len audio_emb_out_size
+        audio_out = audio_out.reshape(bs, -1)
+
+        class_bias = self.classbias[class_id]                  #bs latent_size
+        x_in = torch.cat([ref, pose_emb, audio_out, class_bias], dim=-1) #bs seq_len*(audio_emb_out_size+6)+latent_size
+        x_out = self.MLP(x_in)
+
+        mu = self.linear_means(x_out)
+        logvar = self.linear_means(x_out)                      #bs latent_size 
+
+        batch.update({'mu':mu, 'logvar':logvar})
+        return batch
+
+class DECODER(nn.Module):
+    def __init__(self, layer_sizes, latent_size, num_classes, 
+                audio_emb_in_size, audio_emb_out_size, seq_len):
+        super().__init__()
+
+        self.resunet = ResUnet()
+        self.num_classes = num_classes
+        self.seq_len = seq_len
+
+        self.MLP = nn.Sequential()
+        input_size = latent_size + seq_len*audio_emb_out_size + 6
+        for i, (in_size, out_size) in enumerate(zip([input_size]+layer_sizes[:-1], layer_sizes)):
+            self.MLP.add_module(
+                name="L{:d}".format(i), module=nn.Linear(in_size, out_size))
+            if i+1 < len(layer_sizes):
+                self.MLP.add_module(name="A{:d}".format(i), module=nn.ReLU())
+            else:
+                self.MLP.add_module(name="sigmoid", module=nn.Sigmoid())
+        
+        self.pose_linear = nn.Linear(6, 6)
+        self.linear_audio = nn.Linear(audio_emb_in_size, audio_emb_out_size)
+
+        self.classbias = nn.Parameter(torch.randn(self.num_classes, latent_size))
+
+    def forward(self, batch):
+
+        z = batch['z']                                          #bs latent_size
+        bs = z.shape[0]
+        class_id = batch['class']
+        ref = batch['ref']                             #bs 6
+        audio_in = batch['audio_emb']                           # bs seq_len audio_emb_in_size
+        #print('audio_in: ', audio_in[:, :, :10])
+
+        audio_out = self.linear_audio(audio_in)                 # bs seq_len audio_emb_out_size
+        #print('audio_out: ', audio_out[:, :, :10])
+        audio_out = audio_out.reshape([bs, -1])                 # bs seq_len*audio_emb_out_size
+        class_bias = self.classbias[class_id]                   #bs latent_size
+
+        z = z + class_bias
+        x_in = torch.cat([ref, z, audio_out], dim=-1)
+        x_out = self.MLP(x_in)                                  # bs layer_sizes[-1]
+        x_out = x_out.reshape((bs, self.seq_len, -1))
+
+        #print('x_out: ', x_out)
+
+        pose_emb = self.resunet(x_out.unsqueeze(1))             #bs 1 seq_len 6
+
+        pose_motion_pred = self.pose_linear(pose_emb.squeeze(1))       #bs seq_len 6
+
+        batch.update({'pose_motion_pred':pose_motion_pred})
+        return batch

src/audio2pose_models/discriminator.py

@@ -0,0 +1,76 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+class ConvNormRelu(nn.Module):
+    def __init__(self, conv_type='1d', in_channels=3, out_channels=64, downsample=False,
+                 kernel_size=None, stride=None, padding=None, norm='BN', leaky=False):
+        super().__init__()
+        if kernel_size is None:
+            if downsample:
+                kernel_size, stride, padding = 4, 2, 1
+            else:
+                kernel_size, stride, padding = 3, 1, 1
+
+        if conv_type == '2d':
+            self.conv = nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride,
+                padding,
+                bias=False,
+            )
+            if norm == 'BN':
+                self.norm = nn.BatchNorm2d(out_channels)
+            elif norm == 'IN':
+                self.norm = nn.InstanceNorm2d(out_channels)
+            else:
+                raise NotImplementedError
+        elif conv_type == '1d':
+            self.conv = nn.Conv1d(
+                in_channels,
+                out_channels,
+                kernel_size,
+                stride,
+                padding,
+                bias=False,
+            )
+            if norm == 'BN':
+                self.norm = nn.BatchNorm1d(out_channels)
+            elif norm == 'IN':
+                self.norm = nn.InstanceNorm1d(out_channels)
+            else:
+                raise NotImplementedError
+        nn.init.kaiming_normal_(self.conv.weight)
+
+        self.act = nn.LeakyReLU(negative_slope=0.2, inplace=False) if leaky else nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.conv(x)
+        if isinstance(self.norm, nn.InstanceNorm1d):
+            x = self.norm(x.permute((0, 2, 1))).permute((0, 2, 1))  # normalize on [C]
+        else:
+            x = self.norm(x)
+        x = self.act(x)
+        return x
+
+
+class PoseSequenceDiscriminator(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        self.cfg = cfg
+        leaky = self.cfg.MODEL.DISCRIMINATOR.LEAKY_RELU
+
+        self.seq = nn.Sequential(
+            ConvNormRelu('1d', cfg.MODEL.DISCRIMINATOR.INPUT_CHANNELS, 256, downsample=True, leaky=leaky),  # B, 256, 64
+            ConvNormRelu('1d', 256, 512, downsample=True, leaky=leaky),  # B, 512, 32
+            ConvNormRelu('1d', 512, 1024, kernel_size=3, stride=1, padding=1, leaky=leaky),  # B, 1024, 16
+            nn.Conv1d(1024, 1, kernel_size=3, stride=1, padding=1, bias=True)  # B, 1, 16
+        )
+
+    def forward(self, x):
+        x = x.reshape(x.size(0), x.size(1), -1).transpose(1, 2)
+        x = self.seq(x)
+        x = x.squeeze(1)
+        return x

src/audio2pose_models/networks.py

@@ -0,0 +1,140 @@
+import torch.nn as nn
+import torch
+
+
+class ResidualConv(nn.Module):
+    def __init__(self, input_dim, output_dim, stride, padding):
+        super(ResidualConv, self).__init__()
+
+        self.conv_block = nn.Sequential(
+            nn.BatchNorm2d(input_dim),
+            nn.ReLU(),
+            nn.Conv2d(
+                input_dim, output_dim, kernel_size=3, stride=stride, padding=padding
+            ),
+            nn.BatchNorm2d(output_dim),
+            nn.ReLU(),
+            nn.Conv2d(output_dim, output_dim, kernel_size=3, padding=1),
+        )
+        self.conv_skip = nn.Sequential(
+            nn.Conv2d(input_dim, output_dim, kernel_size=3, stride=stride, padding=1),
+            nn.BatchNorm2d(output_dim),
+        )
+
+    def forward(self, x):
+
+        return self.conv_block(x) + self.conv_skip(x)
+
+
+class Upsample(nn.Module):
+    def __init__(self, input_dim, output_dim, kernel, stride):
+        super(Upsample, self).__init__()
+
+        self.upsample = nn.ConvTranspose2d(
+            input_dim, output_dim, kernel_size=kernel, stride=stride
+        )
+
+    def forward(self, x):
+        return self.upsample(x)
+
+
+class Squeeze_Excite_Block(nn.Module):
+    def __init__(self, channel, reduction=16):
+        super(Squeeze_Excite_Block, 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)
+        return x * y.expand_as(x)
+
+
+class ASPP(nn.Module):
+    def __init__(self, in_dims, out_dims, rate=[6, 12, 18]):
+        super(ASPP, self).__init__()
+
+        self.aspp_block1 = nn.Sequential(
+            nn.Conv2d(
+                in_dims, out_dims, 3, stride=1, padding=rate[0], dilation=rate[0]
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm2d(out_dims),
+        )
+        self.aspp_block2 = nn.Sequential(
+            nn.Conv2d(
+                in_dims, out_dims, 3, stride=1, padding=rate[1], dilation=rate[1]
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm2d(out_dims),
+        )
+        self.aspp_block3 = nn.Sequential(
+            nn.Conv2d(
+                in_dims, out_dims, 3, stride=1, padding=rate[2], dilation=rate[2]
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm2d(out_dims),
+        )
+
+        self.output = nn.Conv2d(len(rate) * out_dims, out_dims, 1)
+        self._init_weights()
+
+    def forward(self, x):
+        x1 = self.aspp_block1(x)
+        x2 = self.aspp_block2(x)
+        x3 = self.aspp_block3(x)
+        out = torch.cat([x1, x2, x3], dim=1)
+        return self.output(out)
+
+    def _init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+
+class Upsample_(nn.Module):
+    def __init__(self, scale=2):
+        super(Upsample_, self).__init__()
+
+        self.upsample = nn.Upsample(mode="bilinear", scale_factor=scale)
+
+    def forward(self, x):
+        return self.upsample(x)
+
+
+class AttentionBlock(nn.Module):
+    def __init__(self, input_encoder, input_decoder, output_dim):
+        super(AttentionBlock, self).__init__()
+
+        self.conv_encoder = nn.Sequential(
+            nn.BatchNorm2d(input_encoder),
+            nn.ReLU(),
+            nn.Conv2d(input_encoder, output_dim, 3, padding=1),
+            nn.MaxPool2d(2, 2),
+        )
+
+        self.conv_decoder = nn.Sequential(
+            nn.BatchNorm2d(input_decoder),
+            nn.ReLU(),
+            nn.Conv2d(input_decoder, output_dim, 3, padding=1),
+        )
+
+        self.conv_attn = nn.Sequential(
+            nn.BatchNorm2d(output_dim),
+            nn.ReLU(),
+            nn.Conv2d(output_dim, 1, 1),
+        )
+
+    def forward(self, x1, x2):
+        out = self.conv_encoder(x1) + self.conv_decoder(x2)
+        out = self.conv_attn(out)
+        return out * x2

src/audio2pose_models/res_unet.py

@@ -0,0 +1,65 @@
+import torch
+import torch.nn as nn
+from src.audio2pose_models.networks import ResidualConv, Upsample
+
+
+class ResUnet(nn.Module):
+    def __init__(self, channel=1, filters=[32, 64, 128, 256]):
+        super(ResUnet, self).__init__()
+
+        self.input_layer = nn.Sequential(
+            nn.Conv2d(channel, filters[0], kernel_size=3, padding=1),
+            nn.BatchNorm2d(filters[0]),
+            nn.ReLU(),
+            nn.Conv2d(filters[0], filters[0], kernel_size=3, padding=1),
+        )
+        self.input_skip = nn.Sequential(
+            nn.Conv2d(channel, filters[0], kernel_size=3, padding=1)
+        )
+
+        self.residual_conv_1 = ResidualConv(filters[0], filters[1], stride=(2,1), padding=1)
+        self.residual_conv_2 = ResidualConv(filters[1], filters[2], stride=(2,1), padding=1)
+
+        self.bridge = ResidualConv(filters[2], filters[3], stride=(2,1), padding=1)
+
+        self.upsample_1 = Upsample(filters[3], filters[3], kernel=(2,1), stride=(2,1))
+        self.up_residual_conv1 = ResidualConv(filters[3] + filters[2], filters[2], stride=1, padding=1)
+
+        self.upsample_2 = Upsample(filters[2], filters[2], kernel=(2,1), stride=(2,1))
+        self.up_residual_conv2 = ResidualConv(filters[2] + filters[1], filters[1], stride=1, padding=1)
+
+        self.upsample_3 = Upsample(filters[1], filters[1], kernel=(2,1), stride=(2,1))
+        self.up_residual_conv3 = ResidualConv(filters[1] + filters[0], filters[0], stride=1, padding=1)
+
+        self.output_layer = nn.Sequential(
+            nn.Conv2d(filters[0], 1, 1, 1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        # Encode
+        x1 = self.input_layer(x) + self.input_skip(x)
+        x2 = self.residual_conv_1(x1)
+        x3 = self.residual_conv_2(x2)
+        # Bridge
+        x4 = self.bridge(x3)
+
+        # Decode
+        x4 = self.upsample_1(x4)
+        x5 = torch.cat([x4, x3], dim=1)
+
+        x6 = self.up_residual_conv1(x5)
+
+        x6 = self.upsample_2(x6)
+        x7 = torch.cat([x6, x2], dim=1)
+
+        x8 = self.up_residual_conv2(x7)
+
+        x8 = self.upsample_3(x8)
+        x9 = torch.cat([x8, x1], dim=1)
+
+        x10 = self.up_residual_conv3(x9)
+
+        output = self.output_layer(x10)
+
+        return output

src/config/auido2exp.yaml

@@ -0,0 +1,58 @@
+DATASET:
+  TRAIN_FILE_LIST: /apdcephfs_cq2/share_1290939/wenxuazhang/code/file_list/train.txt
+  EVAL_FILE_LIST: /apdcephfs_cq2/share_1290939/wenxuazhang/code/file_list/val.txt
+  TRAIN_BATCH_SIZE: 32
+  EVAL_BATCH_SIZE: 32
+  EXP: True
+  EXP_DIM: 64
+  FRAME_LEN: 32
+  COEFF_LEN: 73
+  NUM_CLASSES: 46
+  AUDIO_ROOT_PATH: /apdcephfs_cq2/share_1290939/wenxuazhang/voxceleb1/wav
+  COEFF_ROOT_PATH: /apdcephfs_cq2/share_1290939/wenxuazhang/voxceleb1/wav2lip_3dmm
+  LMDB_PATH: /apdcephfs_cq2/share_1290939/shadowcun/datasets/VoxCeleb/v1/imdb
+  DEBUG: True
+  NUM_REPEATS: 2
+  T: 40
+  
+
+MODEL:
+  FRAMEWORK: V2
+  AUDIOENCODER:
+    LEAKY_RELU: True
+    NORM: 'IN'
+  DISCRIMINATOR:
+    LEAKY_RELU: False
+    INPUT_CHANNELS: 6
+  CVAE:
+    AUDIO_EMB_IN_SIZE: 512
+    AUDIO_EMB_OUT_SIZE: 128
+    SEQ_LEN: 32
+    LATENT_SIZE: 256
+    ENCODER_LAYER_SIZES: [192, 1024]
+    DECODER_LAYER_SIZES: [1024, 192]
+    
+
+TRAIN:
+  MAX_EPOCH: 300
+  GENERATOR:
+    LR: 2.0e-5
+  DISCRIMINATOR:
+    LR: 1.0e-5
+  LOSS:
+    W_FEAT: 0
+    W_COEFF_EXP: 2
+    W_LM: 1.0e-2
+    W_LM_MOUTH: 0
+    W_REG: 0
+    W_SYNC: 0
+    W_COLOR: 0
+    W_EXPRESSION: 0
+    W_LIPREADING: 0.01
+    W_LIPREADING_VV: 0
+    W_EYE_BLINK: 4
+
+TAG:
+  NAME:  small_dataset
+
+

src/config/auido2pose.yaml

@@ -0,0 +1,49 @@
+DATASET:
+  TRAIN_FILE_LIST: /apdcephfs_cq2/share_1290939/wenxuazhang/code/audio2pose_unet_noAudio/dataset/train_33.txt
+  EVAL_FILE_LIST: /apdcephfs_cq2/share_1290939/wenxuazhang/code/audio2pose_unet_noAudio/dataset/val.txt
+  TRAIN_BATCH_SIZE: 64
+  EVAL_BATCH_SIZE: 1
+  EXP: True
+  EXP_DIM: 64
+  FRAME_LEN: 32
+  COEFF_LEN: 73
+  NUM_CLASSES: 46
+  AUDIO_ROOT_PATH: /apdcephfs_cq2/share_1290939/wenxuazhang/voxceleb1/wav
+  COEFF_ROOT_PATH: /apdcephfs_cq2/share_1290939/shadowcun/datasets/VoxCeleb/v1/imdb
+  DEBUG: True
+  
+
+MODEL:
+  AUDIOENCODER:
+    LEAKY_RELU: True
+    NORM: 'IN'
+  DISCRIMINATOR:
+    LEAKY_RELU: False
+    INPUT_CHANNELS: 6
+  CVAE:
+    AUDIO_EMB_IN_SIZE: 512
+    AUDIO_EMB_OUT_SIZE: 6
+    SEQ_LEN: 32
+    LATENT_SIZE: 64
+    ENCODER_LAYER_SIZES: [192, 128]
+    DECODER_LAYER_SIZES: [128, 192]
+    
+
+TRAIN:
+  MAX_EPOCH: 150
+  GENERATOR:
+    LR: 1.0e-4
+  DISCRIMINATOR:
+    LR: 1.0e-4
+  LOSS:
+    LAMBDA_REG: 1
+    LAMBDA_LANDMARKS: 0
+    LAMBDA_VERTICES: 0
+    LAMBDA_GAN_MOTION: 0.7
+    LAMBDA_GAN_COEFF: 0
+    LAMBDA_KL: 1
+
+TAG:
+  NAME: cvae_UNET_useAudio_usewav2lipAudioEncoder
+
+

src/config/facerender.yaml

@@ -0,0 +1,45 @@
+model_params:
+  common_params:
+    num_kp: 15 
+    image_channel: 3                    
+    feature_channel: 32
+    estimate_jacobian: False   # True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32            
+     max_features: 1024
+     scale_factor: 0.25         # 0.25
+     num_blocks: 5
+     reshape_channel: 16384  # 16384 = 1024 * 16
+     reshape_depth: 16
+  he_estimator_params:
+     block_expansion: 64            
+     max_features: 2048
+     num_bins: 66
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    reshape_channel: 32
+    reshape_depth: 16         # 512 = 32 * 16
+    num_resblocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 32
+      max_features: 1024
+      num_blocks: 5
+      reshape_depth: 16
+      compress: 4
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32                 
+    max_features: 512
+    num_blocks: 4
+    sn: True
+  mapping_params:
+      coeff_nc: 70
+      descriptor_nc: 1024
+      layer: 3
+      num_kp: 15
+      num_bins: 66
+

src/config/facerender_pirender.yaml

@@ -0,0 +1,83 @@
+# How often do you want to log the training stats.
+# network_list: 
+#     gen: gen_optimizer
+#     dis: dis_optimizer
+
+distributed: False
+image_to_tensorboard: True
+snapshot_save_iter: 40000
+snapshot_save_epoch: 20
+snapshot_save_start_iter: 20000
+snapshot_save_start_epoch: 10
+image_save_iter: 1000
+max_epoch: 200
+logging_iter: 100
+results_dir: ./eval_results
+
+gen_optimizer:
+    type: adam
+    lr: 0.0001
+    adam_beta1: 0.5
+    adam_beta2: 0.999
+    lr_policy:
+        iteration_mode: True
+        type: step
+        step_size: 300000
+        gamma: 0.2
+
+trainer:
+    type: trainers.face_trainer::FaceTrainer
+    pretrain_warp_iteration: 200000
+    loss_weight:
+      weight_perceptual_warp: 2.5
+      weight_perceptual_final: 4
+    vgg_param_warp:
+      network: vgg19
+      layers: ['relu_1_1', 'relu_2_1', 'relu_3_1', 'relu_4_1', 'relu_5_1']
+      use_style_loss: False
+      num_scales: 4
+    vgg_param_final:
+      network: vgg19
+      layers: ['relu_1_1', 'relu_2_1', 'relu_3_1', 'relu_4_1', 'relu_5_1']
+      use_style_loss: True
+      num_scales: 4      
+      style_to_perceptual: 250
+    init:
+      type: 'normal'
+      gain: 0.02
+gen:
+    type: generators.face_model::FaceGenerator
+    param:
+      mapping_net:
+        coeff_nc: 73
+        descriptor_nc: 256
+        layer: 3
+      warpping_net:
+        encoder_layer: 5
+        decoder_layer: 3
+        base_nc: 32
+      editing_net:
+        layer: 3
+        num_res_blocks: 2
+        base_nc: 64
+      common:
+        image_nc: 3
+        descriptor_nc: 256
+        max_nc: 256
+        use_spect: False
+                
+
+# Data options.
+data:
+    type: data.vox_dataset::VoxDataset
+    path: ./dataset/vox_lmdb
+    resolution: 256
+    semantic_radius: 13
+    train:
+      batch_size: 5
+      distributed: True
+    val:
+      batch_size: 8
+      distributed: True
+
+

src/config/facerender_still.yaml

@@ -0,0 +1,45 @@
+model_params:
+  common_params:
+    num_kp: 15 
+    image_channel: 3                    
+    feature_channel: 32
+    estimate_jacobian: False   # True
+  kp_detector_params:
+     temperature: 0.1
+     block_expansion: 32            
+     max_features: 1024
+     scale_factor: 0.25         # 0.25
+     num_blocks: 5
+     reshape_channel: 16384  # 16384 = 1024 * 16
+     reshape_depth: 16
+  he_estimator_params:
+     block_expansion: 64            
+     max_features: 2048
+     num_bins: 66
+  generator_params:
+    block_expansion: 64
+    max_features: 512
+    num_down_blocks: 2
+    reshape_channel: 32
+    reshape_depth: 16         # 512 = 32 * 16
+    num_resblocks: 6
+    estimate_occlusion_map: True
+    dense_motion_params:
+      block_expansion: 32
+      max_features: 1024
+      num_blocks: 5
+      reshape_depth: 16
+      compress: 4
+  discriminator_params:
+    scales: [1]
+    block_expansion: 32                 
+    max_features: 512
+    num_blocks: 4
+    sn: True
+  mapping_params:
+      coeff_nc: 73
+      descriptor_nc: 1024
+      layer: 3
+      num_kp: 15
+      num_bins: 66
+

src/face3d/data/__init__.py

@@ -0,0 +1,116 @@
+"""This package includes all the modules related to data loading and preprocessing
+
+ To add a custom dataset class called 'dummy', you need to add a file called 'dummy_dataset.py' and define a subclass 'DummyDataset' inherited from BaseDataset.
+ You need to implement four functions:
+    -- <__init__>:                      initialize the class, first call BaseDataset.__init__(self, opt).
+    -- <__len__>:                       return the size of dataset.
+    -- <__getitem__>:                   get a data point from data loader.
+    -- <modify_commandline_options>:    (optionally) add dataset-specific options and set default options.
+
+Now you can use the dataset class by specifying flag '--dataset_mode dummy'.
+See our template dataset class 'template_dataset.py' for more details.
+"""
+import numpy as np
+import importlib
+import torch.utils.data
+from face3d.data.base_dataset import BaseDataset
+
+
+def find_dataset_using_name(dataset_name):
+    """Import the module "data/[dataset_name]_dataset.py".
+
+    In the file, the class called DatasetNameDataset() will
+    be instantiated. It has to be a subclass of BaseDataset,
+    and it is case-insensitive.
+    """
+    dataset_filename = "data." + dataset_name + "_dataset"
+    datasetlib = importlib.import_module(dataset_filename)
+
+    dataset = None
+    target_dataset_name = dataset_name.replace('_', '') + 'dataset'
+    for name, cls in datasetlib.__dict__.items():
+        if name.lower() == target_dataset_name.lower() \
+           and issubclass(cls, BaseDataset):
+            dataset = cls
+
+    if dataset is None:
+        raise NotImplementedError("In %s.py, there should be a subclass of BaseDataset with class name that matches %s in lowercase." % (dataset_filename, target_dataset_name))
+
+    return dataset
+
+
+def get_option_setter(dataset_name):
+    """Return the static method <modify_commandline_options> of the dataset class."""
+    dataset_class = find_dataset_using_name(dataset_name)
+    return dataset_class.modify_commandline_options
+
+
+def create_dataset(opt, rank=0):
+    """Create a dataset given the option.
+
+    This function wraps the class CustomDatasetDataLoader.
+        This is the main interface between this package and 'train.py'/'test.py'
+
+    Example:
+        >>> from data import create_dataset
+        >>> dataset = create_dataset(opt)
+    """
+    data_loader = CustomDatasetDataLoader(opt, rank=rank)
+    dataset = data_loader.load_data()
+    return dataset
+
+class CustomDatasetDataLoader():
+    """Wrapper class of Dataset class that performs multi-threaded data loading"""
+
+    def __init__(self, opt, rank=0):
+        """Initialize this class
+
+        Step 1: create a dataset instance given the name [dataset_mode]
+        Step 2: create a multi-threaded data loader.
+        """
+        self.opt = opt
+        dataset_class = find_dataset_using_name(opt.dataset_mode)
+        self.dataset = dataset_class(opt)
+        self.sampler = None
+        print("rank %d %s dataset [%s] was created" % (rank, self.dataset.name, type(self.dataset).__name__))
+        if opt.use_ddp and opt.isTrain:
+            world_size = opt.world_size
+            self.sampler = torch.utils.data.distributed.DistributedSampler(
+                    self.dataset,
+                    num_replicas=world_size,
+                    rank=rank,
+                    shuffle=not opt.serial_batches
+                )
+            self.dataloader = torch.utils.data.DataLoader(
+                        self.dataset,
+                        sampler=self.sampler,
+                        num_workers=int(opt.num_threads / world_size), 
+                        batch_size=int(opt.batch_size / world_size), 
+                        drop_last=True)
+        else:
+            self.dataloader = torch.utils.data.DataLoader(
+                self.dataset,
+                batch_size=opt.batch_size,
+                shuffle=(not opt.serial_batches) and opt.isTrain,
+                num_workers=int(opt.num_threads),
+                drop_last=True
+            )
+
+    def set_epoch(self, epoch):
+        self.dataset.current_epoch = epoch
+        if self.sampler is not None:
+            self.sampler.set_epoch(epoch)
+
+    def load_data(self):
+        return self
+
+    def __len__(self):
+        """Return the number of data in the dataset"""
+        return min(len(self.dataset), self.opt.max_dataset_size)
+
+    def __iter__(self):
+        """Return a batch of data"""
+        for i, data in enumerate(self.dataloader):
+            if i * self.opt.batch_size >= self.opt.max_dataset_size:
+                break
+            yield data

src/face3d/data/base_dataset.py

@@ -0,0 +1,125 @@
+"""This module implements an abstract base class (ABC) 'BaseDataset' for datasets.
+
+It also includes common transformation functions (e.g., get_transform, __scale_width), which can be later used in subclasses.
+"""
+import random
+import numpy as np
+import torch.utils.data as data
+from PIL import Image
+import torchvision.transforms as transforms
+from abc import ABC, abstractmethod
+
+
+class BaseDataset(data.Dataset, ABC):
+    """This class is an abstract base class (ABC) for datasets.
+
+    To create a subclass, you need to implement the following four functions:
+    -- <__init__>:                      initialize the class, first call BaseDataset.__init__(self, opt).
+    -- <__len__>:                       return the size of dataset.
+    -- <__getitem__>:                   get a data point.
+    -- <modify_commandline_options>:    (optionally) add dataset-specific options and set default options.
+    """
+
+    def __init__(self, opt):
+        """Initialize the class; save the options in the class
+
+        Parameters:
+            opt (Option class)-- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        self.opt = opt
+        # self.root = opt.dataroot
+        self.current_epoch = 0
+
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        """Add new dataset-specific options, and rewrite default values for existing options.
+
+        Parameters:
+            parser          -- original option parser
+            is_train (bool) -- whether training phase or test phase. You can use this flag to add training-specific or test-specific options.
+
+        Returns:
+            the modified parser.
+        """
+        return parser
+
+    @abstractmethod
+    def __len__(self):
+        """Return the total number of images in the dataset."""
+        return 0
+
+    @abstractmethod
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index - - a random integer for data indexing
+
+        Returns:
+            a dictionary of data with their names. It ususally contains the data itself and its metadata information.
+        """
+        pass
+
+
+def get_transform(grayscale=False):
+    transform_list = []
+    if grayscale:
+        transform_list.append(transforms.Grayscale(1))
+    transform_list += [transforms.ToTensor()]
+    return transforms.Compose(transform_list)
+
+def get_affine_mat(opt, size):
+    shift_x, shift_y, scale, rot_angle, flip = 0., 0., 1., 0., False
+    w, h = size
+
+    if 'shift' in opt.preprocess:
+        shift_pixs = int(opt.shift_pixs)
+        shift_x = random.randint(-shift_pixs, shift_pixs)
+        shift_y = random.randint(-shift_pixs, shift_pixs)
+    if 'scale' in opt.preprocess:
+        scale = 1 + opt.scale_delta * (2 * random.random() - 1)
+    if 'rot' in opt.preprocess:
+        rot_angle = opt.rot_angle * (2 * random.random() - 1)
+        rot_rad = -rot_angle * np.pi/180
+    if 'flip' in opt.preprocess:
+        flip = random.random() > 0.5
+
+    shift_to_origin = np.array([1, 0, -w//2, 0, 1, -h//2, 0, 0, 1]).reshape([3, 3])
+    flip_mat = np.array([-1 if flip else 1, 0, 0, 0, 1, 0, 0, 0, 1]).reshape([3, 3])
+    shift_mat = np.array([1, 0, shift_x, 0, 1, shift_y, 0, 0, 1]).reshape([3, 3])
+    rot_mat = np.array([np.cos(rot_rad), np.sin(rot_rad), 0, -np.sin(rot_rad), np.cos(rot_rad), 0, 0, 0, 1]).reshape([3, 3])
+    scale_mat = np.array([scale, 0, 0, 0, scale, 0, 0, 0, 1]).reshape([3, 3])
+    shift_to_center = np.array([1, 0, w//2, 0, 1, h//2, 0, 0, 1]).reshape([3, 3])
+    
+    affine = shift_to_center @ scale_mat @ rot_mat @ shift_mat @ flip_mat @ shift_to_origin    
+    affine_inv = np.linalg.inv(affine)
+    return affine, affine_inv, flip
+
+def apply_img_affine(img, affine_inv, method=Image.BICUBIC):
+    return img.transform(img.size, Image.AFFINE, data=affine_inv.flatten()[:6], resample=Image.BICUBIC)
+
+def apply_lm_affine(landmark, affine, flip, size):
+    _, h = size
+    lm = landmark.copy()
+    lm[:, 1] = h - 1 - lm[:, 1]
+    lm = np.concatenate((lm, np.ones([lm.shape[0], 1])), -1)
+    lm = lm @ np.transpose(affine)
+    lm[:, :2] = lm[:, :2] / lm[:, 2:]
+    lm = lm[:, :2]
+    lm[:, 1] = h - 1 - lm[:, 1]
+    if flip:
+        lm_ = lm.copy()
+        lm_[:17] = lm[16::-1]
+        lm_[17:22] = lm[26:21:-1]
+        lm_[22:27] = lm[21:16:-1]
+        lm_[31:36] = lm[35:30:-1]
+        lm_[36:40] = lm[45:41:-1]
+        lm_[40:42] = lm[47:45:-1]
+        lm_[42:46] = lm[39:35:-1]
+        lm_[46:48] = lm[41:39:-1]
+        lm_[48:55] = lm[54:47:-1]
+        lm_[55:60] = lm[59:54:-1]
+        lm_[60:65] = lm[64:59:-1]
+        lm_[65:68] = lm[67:64:-1]
+        lm = lm_
+    return lm

src/face3d/data/flist_dataset.py

@@ -0,0 +1,125 @@
+"""This script defines the custom dataset for Deep3DFaceRecon_pytorch
+"""
+
+import os.path
+from data.base_dataset import BaseDataset, get_transform, get_affine_mat, apply_img_affine, apply_lm_affine
+from data.image_folder import make_dataset
+from PIL import Image
+import random
+import util.util as util
+import numpy as np
+import json
+import torch
+from scipy.io import loadmat, savemat
+import pickle
+from util.preprocess import align_img, estimate_norm
+from util.load_mats import load_lm3d
+
+
+def default_flist_reader(flist):
+    """
+    flist format: impath label\nimpath label\n ...(same to caffe's filelist)
+    """
+    imlist = []
+    with open(flist, 'r') as rf:
+        for line in rf.readlines():
+            impath = line.strip()
+            imlist.append(impath)
+
+    return imlist
+
+def jason_flist_reader(flist):
+    with open(flist, 'r') as fp:
+        info = json.load(fp)
+    return info
+
+def parse_label(label):
+    return torch.tensor(np.array(label).astype(np.float32))
+
+
+class FlistDataset(BaseDataset):
+    """
+    It requires one directories to host training images '/path/to/data/train'
+    You can train the model with the dataset flag '--dataroot /path/to/data'.
+    """
+
+    def __init__(self, opt):
+        """Initialize this dataset class.
+
+        Parameters:
+            opt (Option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        BaseDataset.__init__(self, opt)
+        
+        self.lm3d_std = load_lm3d(opt.bfm_folder)
+        
+        msk_names = default_flist_reader(opt.flist)
+        self.msk_paths = [os.path.join(opt.data_root, i) for i in msk_names]
+
+        self.size = len(self.msk_paths) 
+        self.opt = opt
+        
+        self.name = 'train' if opt.isTrain else 'val'
+        if '_' in opt.flist:
+            self.name += '_' + opt.flist.split(os.sep)[-1].split('_')[0]
+        
+
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index (int)      -- a random integer for data indexing
+
+        Returns a dictionary that contains A, B, A_paths and B_paths
+            img (tensor)       -- an image in the input domain
+            msk (tensor)       -- its corresponding attention mask
+            lm  (tensor)       -- its corresponding 3d landmarks
+            im_paths (str)     -- image paths
+            aug_flag (bool)    -- a flag used to tell whether its raw or augmented
+        """
+        msk_path = self.msk_paths[index % self.size]  # make sure index is within then range
+        img_path = msk_path.replace('mask/', '')
+        lm_path = '.'.join(msk_path.replace('mask', 'landmarks').split('.')[:-1]) + '.txt'
+
+        raw_img = Image.open(img_path).convert('RGB')
+        raw_msk = Image.open(msk_path).convert('RGB')
+        raw_lm = np.loadtxt(lm_path).astype(np.float32)
+
+        _, img, lm, msk = align_img(raw_img, raw_lm, self.lm3d_std, raw_msk)
+        
+        aug_flag = self.opt.use_aug and self.opt.isTrain
+        if aug_flag:
+            img, lm, msk = self._augmentation(img, lm, self.opt, msk)
+        
+        _, H = img.size
+        M = estimate_norm(lm, H)
+        transform = get_transform()
+        img_tensor = transform(img)
+        msk_tensor = transform(msk)[:1, ...]
+        lm_tensor = parse_label(lm)
+        M_tensor = parse_label(M)
+
+
+        return {'imgs': img_tensor, 
+                'lms': lm_tensor, 
+                'msks': msk_tensor, 
+                'M': M_tensor,
+                'im_paths': img_path, 
+                'aug_flag': aug_flag,
+                'dataset': self.name}
+
+    def _augmentation(self, img, lm, opt, msk=None):
+        affine, affine_inv, flip = get_affine_mat(opt, img.size)
+        img = apply_img_affine(img, affine_inv)
+        lm = apply_lm_affine(lm, affine, flip, img.size)
+        if msk is not None:
+            msk = apply_img_affine(msk, affine_inv, method=Image.BILINEAR)
+        return img, lm, msk
+    
+
+
+
+    def __len__(self):
+        """Return the total number of images in the dataset.
+        """
+        return self.size

src/face3d/data/image_folder.py

@@ -0,0 +1,66 @@
+"""A modified image folder class
+
+We modify the official PyTorch image folder (https://github.com/pytorch/vision/blob/master/torchvision/datasets/folder.py)
+so that this class can load images from both current directory and its subdirectories.
+"""
+import numpy as np
+import torch.utils.data as data
+
+from PIL import Image
+import os
+import os.path
+
+IMG_EXTENSIONS = [
+    '.jpg', '.JPG', '.jpeg', '.JPEG',
+    '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP',
+    '.tif', '.TIF', '.tiff', '.TIFF',
+]
+
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+
+def make_dataset(dir, max_dataset_size=float("inf")):
+    images = []
+    assert os.path.isdir(dir) or os.path.islink(dir), '%s is not a valid directory' % dir
+
+    for root, _, fnames in sorted(os.walk(dir, followlinks=True)):
+        for fname in fnames:
+            if is_image_file(fname):
+                path = os.path.join(root, fname)
+                images.append(path)
+    return images[:min(max_dataset_size, len(images))]
+
+
+def default_loader(path):
+    return Image.open(path).convert('RGB')
+
+
+class ImageFolder(data.Dataset):
+
+    def __init__(self, root, transform=None, return_paths=False,
+                 loader=default_loader):
+        imgs = make_dataset(root)
+        if len(imgs) == 0:
+            raise(RuntimeError("Found 0 images in: " + root + "\n"
+                               "Supported image extensions are: " + ",".join(IMG_EXTENSIONS)))
+
+        self.root = root
+        self.imgs = imgs
+        self.transform = transform
+        self.return_paths = return_paths
+        self.loader = loader
+
+    def __getitem__(self, index):
+        path = self.imgs[index]
+        img = self.loader(path)
+        if self.transform is not None:
+            img = self.transform(img)
+        if self.return_paths:
+            return img, path
+        else:
+            return img
+
+    def __len__(self):
+        return len(self.imgs)

src/face3d/data/template_dataset.py

@@ -0,0 +1,75 @@
+"""Dataset class template
+
+This module provides a template for users to implement custom datasets.
+You can specify '--dataset_mode template' to use this dataset.
+The class name should be consistent with both the filename and its dataset_mode option.
+The filename should be <dataset_mode>_dataset.py
+The class name should be <Dataset_mode>Dataset.py
+You need to implement the following functions:
+    -- <modify_commandline_options>:　Add dataset-specific options and rewrite default values for existing options.
+    -- <__init__>: Initialize this dataset class.
+    -- <__getitem__>: Return a data point and its metadata information.
+    -- <__len__>: Return the number of images.
+"""
+from data.base_dataset import BaseDataset, get_transform
+# from data.image_folder import make_dataset
+# from PIL import Image
+
+
+class TemplateDataset(BaseDataset):
+    """A template dataset class for you to implement custom datasets."""
+    @staticmethod
+    def modify_commandline_options(parser, is_train):
+        """Add new dataset-specific options, and rewrite default values for existing options.
+
+        Parameters:
+            parser          -- original option parser
+            is_train (bool) -- whether training phase or test phase. You can use this flag to add training-specific or test-specific options.
+
+        Returns:
+            the modified parser.
+        """
+        parser.add_argument('--new_dataset_option', type=float, default=1.0, help='new dataset option')
+        parser.set_defaults(max_dataset_size=10, new_dataset_option=2.0)  # specify dataset-specific default values
+        return parser
+
+    def __init__(self, opt):
+        """Initialize this dataset class.
+
+        Parameters:
+            opt (Option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions
+
+        A few things can be done here.
+        - save the options (have been done in BaseDataset)
+        - get image paths and meta information of the dataset.
+        - define the image transformation.
+        """
+        # save the option and dataset root
+        BaseDataset.__init__(self, opt)
+        # get the image paths of your dataset;
+        self.image_paths = []  # You can call sorted(make_dataset(self.root, opt.max_dataset_size)) to get all the image paths under the directory self.root
+        # define the default transform function. You can use <base_dataset.get_transform>; You can also define your custom transform function
+        self.transform = get_transform(opt)
+
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index -- a random integer for data indexing
+
+        Returns:
+            a dictionary of data with their names. It usually contains the data itself and its metadata information.
+
+        Step 1: get a random image path: e.g., path = self.image_paths[index]
+        Step 2: load your data from the disk: e.g., image = Image.open(path).convert('RGB').
+        Step 3: convert your data to a PyTorch tensor. You can use helpder functions such as self.transform. e.g., data = self.transform(image)
+        Step 4: return a data point as a dictionary.
+        """
+        path = 'temp'    # needs to be a string
+        data_A = None    # needs to be a tensor
+        data_B = None    # needs to be a tensor
+        return {'data_A': data_A, 'data_B': data_B, 'path': path}
+
+    def __len__(self):
+        """Return the total number of images."""
+        return len(self.image_paths)

src/face3d/extract_kp_videos.py

@@ -0,0 +1,108 @@
+import os
+import cv2
+import time
+import glob
+import argparse
+import face_alignment
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from itertools import cycle
+
+from torch.multiprocessing import Pool, Process, set_start_method
+
+class KeypointExtractor():
+    def __init__(self, device):
+        self.detector = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, 
+                                                     device=device)   
+
+    def extract_keypoint(self, images, name=None, info=True):
+        if isinstance(images, list):
+            keypoints = []
+            if info:
+                i_range = tqdm(images,desc='landmark Det:')
+            else:
+                i_range = images
+
+            for image in i_range:
+                current_kp = self.extract_keypoint(image)
+                if np.mean(current_kp) == -1 and keypoints:
+                    keypoints.append(keypoints[-1])
+                else:
+                    keypoints.append(current_kp[None])
+
+            keypoints = np.concatenate(keypoints, 0)
+            np.savetxt(os.path.splitext(name)[0]+'.txt', keypoints.reshape(-1))
+            return keypoints
+        else:
+            while True:
+                try:
+                    keypoints = self.detector.get_landmarks_from_image(np.array(images))[0]
+                    break
+                except RuntimeError as e:
+                    if str(e).startswith('CUDA'):
+                        print("Warning: out of memory, sleep for 1s")
+                        time.sleep(1)
+                    else:
+                        print(e)
+                        break    
+                except TypeError:
+                    print('No face detected in this image')
+                    shape = [68, 2]
+                    keypoints = -1. * np.ones(shape)                    
+                    break
+            if name is not None:
+                np.savetxt(os.path.splitext(name)[0]+'.txt', keypoints.reshape(-1))
+            return keypoints
+
+def read_video(filename):
+    frames = []
+    cap = cv2.VideoCapture(filename)
+    while cap.isOpened():
+        ret, frame = cap.read()
+        if ret:
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frame = Image.fromarray(frame)
+            frames.append(frame)
+        else:
+            break
+    cap.release()
+    return frames
+
+def run(data):
+    filename, opt, device = data
+    os.environ['CUDA_VISIBLE_DEVICES'] = device
+    kp_extractor = KeypointExtractor()
+    images = read_video(filename)
+    name = filename.split('/')[-2:]
+    os.makedirs(os.path.join(opt.output_dir, name[-2]), exist_ok=True)
+    kp_extractor.extract_keypoint(
+        images, 
+        name=os.path.join(opt.output_dir, name[-2], name[-1])
+    )
+
+if __name__ == '__main__':
+    set_start_method('spawn')
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--input_dir', type=str, help='the folder of the input files')
+    parser.add_argument('--output_dir', type=str, help='the folder of the output files')
+    parser.add_argument('--device_ids', type=str, default='0,1')
+    parser.add_argument('--workers', type=int, default=4)
+
+    opt = parser.parse_args()
+    filenames = list()
+    VIDEO_EXTENSIONS_LOWERCASE = {'mp4'}
+    VIDEO_EXTENSIONS = VIDEO_EXTENSIONS_LOWERCASE.union({f.upper() for f in VIDEO_EXTENSIONS_LOWERCASE})
+    extensions = VIDEO_EXTENSIONS
+    
+    for ext in extensions:
+        os.listdir(f'{opt.input_dir}')
+        print(f'{opt.input_dir}/*.{ext}')
+        filenames = sorted(glob.glob(f'{opt.input_dir}/*.{ext}'))
+    print('Total number of videos:', len(filenames))
+    pool = Pool(opt.workers)
+    args_list = cycle([opt])
+    device_ids = opt.device_ids.split(",")
+    device_ids = cycle(device_ids)
+    for data in tqdm(pool.imap_unordered(run, zip(filenames, args_list, device_ids))):
+        None

src/face3d/extract_kp_videos_safe.py

@@ -0,0 +1,151 @@
+import os
+import cv2
+import time
+import glob
+import argparse
+import numpy as np
+from PIL import Image
+import torch
+from tqdm import tqdm
+from itertools import cycle
+from torch.multiprocessing import Pool, Process, set_start_method
+
+from facexlib.alignment import landmark_98_to_68
+from facexlib.detection import init_detection_model
+
+from facexlib.utils import load_file_from_url
+from facexlib.alignment.awing_arch import FAN
+
+def init_alignment_model(model_name, half=False, device='cuda', model_rootpath=None):
+    if model_name == 'awing_fan':
+        model = FAN(num_modules=4, num_landmarks=98, device=device)
+        model_url = 'https://github.com/xinntao/facexlib/releases/download/v0.1.0/alignment_WFLW_4HG.pth'
+    else:
+        raise NotImplementedError(f'{model_name} is not implemented.')
+
+    model_path = load_file_from_url(
+        url=model_url, model_dir='facexlib/weights', progress=True, file_name=None, save_dir=model_rootpath)
+    model.load_state_dict(torch.load(model_path, map_location=device)['state_dict'], strict=True)
+    model.eval()
+    model = model.to(device)
+    return model
+
+
+class KeypointExtractor():
+    def __init__(self, device='cuda'):
+
+        ### gfpgan/weights
+        try:
+            import webui  # in webui
+            root_path = 'extensions/SadTalker/gfpgan/weights' 
+
+        except:
+            root_path = 'gfpgan/weights'
+
+        self.detector = init_alignment_model('awing_fan',device=device, model_rootpath=root_path)   
+        self.det_net = init_detection_model('retinaface_resnet50', half=False,device=device, model_rootpath=root_path)
+
+    def extract_keypoint(self, images, name=None, info=True):
+        if isinstance(images, list):
+            keypoints = []
+            if info:
+                i_range = tqdm(images,desc='landmark Det:')
+            else:
+                i_range = images
+
+            for image in i_range:
+                current_kp = self.extract_keypoint(image)
+                # current_kp = self.detector.get_landmarks(np.array(image))
+                if np.mean(current_kp) == -1 and keypoints:
+                    keypoints.append(keypoints[-1])
+                else:
+                    keypoints.append(current_kp[None])
+
+            keypoints = np.concatenate(keypoints, 0)
+            np.savetxt(os.path.splitext(name)[0]+'.txt', keypoints.reshape(-1))
+            return keypoints
+        else:
+            while True:
+                try:
+                    with torch.no_grad():
+                        # face detection -> face alignment.
+                        img = np.array(images)
+                        bboxes = self.det_net.detect_faces(images, 0.97)
+                        
+                        bboxes = bboxes[0]
+                        img = img[int(bboxes[1]):int(bboxes[3]), int(bboxes[0]):int(bboxes[2]), :]
+
+                        keypoints = landmark_98_to_68(self.detector.get_landmarks(img)) # [0]
+
+                        #### keypoints to the original location
+                        keypoints[:,0] += int(bboxes[0])
+                        keypoints[:,1] += int(bboxes[1])
+
+                        break
+                except RuntimeError as e:
+                    if str(e).startswith('CUDA'):
+                        print("Warning: out of memory, sleep for 1s")
+                        time.sleep(1)
+                    else:
+                        print(e)
+                        break    
+                except TypeError:
+                    print('No face detected in this image')
+                    shape = [68, 2]
+                    keypoints = -1. * np.ones(shape)                    
+                    break
+            if name is not None:
+                np.savetxt(os.path.splitext(name)[0]+'.txt', keypoints.reshape(-1))
+            return keypoints
+
+def read_video(filename):
+    frames = []
+    cap = cv2.VideoCapture(filename)
+    while cap.isOpened():
+        ret, frame = cap.read()
+        if ret:
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frame = Image.fromarray(frame)
+            frames.append(frame)
+        else:
+            break
+    cap.release()
+    return frames
+
+def run(data):
+    filename, opt, device = data
+    os.environ['CUDA_VISIBLE_DEVICES'] = device
+    kp_extractor = KeypointExtractor()
+    images = read_video(filename)
+    name = filename.split('/')[-2:]
+    os.makedirs(os.path.join(opt.output_dir, name[-2]), exist_ok=True)
+    kp_extractor.extract_keypoint(
+        images, 
+        name=os.path.join(opt.output_dir, name[-2], name[-1])
+    )
+
+if __name__ == '__main__':
+    set_start_method('spawn')
+    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+    parser.add_argument('--input_dir', type=str, help='the folder of the input files')
+    parser.add_argument('--output_dir', type=str, help='the folder of the output files')
+    parser.add_argument('--device_ids', type=str, default='0,1')
+    parser.add_argument('--workers', type=int, default=4)
+
+    opt = parser.parse_args()
+    filenames = list()
+    VIDEO_EXTENSIONS_LOWERCASE = {'mp4'}
+    VIDEO_EXTENSIONS = VIDEO_EXTENSIONS_LOWERCASE.union({f.upper() for f in VIDEO_EXTENSIONS_LOWERCASE})
+    extensions = VIDEO_EXTENSIONS
+    
+    for ext in extensions:
+        os.listdir(f'{opt.input_dir}')
+        print(f'{opt.input_dir}/*.{ext}')
+        filenames = sorted(glob.glob(f'{opt.input_dir}/*.{ext}'))
+    print('Total number of videos:', len(filenames))
+    pool = Pool(opt.workers)
+    args_list = cycle([opt])
+    device_ids = opt.device_ids.split(",")
+    device_ids = cycle(device_ids)
+    for data in tqdm(pool.imap_unordered(run, zip(filenames, args_list, device_ids))):
+        None

src/face3d/models/__init__.py

@@ -0,0 +1,67 @@
+"""This package contains modules related to objective functions, optimizations, and network architectures.
+
+To add a custom model class called 'dummy', you need to add a file called 'dummy_model.py' and define a subclass DummyModel inherited from BaseModel.
+You need to implement the following five functions:
+    -- <__init__>:                      initialize the class; first call BaseModel.__init__(self, opt).
+    -- <set_input>:                     unpack data from dataset and apply preprocessing.
+    -- <forward>:                       produce intermediate results.
+    -- <optimize_parameters>:           calculate loss, gradients, and update network weights.
+    -- <modify_commandline_options>:    (optionally) add model-specific options and set default options.
+
+In the function <__init__>, you need to define four lists:
+    -- self.loss_names (str list):          specify the training losses that you want to plot and save.
+    -- self.model_names (str list):         define networks used in our training.
+    -- self.visual_names (str list):        specify the images that you want to display and save.
+    -- self.optimizers (optimizer list):    define and initialize optimizers. You can define one optimizer for each network. If two networks are updated at the same time, you can use itertools.chain to group them. See cycle_gan_model.py for an usage.
+
+Now you can use the model class by specifying flag '--model dummy'.
+See our template model class 'template_model.py' for more details.
+"""
+
+import importlib
+from src.face3d.models.base_model import BaseModel
+
+
+def find_model_using_name(model_name):
+    """Import the module "models/[model_name]_model.py".
+
+    In the file, the class called DatasetNameModel() will
+    be instantiated. It has to be a subclass of BaseModel,
+    and it is case-insensitive.
+    """
+    model_filename = "face3d.models." + model_name + "_model"
+    modellib = importlib.import_module(model_filename)
+    model = None
+    target_model_name = model_name.replace('_', '') + 'model'
+    for name, cls in modellib.__dict__.items():
+        if name.lower() == target_model_name.lower() \
+           and issubclass(cls, BaseModel):
+            model = cls
+
+    if model is None:
+        print("In %s.py, there should be a subclass of BaseModel with class name that matches %s in lowercase." % (model_filename, target_model_name))
+        exit(0)
+
+    return model
+
+
+def get_option_setter(model_name):
+    """Return the static method <modify_commandline_options> of the model class."""
+    model_class = find_model_using_name(model_name)
+    return model_class.modify_commandline_options
+
+
+def create_model(opt):
+    """Create a model given the option.
+
+    This function warps the class CustomDatasetDataLoader.
+    This is the main interface between this package and 'train.py'/'test.py'
+
+    Example:
+        >>> from models import create_model
+        >>> model = create_model(opt)
+    """
+    model = find_model_using_name(opt.model)
+    instance = model(opt)
+    print("model [%s] was created" % type(instance).__name__)
+    return instance

src/face3d/models/arcface_torch/README.md

@@ -0,0 +1,164 @@
+# Distributed Arcface Training in Pytorch
+
+This is a deep learning library that makes face recognition efficient, and effective, which can train tens of millions
+identity on a single server.
+
+## Requirements
+
+- Install [pytorch](http://pytorch.org) (torch>=1.6.0), our doc for [install.md](docs/install.md).
+- `pip install -r requirements.txt`.
+- Download the dataset
+  from [https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_](https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_)
+  .
+
+## How to Training
+
+To train a model, run `train.py` with the path to the configs:
+
+### 1. Single node, 8 GPUs:
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py configs/ms1mv3_r50
+```
+
+### 2. Multiple nodes, each node 8 GPUs:
+
+Node 0:
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=0 --master_addr="ip1" --master_port=1234 train.py train.py configs/ms1mv3_r50
+```
+
+Node 1:
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=2 --node_rank=1 --master_addr="ip1" --master_port=1234 train.py train.py configs/ms1mv3_r50
+```
+
+### 3.Training resnet2060 with 8 GPUs:
+
+```shell
+python -m torch.distributed.launch --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr="127.0.0.1" --master_port=1234 train.py configs/ms1mv3_r2060.py
+```
+
+## Model Zoo
+
+- The models are available for non-commercial research purposes only.  
+- All models can be found in here.  
+- [Baidu Yun Pan](https://pan.baidu.com/s/1CL-l4zWqsI1oDuEEYVhj-g):   e8pw  
+- [onedrive](https://1drv.ms/u/s!AswpsDO2toNKq0lWY69vN58GR6mw?e=p9Ov5d)
+
+### Performance on [**ICCV2021-MFR**](http://iccv21-mfr.com/)
+
+ICCV2021-MFR testset consists of non-celebrities so we can ensure that it has very few overlap with public available face 
+recognition training set, such as MS1M and CASIA as they mostly collected from online celebrities. 
+As the result, we can evaluate the FAIR performance for different algorithms.  
+
+For **ICCV2021-MFR-ALL** set, TAR is measured on all-to-all 1:1 protocal, with FAR less than 0.000001(e-6). The 
+globalised multi-racial testset contains 242,143 identities and 1,624,305 images. 
+
+For **ICCV2021-MFR-MASK** set, TAR is measured on mask-to-nonmask 1:1 protocal, with FAR less than 0.0001(e-4). 
+Mask testset contains 6,964 identities, 6,964 masked images and 13,928 non-masked images. 
+There are totally 13,928 positive pairs and 96,983,824 negative pairs.
+
+| Datasets | backbone  | Training throughout | Size / MB  | **ICCV2021-MFR-MASK** | **ICCV2021-MFR-ALL** |
+| :---:    | :---      | :---                | :---       |:---                   |:---                  |     
+| MS1MV3    | r18  | -              | 91   | **47.85** | **68.33** |
+| Glint360k | r18  | 8536           | 91   | **53.32** | **72.07** |
+| MS1MV3    | r34  | -              | 130  | **58.72** | **77.36** |
+| Glint360k | r34  | 6344           | 130  | **65.10** | **83.02** |
+| MS1MV3    | r50  | 5500           | 166  | **63.85** | **80.53** |
+| Glint360k | r50  | 5136           | 166  | **70.23** | **87.08** |
+| MS1MV3    | r100 | -              | 248  | **69.09** | **84.31** |
+| Glint360k | r100 | 3332           | 248  | **75.57** | **90.66** |
+| MS1MV3    | mobilefacenet | 12185 | 7.8  | **41.52** | **65.26** |        
+| Glint360k | mobilefacenet | 11197 | 7.8  | **44.52** | **66.48** |  
+
+### Performance on IJB-C and Verification Datasets
+
+|   Datasets | backbone      | IJBC(1e-05) | IJBC(1e-04) | agedb30 | cfp_fp | lfw  |  log    |
+| :---:      |    :---       | :---          | :---  | :---  |:---   |:---    |:---     |  
+| MS1MV3     | r18      | 92.07 | 94.66 | 97.77 | 97.73 | 99.77 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_arcface_r18_fp16/training.log)|         
+| MS1MV3     | r34      | 94.10 | 95.90 | 98.10 | 98.67 | 99.80 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_arcface_r34_fp16/training.log)|        
+| MS1MV3     | r50      | 94.79 | 96.46 | 98.35 | 98.96 | 99.83 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_arcface_r50_fp16/training.log)|         
+| MS1MV3     | r100     | 95.31 | 96.81 | 98.48 | 99.06 | 99.85 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_arcface_r100_fp16/training.log)|        
+| MS1MV3     | **r2060**| 95.34 | 97.11 | 98.67 | 99.24 | 99.87 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_arcface_r2060_fp16/training.log)|
+| Glint360k  |r18-0.1   | 93.16 | 95.33 | 97.72 | 97.73 | 99.77 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_cosface_r18_fp16_0.1/training.log)| 
+| Glint360k  |r34-0.1   | 95.16 | 96.56 | 98.33 | 98.78 | 99.82 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_cosface_r34_fp16_0.1/training.log)| 
+| Glint360k  |r50-0.1   | 95.61 | 96.97 | 98.38 | 99.20 | 99.83 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_cosface_r50_fp16_0.1/training.log)| 
+| Glint360k  |r100-0.1  | 95.88 | 97.32 | 98.48 | 99.29 | 99.82 |[log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_cosface_r100_fp16_0.1/training.log)|
+
+[comment]: <> (More details see [model.md]&#40;docs/modelzoo.md&#41; in docs.)
+
+
+## [Speed Benchmark](docs/speed_benchmark.md)
+
+**Arcface Torch** can train large-scale face recognition training set efficiently and quickly. When the number of
+classes in training sets is greater than 300K and the training is sufficient, partial fc sampling strategy will get same
+accuracy with several times faster training performance and smaller GPU memory. 
+Partial FC is a sparse variant of the model parallel architecture for large sacle  face recognition. Partial FC use a 
+sparse softmax, where each batch dynamicly sample a subset of class centers for training. In each iteration, only a 
+sparse part of the parameters will be updated, which can reduce a lot of GPU memory and calculations. With Partial FC, 
+we can scale trainset of 29 millions identities, the largest to date. Partial FC also supports multi-machine distributed 
+training and mixed precision training.
+
+![Image text](https://github.com/anxiangsir/insightface_arcface_log/blob/master/partial_fc_v2.png)
+
+More details see 
+[speed_benchmark.md](docs/speed_benchmark.md) in docs.
+
+### 1. Training speed of different parallel methods (samples / second), Tesla V100 32GB * 8. (Larger is better)
+
+`-` means training failed because of gpu memory limitations.
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+| :---    | :--- | :--- | :--- |
+|125000   | 4681         | 4824          | 5004     |
+|1400000  | **1672**     | 3043          | 4738     |
+|5500000  | **-**        | **1389**      | 3975     |
+|8000000  | **-**        | **-**         | 3565     |
+|16000000 | **-**        | **-**         | 2679     |
+|29000000 | **-**        | **-**         | **1855** |
+
+### 2. GPU memory cost of different parallel methods (MB per GPU), Tesla V100 32GB * 8. (Smaller is better)
+
+| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
+| :---    | :---      | :---      | :---  |
+|125000   | 7358      | 5306      | 4868  |
+|1400000  | 32252     | 11178     | 6056  |
+|5500000  | **-**     | 32188     | 9854  |
+|8000000  | **-**     | **-**     | 12310 |
+|16000000 | **-**     | **-**     | 19950 |
+|29000000 | **-**     | **-**     | 32324 |
+
+## Evaluation ICCV2021-MFR and IJB-C
+
+More details see [eval.md](docs/eval.md) in docs.
+
+## Test
+
+We tested many versions of PyTorch. Please create an issue if you are having trouble.  
+
+- [x] torch 1.6.0
+- [x] torch 1.7.1
+- [x] torch 1.8.0
+- [x] torch 1.9.0
+
+## Citation
+
+```
+@inproceedings{deng2019arcface,
+  title={Arcface: Additive angular margin loss for deep face recognition},
+  author={Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
+  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
+  pages={4690--4699},
+  year={2019}
+}
+@inproceedings{an2020partical_fc,
+  title={Partial FC: Training 10 Million Identities on a Single Machine},
+  author={An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and
+  Zhang, Debing and Fu Ying},
+  booktitle={Arxiv 2010.05222},
+  year={2020}
+}
+```

src/face3d/models/arcface_torch/backbones/__init__.py

@@ -0,0 +1,25 @@
+from .iresnet import iresnet18, iresnet34, iresnet50, iresnet100, iresnet200
+from .mobilefacenet import get_mbf
+
+
+def get_model(name, **kwargs):
+    # resnet
+    if name == "r18":
+        return iresnet18(False, **kwargs)
+    elif name == "r34":
+        return iresnet34(False, **kwargs)
+    elif name == "r50":
+        return iresnet50(False, **kwargs)
+    elif name == "r100":
+        return iresnet100(False, **kwargs)
+    elif name == "r200":
+        return iresnet200(False, **kwargs)
+    elif name == "r2060":
+        from .iresnet2060 import iresnet2060
+        return iresnet2060(False, **kwargs)
+    elif name == "mbf":
+        fp16 = kwargs.get("fp16", False)
+        num_features = kwargs.get("num_features", 512)
+        return get_mbf(fp16=fp16, num_features=num_features)
+    else:
+        raise ValueError()

src/face3d/models/arcface_torch/backbones/iresnet.py

@@ -0,0 +1,187 @@
+import torch
+from torch import nn
+
+__all__ = ['iresnet18', 'iresnet34', 'iresnet50', 'iresnet100', 'iresnet200']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05,)
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05,)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05,)
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.layer2(x)
+            x = self.layer3(x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet18(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet18', IBasicBlock, [2, 2, 2, 2], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet34(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet34', IBasicBlock, [3, 4, 6, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet50(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet50', IBasicBlock, [3, 4, 14, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet100(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet100', IBasicBlock, [3, 13, 30, 3], pretrained,
+                    progress, **kwargs)
+
+
+def iresnet200(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet200', IBasicBlock, [6, 26, 60, 6], pretrained,
+                    progress, **kwargs)
+

src/face3d/models/arcface_torch/backbones/iresnet2060.py

@@ -0,0 +1,176 @@
+import torch
+from torch import nn
+
+assert torch.__version__ >= "1.8.1"
+from torch.utils.checkpoint import checkpoint_sequential
+
+__all__ = ['iresnet2060']
+
+
+def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=3,
+                     stride=stride,
+                     padding=dilation,
+                     groups=groups,
+                     bias=False,
+                     dilation=dilation)
+
+
+def conv1x1(in_planes, out_planes, stride=1):
+    """1x1 convolution"""
+    return nn.Conv2d(in_planes,
+                     out_planes,
+                     kernel_size=1,
+                     stride=stride,
+                     bias=False)
+
+
+class IBasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None,
+                 groups=1, base_width=64, dilation=1):
+        super(IBasicBlock, self).__init__()
+        if groups != 1 or base_width != 64:
+            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
+        if dilation > 1:
+            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
+        self.bn1 = nn.BatchNorm2d(inplanes, eps=1e-05, )
+        self.conv1 = conv3x3(inplanes, planes)
+        self.bn2 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.prelu = nn.PReLU(planes)
+        self.conv2 = conv3x3(planes, planes, stride)
+        self.bn3 = nn.BatchNorm2d(planes, eps=1e-05, )
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        identity = x
+        out = self.bn1(x)
+        out = self.conv1(out)
+        out = self.bn2(out)
+        out = self.prelu(out)
+        out = self.conv2(out)
+        out = self.bn3(out)
+        if self.downsample is not None:
+            identity = self.downsample(x)
+        out += identity
+        return out
+
+
+class IResNet(nn.Module):
+    fc_scale = 7 * 7
+
+    def __init__(self,
+                 block, layers, dropout=0, num_features=512, zero_init_residual=False,
+                 groups=1, width_per_group=64, replace_stride_with_dilation=None, fp16=False):
+        super(IResNet, self).__init__()
+        self.fp16 = fp16
+        self.inplanes = 64
+        self.dilation = 1
+        if replace_stride_with_dilation is None:
+            replace_stride_with_dilation = [False, False, False]
+        if len(replace_stride_with_dilation) != 3:
+            raise ValueError("replace_stride_with_dilation should be None "
+                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
+        self.groups = groups
+        self.base_width = width_per_group
+        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(self.inplanes, eps=1e-05)
+        self.prelu = nn.PReLU(self.inplanes)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=2)
+        self.layer2 = self._make_layer(block,
+                                       128,
+                                       layers[1],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[0])
+        self.layer3 = self._make_layer(block,
+                                       256,
+                                       layers[2],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[1])
+        self.layer4 = self._make_layer(block,
+                                       512,
+                                       layers[3],
+                                       stride=2,
+                                       dilate=replace_stride_with_dilation[2])
+        self.bn2 = nn.BatchNorm2d(512 * block.expansion, eps=1e-05, )
+        self.dropout = nn.Dropout(p=dropout, inplace=True)
+        self.fc = nn.Linear(512 * block.expansion * self.fc_scale, num_features)
+        self.features = nn.BatchNorm1d(num_features, eps=1e-05)
+        nn.init.constant_(self.features.weight, 1.0)
+        self.features.weight.requires_grad = False
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.normal_(m.weight, 0, 0.1)
+            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+        if zero_init_residual:
+            for m in self.modules():
+                if isinstance(m, IBasicBlock):
+                    nn.init.constant_(m.bn2.weight, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
+        downsample = None
+        previous_dilation = self.dilation
+        if dilate:
+            self.dilation *= stride
+            stride = 1
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                conv1x1(self.inplanes, planes * block.expansion, stride),
+                nn.BatchNorm2d(planes * block.expansion, eps=1e-05, ),
+            )
+        layers = []
+        layers.append(
+            block(self.inplanes, planes, stride, downsample, self.groups,
+                  self.base_width, previous_dilation))
+        self.inplanes = planes * block.expansion
+        for _ in range(1, blocks):
+            layers.append(
+                block(self.inplanes,
+                      planes,
+                      groups=self.groups,
+                      base_width=self.base_width,
+                      dilation=self.dilation))
+
+        return nn.Sequential(*layers)
+
+    def checkpoint(self, func, num_seg, x):
+        if self.training:
+            return checkpoint_sequential(func, num_seg, x)
+        else:
+            return func(x)
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.conv1(x)
+            x = self.bn1(x)
+            x = self.prelu(x)
+            x = self.layer1(x)
+            x = self.checkpoint(self.layer2, 20, x)
+            x = self.checkpoint(self.layer3, 100, x)
+            x = self.layer4(x)
+            x = self.bn2(x)
+            x = torch.flatten(x, 1)
+            x = self.dropout(x)
+        x = self.fc(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def _iresnet(arch, block, layers, pretrained, progress, **kwargs):
+    model = IResNet(block, layers, **kwargs)
+    if pretrained:
+        raise ValueError()
+    return model
+
+
+def iresnet2060(pretrained=False, progress=True, **kwargs):
+    return _iresnet('iresnet2060', IBasicBlock, [3, 128, 1024 - 128, 3], pretrained, progress, **kwargs)

src/face3d/models/arcface_torch/backbones/mobilefacenet.py

@@ -0,0 +1,130 @@
+'''
+Adapted from https://github.com/cavalleria/cavaface.pytorch/blob/master/backbone/mobilefacenet.py
+Original author cavalleria
+'''
+
+import torch.nn as nn
+from torch.nn import Linear, Conv2d, BatchNorm1d, BatchNorm2d, PReLU, Sequential, Module
+import torch
+
+
+class Flatten(Module):
+    def forward(self, x):
+        return x.view(x.size(0), -1)
+
+
+class ConvBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(ConvBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, groups=groups, stride=stride, padding=padding, bias=False),
+            BatchNorm2d(num_features=out_c),
+            PReLU(num_parameters=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class LinearBlock(Module):
+    def __init__(self, in_c, out_c, kernel=(1, 1), stride=(1, 1), padding=(0, 0), groups=1):
+        super(LinearBlock, self).__init__()
+        self.layers = nn.Sequential(
+            Conv2d(in_c, out_c, kernel, stride, padding, groups=groups, bias=False),
+            BatchNorm2d(num_features=out_c)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class DepthWise(Module):
+    def __init__(self, in_c, out_c, residual=False, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=1):
+        super(DepthWise, self).__init__()
+        self.residual = residual
+        self.layers = nn.Sequential(
+            ConvBlock(in_c, out_c=groups, kernel=(1, 1), padding=(0, 0), stride=(1, 1)),
+            ConvBlock(groups, groups, groups=groups, kernel=kernel, padding=padding, stride=stride),
+            LinearBlock(groups, out_c, kernel=(1, 1), padding=(0, 0), stride=(1, 1))
+        )
+
+    def forward(self, x):
+        short_cut = None
+        if self.residual:
+            short_cut = x
+        x = self.layers(x)
+        if self.residual:
+            output = short_cut + x
+        else:
+            output = x
+        return output
+
+
+class Residual(Module):
+    def __init__(self, c, num_block, groups, kernel=(3, 3), stride=(1, 1), padding=(1, 1)):
+        super(Residual, self).__init__()
+        modules = []
+        for _ in range(num_block):
+            modules.append(DepthWise(c, c, True, kernel, stride, padding, groups))
+        self.layers = Sequential(*modules)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class GDC(Module):
+    def __init__(self, embedding_size):
+        super(GDC, self).__init__()
+        self.layers = nn.Sequential(
+            LinearBlock(512, 512, groups=512, kernel=(7, 7), stride=(1, 1), padding=(0, 0)),
+            Flatten(),
+            Linear(512, embedding_size, bias=False),
+            BatchNorm1d(embedding_size))
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class MobileFaceNet(Module):
+    def __init__(self, fp16=False, num_features=512):
+        super(MobileFaceNet, self).__init__()
+        scale = 2
+        self.fp16 = fp16
+        self.layers = nn.Sequential(
+            ConvBlock(3, 64 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1)),
+            ConvBlock(64 * scale, 64 * scale, kernel=(3, 3), stride=(1, 1), padding=(1, 1), groups=64),
+            DepthWise(64 * scale, 64 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=128),
+            Residual(64 * scale, num_block=4, groups=128, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(64 * scale, 128 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=256),
+            Residual(128 * scale, num_block=6, groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+            DepthWise(128 * scale, 128 * scale, kernel=(3, 3), stride=(2, 2), padding=(1, 1), groups=512),
+            Residual(128 * scale, num_block=2, groups=256, kernel=(3, 3), stride=(1, 1), padding=(1, 1)),
+        )
+        self.conv_sep = ConvBlock(128 * scale, 512, kernel=(1, 1), stride=(1, 1), padding=(0, 0))
+        self.features = GDC(num_features)
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+                if m.bias is not None:
+                    m.bias.data.zero_()
+
+    def forward(self, x):
+        with torch.cuda.amp.autocast(self.fp16):
+            x = self.layers(x)
+        x = self.conv_sep(x.float() if self.fp16 else x)
+        x = self.features(x)
+        return x
+
+
+def get_mbf(fp16, num_features):
+    return MobileFaceNet(fp16, num_features)

src/face3d/models/arcface_torch/configs/3millions.py

@@ -0,0 +1,23 @@
+from easydict import EasyDict as edict
+
+# configs for test speed
+
+config = edict()
+config.loss = "arcface"
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 1.0
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.rec = "synthetic"
+config.num_classes = 300 * 10000
+config.num_epoch = 30
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = []

src/face3d/models/arcface_torch/configs/3millions_pfc.py

@@ -0,0 +1,23 @@
+from easydict import EasyDict as edict
+
+# configs for test speed
+
+config = edict()
+config.loss = "arcface"
+config.network = "r50"
+config.resume = False
+config.output = None
+config.embedding_size = 512
+config.sample_rate = 0.1
+config.fp16 = True
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+config.rec = "synthetic"
+config.num_classes = 300 * 10000
+config.num_epoch = 30
+config.warmup_epoch = -1
+config.decay_epoch = [10, 16, 22]
+config.val_targets = []

src/face3d/models/arcface_torch/configs/base.py

@@ -0,0 +1,56 @@
+from easydict import EasyDict as edict
+
+# make training faster
+# our RAM is 256G
+# mount -t tmpfs -o size=140G  tmpfs /train_tmp
+
+config = edict()
+config.loss = "arcface"
+config.network = "r50"
+config.resume = False
+config.output = "ms1mv3_arcface_r50"
+
+config.dataset = "ms1m-retinaface-t1"
+config.embedding_size = 512
+config.sample_rate = 1
+config.fp16 = False
+config.momentum = 0.9
+config.weight_decay = 5e-4
+config.batch_size = 128
+config.lr = 0.1  # batch size is 512
+
+if config.dataset == "emore":
+    config.rec = "/train_tmp/faces_emore"
+    config.num_classes = 85742
+    config.num_image = 5822653
+    config.num_epoch = 16
+    config.warmup_epoch = -1
+    config.decay_epoch = [8, 14, ]
+    config.val_targets = ["lfw", ]
+
+elif config.dataset == "ms1m-retinaface-t1":
+    config.rec = "/train_tmp/ms1m-retinaface-t1"
+    config.num_classes = 93431
+    config.num_image = 5179510
+    config.num_epoch = 25
+    config.warmup_epoch = -1
+    config.decay_epoch = [11, 17, 22]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+elif config.dataset == "glint360k":
+    config.rec = "/train_tmp/glint360k"
+    config.num_classes = 360232
+    config.num_image = 17091657
+    config.num_epoch = 20
+    config.warmup_epoch = -1
+    config.decay_epoch = [8, 12, 15, 18]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]
+
+elif config.dataset == "webface":
+    config.rec = "/train_tmp/faces_webface_112x112"
+    config.num_classes = 10572
+    config.num_image = "forget"
+    config.num_epoch = 34
+    config.warmup_epoch = -1
+    config.decay_epoch = [20, 28, 32]
+    config.val_targets = ["lfw", "cfp_fp", "agedb_30"]