DualTalk/main.py at main · ZiqiaoPeng/DualTalk · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
#-*- coding : utf-8-*-
import time
import numpy as np
import argparse
from tqdm import tqdm
import os, shutil
import torch
import torch.nn as nn
from dataset_train import get_loader
from DualTalk import DualTalkModel

def trainer(args, train_loader, dev_loader, model, optimizer, criterion,epoch,last_train):
    save_path = args.save_path
    os.makedirs(save_path, exist_ok=True)

    log_file = os.path.join(save_path, 'training_log.txt')
    iteration = 0
    lr_list = []
    scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size = 50, gamma=0.8) #TODO
    for e in range(epoch):
        loss_log = []
        model.train()
        pbar = tqdm(enumerate(train_loader),total=len(train_loader))
        optimizer.zero_grad()

        for i, (file_name, audio1, audio2, exp1, jawpose1, neck1, exp2, jawpose2, neck2) in pbar:
            iteration += 1
            audio1 = audio1.to(args.device)
            audio2 = audio2.to(args.device)
            exp1 = exp1.float().to(args.device) #[bs,frames,52]
            jawpose1 = jawpose1.float().to(args.device)
            neck1 = neck1.float().to(args.device)
            exp2 = exp2.float().to(args.device)
            jawpose2 = jawpose2.float().to(args.device)
            neck2 = neck2.float().to(args.device)
            if exp1.shape[1] > args.max_seq_len:
                exp1 = exp1[:,:args.max_seq_len,:]
                jawpose1 = jawpose1[:,:args.max_seq_len,:]
                neck1 = neck1[:,:args.max_seq_len,:]
                audio1 = audio1[:,:int(args.max_seq_len/25*16000)]
            if exp2.shape[1] > args.max_seq_len:
                exp2 = exp2[:,:args.max_seq_len,:]
                jawpose2 = jawpose2[:,:args.max_seq_len,:]
                neck2 = neck2[:,:args.max_seq_len,:]
                audio2 = audio2[:,:int(args.max_seq_len/25*16000)]
            blendshape1 = torch.cat((exp1, jawpose1, neck1), dim=2)
            blendshape2 = torch.cat((exp2, jawpose2, neck2), dim=2)
            bs_output = model(audio1,audio2,blendshape2)
            length = min(bs_output.shape[1],blendshape1.shape[1])
            bs_output = bs_output[:,:length,:]
            blendshape1 = blendshape1[:,:length,:]
            loss1 = criterion(bs_output[:,:,:50], blendshape1[:,:,:50])
            loss_jaw = criterion(bs_output[:,:, 50:53], blendshape1[:,:, 50:53])
            loss_neck = criterion(bs_output[:,:, 53:56], blendshape1[:,:, 53:56])
            gt_vel = blendshape1[:,1:, :] - blendshape1[:,:-1, :]
            pred_vel = bs_output[:,1:, :] - bs_output[:,:-1, :]
            loss3 = criterion(pred_vel, gt_vel)
            loss = torch.mean(loss1+loss_jaw+loss_neck+loss3)
            optimizer.zero_grad()
            loss.backward()
            loss_log.append(loss.item())
            if i % args.gradient_accumulation_steps == 0:
                optimizer.step()
                optimizer.zero_grad()
            pbar.set_description("(Epoch {}, iteration {}) TRAIN LOSS:{:.7f} EXP LOSS:{:.4f} JAW LOSS:{:.4f} NECK LOSS:{:.4f} VELOCITY LOSS:{:.4f} LR:{:.7f}".format((e+1+last_train), iteration ,np.mean(loss_log),loss1.item(),loss_jaw.item(),loss_neck.item(),loss3.item(),scheduler.get_last_lr()[0]))

        val_loss_log = []
        model.eval()
        with torch.no_grad():
            for file_name, audio1, audio2, exp1, jawpose1, neck1, exp2, jawpose2, neck2 in dev_loader:
                audio1 = audio1.to(args.device)
                audio2 = audio2.to(args.device)
                exp1 = exp1.float().to(args.device)
                jawpose1 = jawpose1.float().to(args.device)
                neck1 = neck1.float().to(args.device)
                exp2 = exp2.float().to(args.device)
                jawpose2 = jawpose2.float().to(args.device)
                neck2 = neck2.float().to(args.device)
                if exp1.shape[1] > args.max_seq_len:
                    exp1 = exp1[:,:args.max_seq_len,:]
                    jawpose1 = jawpose1[:,:args.max_seq_len,:]
                    neck1 = neck1[:,:args.max_seq_len,:]
                    audio1 = audio1[:,:int(args.max_seq_len/25*16000)]
                if exp2.shape[1] > args.max_seq_len:
                    exp2 = exp2[:,:args.max_seq_len,:]
                    jawpose2 = jawpose2[:,:args.max_seq_len,:]
                    neck2 = neck2[:,:args.max_seq_len,:]
                    audio2 = audio2[:,:int(args.max_seq_len/25*16000)]
                blendshape1 = torch.cat((exp1, jawpose1, neck1), dim=2)
                blendshape2 = torch.cat((exp2, jawpose2, neck2), dim=2)
                bs_output = model(audio1,audio2,blendshape2)
                length = min(bs_output.shape[1],blendshape1.shape[1])
                bs_output = bs_output[:,:length,:]
                blendshape1 = blendshape1[:,:length,:]
                val_exp = criterion(bs_output[:,:,:50], blendshape1[:,:,:50])
                val_jaw = criterion(bs_output[:,:, 50:53], blendshape1[:,:, 50:53])
                val_neck = criterion(bs_output[:,:, 53:56], blendshape1[:,:, 53:56])
                val_loss_log.append(val_exp.item()+val_jaw.item()+val_neck.item())

        print("epoch {} val all loss:{:.7f}, exp loss: {:.7f}, jaw loss: {:.7f}, neck loss: {:.7f}".format(e+1, np.mean(val_loss_log), val_exp.item(), val_jaw.item(), val_neck.item()))

        scheduler.step()
        lr_list.append(scheduler.get_last_lr())
        if (e%10 ==0 and e>=1):
            torch.save(model.state_dict(), os.path.join(save_path, "model_{}.pth".format(e+last_train)))
            print("Model saved at epoch {}".format(e+last_train))

        with open(log_file, 'a') as f:
            f.write(f"Epoch {e+1+last_train}, "
                    f"Train Loss: {np.mean(loss_log):.7f}, "
                    f"Train Exp Loss: {loss1.item():.7f}, "
                    f"Train Jaw Loss: {loss_jaw.item():.7f}, "
                    f"Train Neck Loss: {loss_neck.item():.7f}, "
                    f"Val Loss: {np.mean(val_loss_log):.7f}, "
                    f"Val Exp Loss: {val_exp.item():.7f}, "
                    f"Val Jaw Loss: {val_jaw.item():.7f}, "
                    f"Val Neck Loss: {val_neck.item():.7f}, "
                    f"Learning Rate: {scheduler.get_last_lr()[0]:.7f}\n")

    return model


def main():
    parser = argparse.ArgumentParser(description='DualTalk: Dual-Speaker Interaction for 3D Talking Head Conversations')
    parser.add_argument("--lr", type=float, default=0.0001, help='learning rate')
    parser.add_argument("--train_data_path", type=str, default= "./data/train/", help='path of the training data')
    parser.add_argument("--val_data_path", type=str, default= "./data/test/", help='path of the validation data')
    parser.add_argument("--test_data_path", type=str, default= None, help='path of the test data')
    parser.add_argument("--seed", type=int, default=6666, help='random seed')
    parser.add_argument("--scale",type=str, default="large",help="large or base")
    parser.add_argument("--blendshape_dim", type=int, default=56, help='number of blendshapes:52')
    parser.add_argument("--feature_dim", type=int, default=256, help='64 for vocaset; 128 for BIWI')
    parser.add_argument("--period", type=int, default=25, help='period in PPE - 30 for vocaset; 25 for BIWI')
    parser.add_argument("--gradient_accumulation_steps", type=int, default=1, help='gradient accumulation')
    parser.add_argument("--max_epoch", type=int, default=1000, help='number of epochs')
    parser.add_argument("--last_train", type=int, default=0, help='last train')
    parser.add_argument("--device", type=str, default="cuda:1", help='cuda:0 or cuda:1')
    parser.add_argument("--save_path", type=str, default="save_DualTalk", help='path of the trained models')
    parser.add_argument("--result_path", type=str, default="result_DualTalk", help='path to the predictions')
    parser.add_argument("--load_path", type=str, default=None, help='path to the trained models')
    parser.add_argument("--max_seq_len", type=int, default=1800, help='max_seq_len')
    parser.add_argument("--num_workers", type=int, default=8)
    parser.add_argument("--batch_size", type=int, default=32)

    args = parser.parse_args()
    model = DualTalkModel(args)
    model = model.to(args.device)
    dataset = get_loader(args)

    criterion = nn.MSELoss()
    optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=args.lr)
    model = trainer(args, dataset['train'],dataset['val'],model, optimizer,criterion, args.max_epoch,args.last_train)

if __name__ == "__main__":
    main()