noisyFER/train_emotion_multi_task.py at main · sanweiliti/noisyFER · 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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
import torch
import time
from tqdm import tqdm
import argparse
from loader.dataloader_raf_multi_task import DataloaderRAF_MultiTask
from loader.dataloader_raf import DataloaderRAF
from loader.dataloader_affectnet_multi_task import DataloaderAffectnet_MultiTask
from utils import *
from torch.utils import data
import random
from tensorboardX import SummaryWriter
from metrics import *
from model.multi_task_model import MultiTaskModel


####### training settings #########
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", default=128, type=int)
parser.add_argument("--img_size", default=64, type=int)
parser.add_argument("--log_step", default=100, type=int)
parser.add_argument("--val_step", default=100, type=int)
parser.add_argument("--save_step", default=200, type=int)
parser.add_argument("--num_epoch", default=100000, type=int)
parser.add_argument("--isTrain", default=True, type=bool)
parser.add_argument("--fc_layer", default=500, type=int)
parser.add_argument("--noise_dim", default=100, type=int, help='dimension of noise vector')

parser.add_argument("--base_dataset", default='raf', type=str, help='base dataset, raf-case/affectnet-case')  # raf/affectnet
parser.add_argument("--gan_start_epoch", default=5, type=int, help='start gan loss from which epoch')  # raf: 5 / affectnet: -1

parser.add_argument("--lambda_exp", default=1.0, type=float)
parser.add_argument("--lambda_va", default=1.0, type=float)
parser.add_argument("--lambda_gan", default=1.0, type=float)

# weights for marginal scores in D
parser.add_argument("--lambda_sx", default=1.0, type=float, help='loss weight for marginal score of image')
parser.add_argument("--lambda_sz0", default=1.0, type=float, help='loss weight for mariginal score of noise vector')
parser.add_argument("--lambda_sz1_exp", default=1.0, type=float,
                    help='loss weight for mariginal score of noisy discrete emotion label')
parser.add_argument("--lambda_sz1_va", default=1.0, type=float,
                    help='loss weight for mariginal score of noisy continuous va label')
# weights for joint score in D
parser.add_argument("--lambda_sxz", default=1.0, type=float,
                    help='loss weights for joint score')

# iteration number for G/D training
parser.add_argument("--lr", default=0.0001, type=float)
parser.add_argument("--iter_G", default=1, type=int)
parser.add_argument("--iter_D", default=2, type=int)

parser.add_argument("--save_dir", default='runs_emotion', type=str)
parser.add_argument("--vgg_pretrain", action='store_true')  # raf: False / affectnet: true
parser.add_argument("--pretrain_vggface_dir", default='vgg_face_dag.pth', type=str)
# affectnet: ../datasets/affectnet/
parser.add_argument("--root", default='../datasets/rafd/basic',
                    type=str, help='path to dataset folder')
args = parser.parse_args()
###################################

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def train(writer, logger):
    if args.base_dataset == 'raf':
        img_root = os.path.join(args.root, 'Image/myaligned/imgs')
        # noisy training lbl: predicted lbl on RAF by pretrained affectnet model, label in raf style:
        # surprise, 2: fear, 3: disgust, 4: happiness, 5: sadness, 6: anger, 7: neutral
        train_exp_csv_file = 'noisy_labels/list_label_AffectnetModel.txt'
        train_va_csv_file = 'noisy_labels/list_va_AffectnetModel.txt'
        # val lbl: original lbl of RAF
        val_csv_file = os.path.join(args.root, 'EmoLabel/list_patition_label.txt')

        train_dataset = DataloaderRAF_MultiTask(img_size=args.img_size, is_transform=True, split='train')
        train_dataset.load_data_exp(train_exp_csv_file, img_root)
        train_dataset.load_data_va(train_va_csv_file)
        train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=args.batch_size, shuffle=True,
                                                       num_workers=8)

        val_dataset = DataloaderRAF(img_size=args.img_size, is_transform=False, split='test')
        val_dataset.load_data(val_csv_file, img_root)
        val_dataloader = torch.utils.data.DataLoader(dataset=val_dataset, batch_size=args.batch_size,
                                                     shuffle=False, num_workers=4)

    elif args.base_dataset == 'affectnet':
        img_root = os.path.join(args.root, 'myaligned')
        # noisy exp training lbl: predicted by RAF pretrained model, labeled in affectnet style
        # 0: Neutral, 1: Happy, 2: Sad, 3: Surprise, 4: Fear, 5: Disgust, 6: Anger
        train_exp_csv_file = 'noisy_labels/training_RAFModel.csv'
        train_va_csv_file = os.path.join(args.root, 'training.csv')
        # test: load affectnet original validation exp lbl
        test_csv_file = os.path.join(args.root, 'validate.csv')

        train_dataset = DataloaderAffectnet_MultiTask(img_size=args.img_size, exp_classes=7, is_transform=True)
        train_dataset.load_data(train_exp_csv_file, train_va_csv_file, img_root)
        train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=args.batch_size, shuffle=False,
                                                       num_workers=8)
        val_dataset = DataloaderAffectnet_MultiTask(img_size=args.img_size, exp_classes=7, is_transform=False)
        val_dataset.load_data(test_csv_file, test_csv_file, img_root)
        val_dataloader = torch.utils.data.DataLoader(dataset=val_dataset, batch_size=args.batch_size, shuffle=False,
                                                     num_workers=2)

    model = MultiTaskModel(args)
    if args.vgg_pretrain:
        init_state_dict = model.encoder.init_vggface_params(args.pretrain_vggface_dir)
        model.encoder.load_state_dict(init_state_dict, strict=False)
    model.setup()
    time_meter = averageMeter()

    total_steps = 0
    for epoch in range(args.num_epoch):
        for step, data in tqdm(enumerate(train_dataloader)):
            start_ts = time.time()
            total_steps += 1

            model.set_input(data)
            model.optimize_params(epoch)
            time_meter.update(time.time() - start_ts)

            if total_steps % args.log_step == 0:
                writer.add_scalar('train/loss_exp', model.loss_class.item(), total_steps)
                print_str = 'Step {:d}/ Epoch {:d}]  loss_exp: {:.4f}  Time/Image: {:.4f}'. \
                    format(step, epoch, model.loss_class.item(), time_meter.val / args.batch_size)
                logger.info(print_str)
                print(print_str)

                writer.add_scalar('train/loss_va', model.loss_va.item(), total_steps)
                print_str = 'Step {:d}/ Epoch {:d}]  loss_va: {:.4f}  Time/Image: {:.4f}'. \
                    format(step, epoch, model.loss_va.item(), time_meter.val / args.batch_size)
                logger.info(print_str)
                print(print_str)

                if epoch > args.gan_start_epoch:
                    writer.add_scalar('train/loss_gan', model.loss_gan.item(), total_steps)
                    print_str = 'Step {:d}/ Epoch {:d}]  loss_gan: {:.4f}  Time/Image: {:.4f}'. \
                        format(step, epoch, model.loss_gan.item(), time_meter.val / args.batch_size)
                    logger.info(print_str)
                    print(print_str)

                    writer.add_scalar('train/G_total_loss', model.loss_G.item(), total_steps)
                    print_str = 'Step {:d}/ Epoch {:d}]  G_total_loss: {:.4f}  Time/Image: {:.4f}'. \
                        format(step, epoch, model.loss_G.item(), time_meter.val / args.batch_size)
                    logger.info(print_str)
                    print(print_str)
                    time_meter.reset()

                    # D loss
                    writer.add_scalar('train/D_total_loss', model.loss_D_gan.item(), total_steps)
                    print_str = 'Step {:d}/ Epoch {:d}]  loss_D_gan: {:.4f}  Time/Image: {:.4f}'. \
                        format(step, epoch, model.loss_D_gan.item(), time_meter.val / args.batch_size)
                    logger.info(print_str)
                    print(print_str)


            if total_steps % args.val_step == 0:
                model.encoder.eval()
                total, correct, correct_clean, correct_noise = 0, 0, 0, 0
                # if not args.noise:
                with torch.no_grad():
                    for step_val, data in tqdm(enumerate(val_dataloader)):
                        img = data[0].to(device)
                        clean_lbl = data[1].to(device)
                        _, _, _, z1_enc = model.encoder(img)
                        z1_exp_enc = z1_enc[:, 0:7]

                        _, predicted = z1_exp_enc.max(1)
                        total += clean_lbl.size(0)
                        correct += predicted.eq(clean_lbl).sum().item()

                    acc = 100.0 * correct / total
                    logger.info('acc:{}'.format(acc))
                    writer.add_scalar('val/acc', acc, total_steps)
                    print('acc:', acc)

                if total_steps % args.save_step == 0:
                    save_suffix = 'last'
                    model.save_networks(save_suffix, writer.file_writer.get_logdir())
                    print("[*] models saved.")
                    logger.info("[*] models saved.")


        if epoch == args.num_epoch:
            break


if __name__ == '__main__':
    run_id = random.randint(1, 100000)
    logdir = os.path.join(args.save_dir, str(run_id))  # create new path
    writer = SummaryWriter(log_dir=logdir)
    print('RUNDIR: {}'.format(logdir))

    logger = get_logger(logdir)
    logger.info('Let the games begin')  # write in log file
    save_config(logdir, args)
    train(writer, logger)