ProtoCLR/train_utils.py at main · ilyassmoummad/ProtoCLR · 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
from eval_utils import fewshot_val
from tqdm import tqdm
import torch
import math
import os

def cosine_lr_scheduler(optimizer, epoch, args):
    """Decay the learning rate based on schedule"""
    lr = args.lr
    # cosine lr schedule
    lr *= 0.5 * (1. + math.cos(math.pi * epoch / args.epochs))

    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

def train_epoch(encoder, projector, train_loader, train_transform, loss_fn, optim, scaler, epoch, args):

    encoder.train()
    projector.train()

    print("Epoch {}".format(epoch+1))
    cosine_lr_scheduler(optim, epoch, args)

    tr_loss = 0.

    if args.loss == 'ce':

        for idx, batch in enumerate(tqdm(train_loader)):

            x, y = batch

            optim.zero_grad()

            with torch.autocast(device_type='cuda', dtype=torch.bfloat16):

                x = x.to(args.device)
                y = y.to(args.device)

                with torch.no_grad():

                    x = train_transform(x)

                z = encoder(x)

                h = projector(z)

                loss = loss_fn(h, y)

            tr_loss += loss.item()

            scaler.scale(loss).backward()
            scaler.step(optim)
            scaler.update()

    else:

        for idx, batch in enumerate(tqdm(train_loader)):

            x, y = batch

            optim.zero_grad()

            with torch.autocast(device_type='cuda', dtype=torch.bfloat16):

                x = x.to(args.device)
                y = y.to(args.device)

                with torch.no_grad():

                    x1 = train_transform(x); x2 = train_transform(x)

                z1 = encoder(x1); z2 = encoder(x2)

                h1 = projector(z1); h2 = projector(z2)

                if args.loss in ['supcon', 'protoclr']:
                    loss = loss_fn(h1, h2, y)

                elif args.loss == 'simclr':
                    loss = loss_fn(h1, h2)

            tr_loss += loss.item()

            scaler.scale(loss).backward()
            scaler.step(optim)
            scaler.update()

    tr_loss = tr_loss/len(train_loader)
    print('Train Loss: {}'.format(tr_loss))

    if args.loss == 'ce':
        return encoder, projector, tr_loss
    else:
        return encoder, tr_loss

def train(encoder, projector, train_loader, val_dataset, train_transform, val_transform, loss_fn, optim, args):

    print(f"Training starting on {args.device}")

    num_epochs = args.epochs

    encoder = encoder.to(args.device)
    projector = projector.to(args.device)

    scaler = torch.cuda.amp.GradScaler()

    if args.loss == 'ce':
        best_acc = 0.

    for epoch in range(num_epochs):
        if args.loss == 'ce':
            encoder, projector, tr_loss = train_epoch(encoder, projector, train_loader, train_transform, loss_fn, optim, scaler, epoch, args)
        else:
            encoder, tr_loss = train_epoch(encoder, projector, train_loader, train_transform, loss_fn, optim, scaler, epoch, args)
        val_acc, val_std = fewshot_val(encoder, val_dataset, val_transform, args, shot=1)
        print(f"Val Acc: {val_acc}+-{val_std}")

        if args.loss == 'ce':
            if val_acc > best_acc:
                best_acc = val_acc
                best_state_dict_encoder = encoder.state_dict()
                best_state_dict_projector = projector.state_dict()

        if args.savefreq:
            if (epoch+1) % args.savefreq == 0:
                os.makedirs(args.modelpath, exist_ok=True)
                best_model_path = os.path.join(args.modelpath, args.model + '_' + args.loss + '_pretrain_' + args.pretrainds + '_epochs' + str(args.epochs) + '_lr' + str(args.lr) + '_trial1.pth')
                if os.path.isfile(best_model_path):
                    i = 1
                    while os.path.isfile(best_model_path):
                        i += 1
                        best_model_path = best_model_path[:-5] + str(i) + '.pth'
                strg = 'epochs' + str(args.epochs)
                newstrg = 'epoch' + str(epoch+1)
                freq_model_path = best_model_path.replace(strg, newstrg)

                if args.loss == 'ce':
                    torch.save({'encoder': best_state_dict_encoder, 'classifier': best_state_dict_projector}, freq_model_path)
                else:
                    torch.save(encoder.state_dict(), freq_model_path)

    if args.loss == 'ce':
        return best_state_dict_encoder, best_state_dict_projector
    else:
        return encoder.state_dict()