fpet/main.py at main · kyk120/fpet · 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
import torch
from torch import nn
from torch.optim import AdamW
from torch.nn import functional as F
from tqdm import tqdm
import timm
from timm.models import create_model
from timm.scheduler.cosine_lr import CosineLRScheduler
from argparse import ArgumentParser
from vtab import *
from utils import save, load, load_config, set_seed, QLinear, AverageMeter, log
import adaptformer
import lora
import fpet
from torch.utils.tensorboard import SummaryWriter
import time
import datetime


def train(args, model, dl, opt, scheduler, epoch, log_writer):
    model.train()
    model = model.cuda()
    pbar = tqdm(range(epoch))
    start_train = time.time()
    for ep in pbar:
        model.train()
        model = model.cuda()
        for i, batch in enumerate(dl):
            x, y = batch[0].cuda(), batch[1].cuda()
            out = model(x)
            loss = F.cross_entropy(out, y)
            opt.zero_grad()
            loss.backward()
            opt.step()
        if scheduler is not None:
            scheduler.step(ep)
        if ep % 10 == 9:
            end_train = time.time()
            start_test = time.time()
            acc = test(vit, test_dl)
            end_test = time.time()
            train_time = (end_train-start_train)/(len(train_dl)*10)
            test_time = (end_test-start_test)/len(test_dl)
            log(args, acc, train_time, test_time, loss.item(), ep)
            log_writer.add_scalar('acc', acc, ep)
            log_writer.add_scalar('train_time', train_time, ep)
            log_writer.add_scalar('test_time', test_time, ep)
            log_writer.add_scalar('loss', loss.item(), ep)
            log_writer.flush()
            if acc > args.best_acc:
                args.best_acc = acc
                args.best_ep = ep
                save(args, model)
            pbar.set_description('best_acc ' + str(args.best_acc))
            start_train = time.time()

    model = model.cpu()
    return model


@torch.no_grad()
def test(model, dl):
    model.eval()
    acc = AverageMeter()
    model = model.cuda()
    for batch in tqdm(dl):
        x, y = batch[0].cuda(), batch[1].cuda()
        out = model(x).data
        acc.update(out, y)
    return acc.result().item()


if __name__ == '__main__':
    parser = ArgumentParser()
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--bit', type=int, default=1, choices=[1, 2, 4, 8, 32])
    parser.add_argument('--dim', type=int, default=32)
    parser.add_argument('--scale', type=float, default=1)
    parser.add_argument('--lr', type=float, default=1e-3)
    parser.add_argument('--wd', type=float, default=1e-4)
    parser.add_argument('--model', type=str, default='vit_base_patch16_224_in21k')
    parser.add_argument('--dataset', type=str, default='cifar')
    parser.add_argument('--method', type=str, default='adaptformer',
                        choices=['adaptformer', 'adaptformer-bihead', 'lora', 'lora-bihead'])
    parser.add_argument('--eval', action='store_true', default=False)
    parser.add_argument('--config_path', type=str, default='.')
    parser.add_argument('--model_path', type=str, default='.')
    parser.add_argument('--load_config', action='store_true', default=False)
    parser.add_argument('--log_path', type=str, default='.')
    parser.add_argument('--r_layer', default=False, type=int)
    args = parser.parse_args()
    print(args)
    if args.eval or args.load_config:
        load_config(args)
    set_seed(args.seed)
    args.best_acc = 0
    args.best_ep = 0
    vit = create_model(args.model, checkpoint_path='ViT-B_16.npz', drop_path_rate=0.1)
    train_dl, test_dl = get_data(args.dataset, normalize=False)

    save_folder = os.path.join(args.model_path, args.dataset)
    if not os.path.exists(save_folder):
        os.makedirs(save_folder)
    log_writer = SummaryWriter(log_dir=os.path.join(args.model_path, args.dataset))

    if args.method == 'adaptformer':
        adaptformer.set_adapter(vit, dim=args.dim, s=args.scale, bit=args.bit)
        vit.reset_classifier(get_classes_num(args.dataset))
    elif args.method == 'lora':
        lora.set_adapter(vit, dim=args.dim, s=args.scale, bit=args.bit)
        vit.reset_classifier(get_classes_num(args.dataset))

    if args.r_layer != False:
        fpet.apply_fpet.apply_patch(vit, method = args.method, r_layer = args.r_layer)
        print(f'FPET applied!')

    if not args.eval:
        import shutil
        shutil.copytree('./configs', os.path.join(save_folder, 'codes/configs'))
        shutil.copytree('./fpet', os.path.join(save_folder, 'codes/fpet'))
        shutil.copy('adaptformer.py', os.path.join(save_folder, 'codes/adaptformer.py'))
        shutil.copy('lora.py', os.path.join(save_folder, 'codes/lora.py'))
        shutil.copy('main.py', os.path.join(save_folder, 'codes/main.py'))
        shutil.copy('utils.py', os.path.join(save_folder, 'codes/utils.py'))
        shutil.copy('vtab.py', os.path.join(save_folder, 'codes/vtab.py'))

        import sys
        with open(os.path.join(save_folder, 'params.sh'), 'w+') as out:
            sys.argv[0] = os.path.join(os.getcwd(), sys.argv[0])
            out.write('#!/bin/bash\n')
            out.write('python3 ')
            out.write(' '.join(sys.argv))
            out.write('\n')

        start_total = time.time()

    if not args.eval:
        trainable = []
        trainable_n = []
        for n, p in vit.named_parameters():
            if ('adapter' in n or 'head' in n) and p.requires_grad:
                trainable.append(p)
                trainable_n.append(n)
            else:
                p.requires_grad = False
        print(f'{trainable_n=}')
        opt = AdamW(trainable, lr=args.lr, weight_decay=args.wd)
        scheduler = CosineLRScheduler(opt, t_initial=100,
                                      warmup_t=10, lr_min=1e-5, warmup_lr_init=1e-6, decay_rate=0.1)
        vit = train(args, vit, train_dl, opt, scheduler, epoch=100, log_writer=log_writer)

        end_total = time.time()
        total_time = end_total - start_total
        log_stats = {"best_epoch": args.best_ep, "best_acc": args.best_acc, "total_train_time": str(datetime.timedelta(seconds=int(total_time)))}
        log(args, 0, 0, 0, 0, 0, log_stats=log_stats)

    else:
        load(args, vit)
        args.best_acc = test(vit, test_dl)

    print('best_acc:', args.best_acc)