MoonLite/main.py at main · bikrammajhi/MoonLite · 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
import torch
import numpy as np
import cv2
from tqdm import tqdm

from logger import Logger
from option import get_option
from data import import_loader
from loss import import_loss
from model import import_model


def train(opt, logger):
    logger.info('task: {}, model task: {}'.format(opt.task, opt.model_task))

    train_loader, valid_loader = import_loader(opt)
    lr = float(opt.config['train']['lr'])
    lr_warmup = float(opt.config['train']['lr_warmup'])

    loss_warmup = import_loss('warmup')
    loss_training = import_loss(opt.model_task)
    net = import_model(opt)
    # logger.info(net)

    net.train()
    # Phase Warming-up
    if opt.config['train']['warmup']:
        logger.info('start warming-up')

        optim_warm = torch.optim.Adam(net.parameters(), lr_warmup, weight_decay=0)
        epochs = opt.config['train']['warmup_epoch']
        for epo in range(epochs):
            loss_li = []
            for img_inp, img_gt, _ in tqdm(train_loader, ncols=80):
                optim_warm.zero_grad()
                warmup_out1, warmup_out2 = net.forward_warm(img_inp)
                loss = loss_warmup(img_inp, img_gt, warmup_out1, warmup_out2)
                loss.backward()
                optim_warm.step()
                loss_li.append(loss.item())

            logger.info('epoch: {}, train_loss: {}'.format(epo+1, sum(loss_li)/len(loss_li)))
            torch.save(net.state_dict(), r'{}\model_pre.pkl'.format(opt.save_model_dir))
        logger.info('warming-up phase done')

    # Phase Training
    best_psnr = 0
    epochs = int(opt.config['train']['epoch'])
    optim = torch.optim.Adam(net.parameters(), lr, weight_decay=0)
    lr_sch = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optim, 50, 2, 1e-7)

    logger.info('start training')
    for epo in range(epochs):
        loss_li = []
        test_psnr = []
        net.train()
        for img_inp, img_gt, _ in tqdm(train_loader, ncols=80):
            out = net(img_inp)
            loss = loss_training(out, img_gt)
            optim.zero_grad()
            loss.backward()
            optim.step()
            loss_li.append(loss.item())
        lr_sch.step()

        # Validation
        net.eval()
        for img_inp, img_gt, _ in tqdm(valid_loader, ncols=80):
            with torch.no_grad():
                out = net(img_inp)
                mse = ((out - img_gt)**2).mean((2, 3))
                psnr = (1 / mse).log10().mean() * 10
            test_psnr.append(psnr.item())
        mean_psnr = sum(test_psnr)/len(test_psnr)

        if (epo+1) % int(opt.config['train']['save_every']) == 0:
            torch.save(net.state_dict(), r'{}\model_{}.pkl'.format(opt.save_model_dir, epo+1))

        logger.info('epoch: {}, training loss: {}, validation psnr: {}'.format(
            epo+1, sum(loss_li) / len(loss_li), sum(test_psnr) / len(test_psnr)
        ))

        if mean_psnr > best_psnr:
            best_psnr = mean_psnr
            torch.save(net.state_dict(), r'{}\model_best.pkl'.format(opt.save_model_dir))
            if opt.config['train']['save_slim']:
                net_slim = net.slim().to(opt.device)
                torch.save(net_slim.state_dict(), r'{}\model_best_slim.pkl'.format(opt.save_model_dir))
                logger.info('best model saved and re-parameterized in epoch {}'.format(epo+1))
            else:
                logger.info('best model saved in epoch in epoch {}'.format(epo+1))

    logger.info('training done')


def test(opt, logger):
    test_loader = import_loader(opt)
    net = import_model(opt)
    net.eval()
    psnr_list = []
    logger.info('start testing')
    for (img_inp, img_gt, img_name) in test_loader:

        with torch.no_grad():
            out = net(img_inp)
            mse = ((out - img_gt)**2).mean((2, 3))
            psnr = (1 / mse).log10().mean() * 10

        if opt.config['test']['save']:
            out_img = (out.clip(0, 1)[0] * 255).permute([1, 2, 0]).cpu().numpy().astype(np.uint8)[..., ::-1]
            cv2.imwrite(r'{}\{}.png'.format(opt.save_image_dir, img_name[0]), out_img)

        psnr_list.append(psnr.item())
        logger.info('image name: {}, test psnr: {}'.format(img_name[0], psnr))

    logger.info('testing done, overall psnr: {}'.format(sum(psnr_list) / len(psnr_list)))


def demo(opt, logger):
    demo_loader = import_loader(opt)
    net = import_model(opt)
    net.eval()
    logger.info('start demonstration')
    for img_inp, img_name in demo_loader:

        with torch.no_grad():
            out = net(img_inp)
        out_img = (out.clip(0, 1)[0] * 255).permute([1, 2, 0]).cpu().numpy().astype(np.uint8)[..., ::-1]
        cv2.imwrite(r'{}\{}.png'.format(opt.save_image_dir, img_name[0]), out_img)
        logger.info('image name: {} output generated'.format(img_name[0]))
    logger.info('demonstration done')


if __name__ == "__main__":
    opt = get_option()
    logger = Logger(opt)

    if opt.task == 'train':
        train(opt, logger)
    elif opt.task == 'test':
        test(opt, logger)
    elif opt.task == 'demo':
        demo(opt, logger)
    else:
        raise ValueError('unknown task, please choose from [train, test, demo].')