File size: 8,549 Bytes
ab01e4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
# Change the numbers when you want to train with specific gpus
# os.environ['CUDA_VISIBLE_DEVICES'] = '0, 1, 2, 3'
import torch
from STTNet import STTNet
import torch.nn.functional as F
from Utils.Datasets import get_data_loader
from Utils.Utils import make_numpy_img, inv_normalize_img, encode_onehot_to_mask, get_metrics, Logger
import matplotlib.pyplot as plt
import numpy as np
from collections import OrderedDict
from torch.optim.lr_scheduler import MultiStepLR

if __name__ == '__main__':
    model_infos = {
        # vgg16_bn, resnet50, resnet18
        'backbone': 'resnet50',
        'pretrained': True,
        'out_keys': ['block4'],
        'in_channel': 3,
        'n_classes': 2,
        'top_k_s': 64,
        'top_k_c': 16,
        'encoder_pos': True,
        'decoder_pos': True,
        'model_pattern': ['X', 'A', 'S', 'C'],

        'BATCH_SIZE': 8,
        'IS_SHUFFLE': True,
        'NUM_WORKERS': 0,
        'DATASET': 'Tools/generate_dep_info/train_data.csv',
        'model_path': 'Checkpoints',
        'log_path': 'Results',
        # if you need the validation process.
        'IS_VAL': True,
        'VAL_BATCH_SIZE': 4,
        'VAL_DATASET': 'Tools/generate_dep_info/val_data.csv',
        # if you need the test process.
        'IS_TEST': True,
        'TEST_DATASET': 'Tools/generate_dep_info/test_data.csv',
        'IMG_SIZE': [512, 512],
        'PHASE': 'seg',

        # INRIA Dataset
        'PRIOR_MEAN': [0.40672500537632994, 0.42829032416229895, 0.39331840468605667],
        'PRIOR_STD': [0.029498464618176873, 0.027740088491668233, 0.028246722411879095],
        # # # WHU Dataset
        # 'PRIOR_MEAN': [0.4352682576428411, 0.44523221318154493, 0.41307610541534784],
        # 'PRIOR_STD': [0.026973196780331585, 0.026424642808887323, 0.02791246590291434],

        # if you want to load state dict
        'load_checkpoint_path': r'E:\BuildingExtractionDataset\INRIA_ckpt_latest.pt',
        # if you want to resume a checkpoint
        'resume_checkpoint_path': '',

    }
    os.makedirs(model_infos['model_path'], exist_ok=True)
    if model_infos['IS_VAL']:
        os.makedirs(model_infos['log_path']+'/val', exist_ok=True)
    if model_infos['IS_TEST']:
        os.makedirs(model_infos['log_path']+'/test', exist_ok=True)
    logger = Logger(model_infos['log_path'] + '/log.log')

    data_loaders = get_data_loader(model_infos)
    loss_weight = 0.1
    model = STTNet(**model_infos)

    epoch_start = 0
    if model_infos['load_checkpoint_path'] is not None and os.path.exists(model_infos['load_checkpoint_path']):
        logger.write(f'load checkpoint from {model_infos["load_checkpoint_path"]}\n')
        state_dict = torch.load(model_infos['load_checkpoint_path'], map_location='cpu')
        model_dict = state_dict['model_state_dict']
        try:
            model_dict = OrderedDict({k.replace('module.', ''): v for k, v in model_dict.items()})
            model.load_state_dict(model_dict)
        except Exception as e:
            model.load_state_dict(model_dict)
    if model_infos['resume_checkpoint_path'] is not None and os.path.exists(model_infos['resume_checkpoint_path']):
        logger.write(f'resume checkpoint path from {model_infos["resume_checkpoint_path"]}\n')
        state_dict = torch.load(model_infos['resume_checkpoint_path'], map_location='cpu')
        epoch_start = state_dict['epoch_id']
        model_dict = state_dict['model_state_dict']
        logger.write(f'resume checkpoint from epoch {epoch_start}\n')
        try:
            model_dict = OrderedDict({k.replace('module.', ''): v for k, v in model_dict.items()})
            model.load_state_dict(model_dict)
        except Exception as e:
            model.load_state_dict(model_dict)
    model = model.cuda()
    device_ids = range(torch.cuda.device_count())
    if len(device_ids) > 1:
        model = torch.nn.DataParallel(model, device_ids=device_ids)
        logger.write(f'Use GPUs: {device_ids}\n')
    else:
        logger.write(f'Use GPUs: 1\n')
    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
    max_epoch = 300
    scheduler = MultiStepLR(optimizer, [int(max_epoch*2/3), int(max_epoch*5/6)], 0.5)

    for epoch_id in range(epoch_start, max_epoch):
        pattern = 'train'
        model.train()  # Set model to training mode
        for batch_id, batch in enumerate(data_loaders[pattern]):
            # Get data
            img_batch = batch['img'].cuda()
            label_batch = batch['label'].cuda()

            # inference
            optimizer.zero_grad()
            logits, att_branch_output = model(img_batch)

            # compute loss
            label_downs = F.interpolate(label_batch, att_branch_output.size()[2:], mode='nearest')
            loss_branch = F.binary_cross_entropy_with_logits(att_branch_output, label_downs)
            loss_master = F.binary_cross_entropy_with_logits(logits, label_batch)
            loss = loss_master + loss_weight * loss_branch
            # loss backward
            loss.backward()
            optimizer.step()

            if batch_id % 20 == 1:
                logger.write(
                    f'{pattern}: {epoch_id}/{max_epoch} {batch_id}/{len(data_loaders[pattern])} loss: {loss.item():.4f}\n')

        scheduler.step()
        patterns = ['val', 'test']
        for pattern_id, is_pattern in enumerate([model_infos['IS_VAL'], model_infos['IS_TEST']]):
            if is_pattern:
                # pred: logits, tensor, nBatch * nClass * W * H
                # target: labels, tensor, nBatch * nClass * W * H
                # output, batch['label']
                collect_result = {'pred': [], 'target': []}
                pattern = patterns[pattern_id]
                model.eval()
                for batch_id, batch in enumerate(data_loaders[pattern]):
                    # Get data
                    img_batch = batch['img'].cuda()
                    label_batch = batch['label'].cuda()
                    img_names = batch['img_name']
                    collect_result['target'].append(label_batch.data.cpu())

                    # inference
                    with torch.no_grad():
                        logits, att_branch_output = model(img_batch)

                    collect_result['pred'].append(logits.data.cpu())
                    # get segmentation result, when the phase is test.
                    pred_label = torch.argmax(logits, 1)
                    pred_label *= 255

                    if pattern == 'test' or batch_id % 5 == 1:
                        batch_size = pred_label.size(0)
                        # k = np.clip(int(0.3 * batch_size), a_min=1, a_max=batch_size)
                        # ids = np.random.choice(range(batch_size), k, replace=False)
                        ids = range(batch_size)
                        for img_id in ids:
                            img = img_batch[img_id].detach().cpu()
                            target = label_batch[img_id].detach().cpu()
                            pred = pred_label[img_id].detach().cpu()
                            img_name = img_names[img_id]

                            img = make_numpy_img(
                                inv_normalize_img(img, model_infos['PRIOR_MEAN'], model_infos['PRIOR_STD']))
                            target = make_numpy_img(encode_onehot_to_mask(target)) * 255
                            pred = make_numpy_img(pred)

                            vis = np.concatenate([img / 255., target / 255., pred / 255.], axis=0)
                            vis = np.clip(vis, a_min=0, a_max=1)
                            file_name = os.path.join(model_infos['log_path'], pattern, f'Epoch_{epoch_id}_{img_name.split(".")[0]}.png')
                            plt.imsave(file_name, vis)

                collect_result['pred'] = torch.cat(collect_result['pred'], dim=0)
                collect_result['target'] = torch.cat(collect_result['target'], dim=0)
                IoU, OA, F1_score = get_metrics('seg', **collect_result)
                logger.write(f'{pattern}: {epoch_id}/{max_epoch} Iou:{IoU[-1]:.4f} OA:{OA[-1]:.4f} F1:{F1_score[-1]:.4f}\n')
        if epoch_id % 20 == 1:
            torch.save({
                'epoch_id': epoch_id,
                'model_state_dict': model.state_dict()
            }, os.path.join(model_infos['model_path'], f'ckpt_{epoch_id}.pt'))
        torch.save({
            'epoch_id': epoch_id,
            'model_state_dict': model.state_dict()
        }, os.path.join(model_infos['model_path'], f'ckpt_latest.pt'))