UniLMMV/trainer.py at master · uncbiag/UniLMMV · 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
194
195
196
import os
import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from tensorboardX import SummaryWriter
from torch.nn.utils.clip_grad import clip_grad_norm_
from sklearn.metrics import f1_score, roc_auc_score, balanced_accuracy_score, accuracy_score, average_precision_score

from utils import get_WOMAC_label, get_KLG_label, aug_img


class Trainer():
    def __init__(self, config, net, logger, result_types):
        self.config = config
        self.net = net
        self.logger = logger
        self.result_types = result_types

        self.ckp_folder = os.path.join(config['save_folder'], 'checkpoint')
        os.makedirs(self.ckp_folder, exist_ok=True)
        tensorboard_folder = os.path.join(config['save_folder'], 'tensorboard')
        os.makedirs(tensorboard_folder, exist_ok=True)
        self.writer = SummaryWriter(tensorboard_folder)

        self.best = 0
        self.loss_types = ['cls']
        self.loss_dict, self.result_dict = dict(), dict()

    def clear_loss(self):
        for t in self.loss_types:
            self.loss_dict[t] = 0

    def clear_result(self):
        for t in self.result_types:
            self.result_dict[t] = 0

    def set_zero_grad(self):
        self.net.module.zero_grad()

    def set_train(self):
        self.net.module.train()

    def set_eval(self):
        self.net.module.eval()

    def save_networks(self, save_path):
        torch.save(self.net.state_dict(), save_path)

    def prepare_data(self, data, augmentation):
        months = list(data['label'].keys())
        views = self.config['views'].copy()

        if 'WOMAC' in self.config['data']['label_sheet_name']:
            score = get_WOMAC_label(data['label'])
        else:
            score = get_KLG_label(data['label'])
        label = []
        for m in months:
            label.append(score[m].type(torch.LongTensor).cuda())

        if 'tab' in views:      # tab in views
            tab, tab_mask = [], []
            for m in months:
                t = data['tab'][m]
                tm = data['tab_mask'][m]
                tab.append([t[i].cuda() for i in range(len(t))])      # [cat, cont]
                tab_mask.append([tm[i].cuda() for i in range(len(tm))])
            views.remove('tab')
        else:
            tab = tab_mask = None

        if len(views) > 0:      # image in views
            image, image_mask = [[] for _ in range(len(views))], [[] for _ in range(len(views))]
            zero_img = torch.zeros_like(data[views[0]]['0'][0]).to(torch.float)
            for i in range(len(views)):
                for m in months:
                    if 'knee' in views[i] or 'thickness' in views[i] or (m == '0' or m == '48'):       # (is knee) or (is thickness) or (is 0 or 48m)
                        img = data[views[i]][m].to(torch.float)
                        if augmentation:
                            contract_adjust = False if 'thickness' in views[i] else True
                            img = aug_img(img, zero_img, contract_adjust)
                        image[i].append(img.unsqueeze(1).cuda())
                        image_mask[i].append(data['img_mask'][m][i])
                    else:
                        image[i].append(None)
                        image_mask[i].append(torch.zeros_like(data['img_mask']['0'][0]))
        else:
            image = image_mask = None
        return image, image_mask, tab, tab_mask, label

    def print_loss_result(self, phase, prefix, length, is_train, print_result=True):
        # print loss
        if len(self.loss_dict.keys()) > 0:
            for key in list(self.loss_dict.keys()):
                self.loss_dict[key] /= length
            losses = ''
            for key in self.loss_dict:
                losses += ('%s: %.2f | ' % (key, self.loss_dict[key].data.item()))
                if is_train:
                    self.writer.add_scalar('%s/loss_%s' % (phase, key), self.loss_dict[key].data.item(), self.epoch)
            output = '{} \n\t{}'.format(prefix, losses)
        else:
            output = ''

        # print result
        if print_result:
            result = ''
            for key in self.result_dict:
                result += ('%s: %.4f | ' % (key, self.result_dict[key]))
                if is_train:
                    self.writer.add_scalar('%s/%s' % (phase, key), self.result_dict[key], self.epoch)
            output = "{}\n\t{}".format(output, result)

        self.logger.info(output)

    def get_result(self, pred, label):
        pred = np.array(pred)
        label = np.array(label)
        if 'roc' in self.result_types:
            roc = roc_auc_score(y_true=label, y_score=pred[:, 1]) if self.config['attention']['output_dim'] == 2 \
                else roc_auc_score(y_true=label, y_score=pred, multi_class='ovr')
            self.result_dict['roc'] += roc
        if 'ap' in self.result_types:
            ap = average_precision_score(y_true=label, y_score=pred[:, 1]) if self.config['attention']['output_dim'] == 2 \
                else average_precision_score(y_true=label, y_score=pred)
            self.result_dict['ap'] += ap
        pred = np.argmax(pred, axis=1)
        if 'f1score' in self.result_types:
            f1score = f1_score(label, pred, average='macro')
            self.result_dict['f1score'] += f1score
        if 'bal_acc' in self.result_types:
            bal_acc = balanced_accuracy_score(label, pred)
            self.result_dict['bal_acc'] += bal_acc
        if 'acc' in self.result_types:
            acc = accuracy_score(label, pred)
            self.result_dict['acc'] += acc

    def one_iter(self, data, class_weight=None, augmentation=False):
        image, image_mask, tab, tab_mask, label = self.prepare_data(data, augmentation=augmentation)

        label = torch.stack(label).permute(1, 0).reshape(-1)
        pred, _, _, _ = self.net(image, image_mask, tab, tab_mask, is_train=augmentation)
        pred = pred.view(-1, pred.size(-1))

        pred = pred[label != -1]        # remove those that do not have a ground truth label
        label = label[label != -1]
        cls_loss = F.cross_entropy(pred, label, weight=class_weight) if len(label) > 0 else 0
        self.loss_dict['cls'] += cls_loss
        return cls_loss, nn.Softmax(dim=1)(pred), label

    def train(self, train_loader, val_loader, optimizer, scheduler, class_weight, params):
        self.set_train()
        for self.epoch in range(self.config['training']['epoch']):
            self.clear_loss()
            self.clear_result()
            all_pred, all_label = [], []
            for batch_idx, data in enumerate(train_loader):
                self.set_zero_grad()
                train_loss, pred, label = self.one_iter(data, class_weight, augmentation=self.config['training']['augmentation'])
                train_loss.backward()
                if self.config['training']['grad_clip'] > 0:
                    clip_grad_norm_(params, self.config['training']['grad_clip'])
                for i in range(len(optimizer)):
                    optimizer[i].step()
                for i in range(len(scheduler)):
                    scheduler[i].step()
                all_pred.extend(pred.data.cpu().numpy().tolist())
                all_label.extend(label.data.cpu().numpy().tolist())
            self.get_result(all_pred, all_label)
            if (self.epoch + 1) % self.config['eval_freq'] == 0:
                prefix = "Epoch : {:.0f}/{:.0f}\n  Train".format((self.epoch + 1), self.config['training']['epoch'])
                self.print_loss_result('train', prefix, len(train_loader), is_train=True)
                self.eval(val_loader, is_train=True)    # evalution

    def eval(self, data_loader, is_train):
        self.set_eval()
        with torch.no_grad():
            self.clear_loss()
            self.clear_result()
            all_pred, all_label = [], []
            for batch_idx, data in enumerate(data_loader):
                self.set_zero_grad()
                _, pred, label = self.one_iter(data, augmentation=False)
                all_pred.extend(pred.data.cpu().numpy().tolist())
                all_label.extend(label.data.cpu().numpy())
            self.get_result(all_pred, all_label)
            self.print_loss_result('val', prefix="  Evaluation", length=len(data_loader), is_train=is_train)

        if is_train:
            t = 'ap' if 'ap' in self.result_types else 'bal_acc'
            if self.result_dict[t] > self.best:
                self.best = self.result_dict[t]
                self.save_networks(save_path=os.path.join(self.ckp_folder, 'best_' + t + '.pth'))
        self.set_train()
        return self.result_dict