STCAE/train.py at master · spthermo/STCAE · 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
import torch
import torch.nn as nn
import datetime
import sys
import opt
import os
import metrics.segmentation_metrics
import metrics.reasoning_metrics

from models import get_model
from options import parse_arguments
from data import SOR3DLoader, SOR3DLoaderParams
from metrics.segmentation_metrics import F1_Score
from utils import NullVisualizer, VisdomVisualizer, initialize_vgg_weights, initialize_weights, generate_gt_heatmap


if __name__ == '__main__':
    print("{} | Torch Version: {}".format(datetime.datetime.now(), torch.__version__))
    args, uknown = parse_arguments(sys.argv)
    device = torch.device("cuda:0" if args.cuda else "cpu")
    torch.manual_seed(667)
    if device.type == 'cuda':
        torch.cuda.manual_seed(667)

    # visdom init
    visualizer = NullVisualizer() if args.visdom is None\
        else VisdomVisualizer(args.name, args.visdom, count=4)
    if args.visdom is None:
        args.visdom_iters = 0

    # data
    train_data_params = SOR3DLoaderParams(root_path = os.path.join(args.train_path, 'train'))
    train_data_iterator = SOR3DLoader(train_data_params)

    train_set = torch.utils.data.DataLoader(train_data_iterator,\
        batch_size = args.batch_size, shuffle=True,\
        num_workers = 0, pin_memory=False)

    val_data_params = SOR3DLoaderParams(root_path = os.path.join(args.train_path, 'val'))
    val_data_iterator = SOR3DLoader(val_data_params)

    val_set = torch.utils.data.DataLoader(val_data_iterator,\
        batch_size = 1, shuffle=False,\
        num_workers = 0, pin_memory=False)

    # create & init model
    model_params = {
        'dim': args.crop_size,
        'ndf': args.ndf,
        'affordance_classes': args.action_clasees,
        'ngroups': args.ngroups,
        'nchannels': 3
    }
    encoder_clstm, decoder = get_model(args.model, model_params)
    initialize_vgg_weights(encoder_clstm, args.weight_init)
    encoder_clstm_params = sum(p.numel() for p in encoder_clstm.parameters() if p.requires_grad)
    encoder_clstm.to(device)
    initialize_weights(decoder, args.weight_init)
    decoder_params = sum(p.numel() for p in decoder.parameters() if p.requires_grad)
    decoder.to(device)
    print ('Model params: ', encoder_clstm_params + decoder_params)

    # create and init optimizer
    opt_params = opt.OptimizerParameters(learning_rate=args.lr, momentum=args.momentum,\
        momentum2=args.momentum2, epsilon=args.epsilon)
    optimizer = opt.get_optimizer(args.optimizer, encoder_clstm.parameters(), opt_params)

    opt_params2 = opt.OptimizerParameters(learning_rate=args.lr, momentum=args.momentum,\
        momentum2=args.momentum2, epsilon=args.epsilon)
    optimizer2 = opt.get_optimizer(args.optimizer, decoder.parameters(), opt_params2)

    affordance_class_loss = nn.CrossEntropyLoss().to(device)

    affordance_seg_loss = nn.NLLLoss(reduction='mean').to(device) #ignore_index=-1
    kld_loss = nn.KLDivLoss(reduction='batchmean')

    f1_score = F1_Score()
    kld_score = nn.KLDivLoss(reduction='batchmean')

    # training loop
    target = torch.zeros(args.batch_size)
    iterations = 0
    for epoch in range(args.epochs):
        print("Training | Epoch: {}".format(epoch))
        encoder_clstm.train()
        decoder.train()
        total_data_num = len(train_set.dataset.fdata)
        for batch_id, batch in enumerate(train_set):
            if batch_id > ((total_data_num // args.batch_size) - 1):
                break
            optimizer.zero_grad()
            for in_batch_cnt in range(args.batch_size):
                target[in_batch_cnt] = batch["frame_001"]["action"][in_batch_cnt].item()
            sequence_action_loss = 0
            for frame in batch:
                frame_action_loss = 0
                attention_mask, out_list, action_pred = encoder_clstm.forward(batch[frame]["color"].to(device)) #, object_pred
                frame_action_loss = affordance_class_loss(action_pred.to(device), target.long().to(device))
                sequence_action_loss += frame_action_loss
            # predictions
            pred_mask, pred_heatmap = decoder.forward(out_list, attention_mask)
            # segmentation loss
            target_mask = batch[frame]["target"]
            seg_loss = affordance_seg_loss(pred_mask, target_mask.to(device))
            affordance_loss = sequence_action_loss / len(batch)
            # reasoning loss
            heat = batch[frame]["heatmap"]
            heat_processed = generate_gt_heatmap(heat, 5)
            reasoning_loss = kld_loss(pred_heatmap, heat_processed.to(device))
            # total loss
            total_loss =  0.1 * reasoning_loss + 0.3 * seg_loss + 0.6 * affordance_loss
            total_loss.backward()
            optimizer.step()
            optimizer2.step()
            iterations += args.batch_size
            print("Epoch: {}, iteration: {}, learning rate: {}, Total Loss: {}\n"\
                    .format(epoch, iterations, optimizer.param_groups[0]['lr'], total_loss.item()))

            #visualization
            if (iterations) % args.visdom_iters == 0:
                visualizer.show_seg_map(torch.exp(pred_heatmap[0]), 'heatmap prediction 1')
                visualizer.show_seg_map(heat_processed[0], 'heatmap gt 1')
                visualizer.show_seg_map(torch.exp(pred_heatmap[1]), 'heatmap prediction 2')
                visualizer.show_seg_map(heat_processed[1], 'heatmap gt 2')
                visualizer.show_seg_map(pred_mask[0].argmax(0), 'segmentation prediction 1') #[target[1].long()]
                visualizer.show_seg_map(target_mask[0], 'segmentation gt 1')
                visualizer.show_seg_map(pred_mask[1].argmax(0), 'segmentation prediction 2') #[target[1].long()]
                visualizer.show_seg_map(target_mask[1], 'segmentation gt 2')

            if (iterations + 1) % args.disp_iters == 0:
                visualizer.append_loss(epoch + 1, iterations, total_loss.item(), "total")
                visualizer.append_loss(epoch + 1, iterations, affordance_loss.item(), "affordance")
                visualizer.append_loss(epoch + 1, iterations, seg_loss.item(), "segmentation")
                visualizer.append_loss(epoch + 1, iterations, reasoning_loss.item(), "reasoning")

        print("Validation | Epoch: {}".format(epoch))
        encoder_clstm.eval()
        decoder.eval()
        total_jaccard =  0
        total_f1 = 0
        total_kld = 0
        total_data_num = len(val_set.dataset.fdata)
        for batch_id, batch in enumerate(val_set):
            if batch_id > ((total_data_num // args.batch_size) - 1):
                break
            target[0] = batch["frame_001"]["action"][0].item()
            for frame in batch:
                attention_mask, out_list, action_pred = encoder_clstm.forward(batch[frame]["color"].to(device)) #, object_pred
            # predictions
            target_mask = batch[frame]["target"]
            heat = batch[frame]["heatmap"]
            heat_processed = generate_gt_heatmap(heat, 5)
            pred_mask, pred_heatmap = decoder.forward(out_list, attention_mask)
            _, collapsed_mask = torch.max(torch.exp(pred_mask), 1)
            jaccard = metrics.segmentation_metrics.compute_jaccard(collapsed_mask.cpu().detach().numpy().reshape(-1), target_mask.cpu().detach().numpy().reshape(-1))

            binary_pred_mask = torch.exp(pred_mask.squeeze(0))
            binary_pred_mask[binary_pred_mask > 0.75] = 1
            binary_pred_mask[binary_pred_mask <= 0.75] = 0
            binary_target_mask = torch.zeros(model_params['affordance_classes'], target_mask.shape[1], target_mask.shape[2])
            for i in range(target_mask.shape[1]):
                for j in range(target_mask.shape[2]):
                    t_class = target_mask[0][i][j]
                    binary_target_mask[t_class][i][j] = 1
            f1 = f1_score(binary_pred_mask.detach().cpu(), binary_target_mask)
            kld = kld_score(pred_heatmap.cpu(), heat_processed)

            total_jaccard += jaccard
            total_f1 += f1
            total_kld += kld

        print("Epoch: {}, IoU: {}, F1: {}, KLD: {}\n"\
                    .format(epoch, total_jaccard / total_data_num, total_f1 / total_data_num, total_kld / total_data_num))

        #save model params
        if epoch in args.save_scheduler:
            opt.save_checkpoint({
                'epoch': epoch,
                'batch_size': args.batch_size,
                'task': args.model,
                'state_dict_en': encoder_clstm.state_dict(),
                'state_dict_de': decoder.state_dict(),
                'optimizer_en': optimizer.state_dict(),
                'optimizer_de': optimizer.state_dict(),
            }, epoch)