RAFNet/process.py at main · NihilDigit/RAFNet · 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
'''
 @FileName    : process.py
 @EditTime    : 2022-09-27 16:18:51
 @Author      : Buzhen Huang
 @Email       : hbz@seu.edu.cn
 @Description :
'''

import torch
import numpy as np
import cv2
from tqdm import tqdm
import time

def extract_valid(data):
    batch_size, agent_num, d = data['keypoints'].shape[:3]
    valid = data['valid'].reshape(-1,)

    data['center'] = data['center'] #.reshape(batch_size*agent_num, -1)[valid == 1]
    data['scale'] = data['scale'] #.reshape(batch_size*agent_num,)[valid == 1]
    data['img_h'] = data['img_h'] #.reshape(batch_size*agent_num,)[valid == 1]
    data['img_w'] = data['img_w'] #.reshape(batch_size*agent_num,)[valid == 1]
    data['focal_length'] = data['focal_length'] #.reshape(batch_size*agent_num,)[valid == 1]

    data['valid_img_h'] = data['img_h'].reshape(batch_size*agent_num,)[valid == 1]
    data['valid_img_w'] = data['img_w'].reshape(batch_size*agent_num,)[valid == 1]
    data['valid_focal_length'] = data['focal_length'].reshape(batch_size*agent_num,)[valid == 1]
    data['has_3d'] = data['has_3d'].reshape(batch_size*agent_num,1)[valid == 1]
    data['has_smpl'] = data['has_smpl'].reshape(batch_size*agent_num,1)[valid == 1]
    data['verts'] = data['verts'].reshape(batch_size*agent_num, 6890, 3)[valid == 1]
    data['gt_joints'] = data['gt_joints'].reshape(batch_size*agent_num, -1, 4)[valid == 1]
    data['pose'] = data['pose'].reshape(batch_size*agent_num, 72)[valid == 1]
    data['betas'] = data['betas'].reshape(batch_size*agent_num, 10)[valid == 1]
    data['keypoints'] = data['keypoints'].reshape(batch_size*agent_num, 26, 3)[valid == 1]
    data['gt_cam_t'] = data['gt_cam_t'].reshape(batch_size*agent_num, 3)[valid == 1]

    data['ori_imgname'] = data['imgname']
    imgname = (np.array(data['imgname']).T).reshape(batch_size*agent_num,)[valid.detach().cpu().numpy() == 1]
    data['imgname'] = imgname.tolist()

    return data

def extract_valid_demo(data):
    batch_size, agent_num, _, _, _ = data['img'].shape
    valid = data['valid'].reshape(-1,)

    data['center'] = data['center']
    data['scale'] = data['scale']
    data['img_h'] = data['img_h']
    data['img_w'] = data['img_w']
    data['focal_length'] = data['focal_length']

    data['valid_focal_length'] = data['focal_length'].reshape(batch_size*agent_num,)[valid == 1]

    return data

def to_device(data, device):
    imnames = {'imgname':data['imgname']}
    data = {k:v.to(device).float() for k, v in data.items() if k not in ['imgname']}
    data = {**imnames, **data}

    return data

def relation_train(model, loss_func, train_loader, epoch, num_epoch, device=torch.device('cpu')):

    print('-' * 10 + 'model training' + '-' * 10)
    len_data = len(train_loader)
    model.model.train(mode=True)
    if model.scheduler is not None:
        model.scheduler.step()

    train_loss = 0.
    for i, data in enumerate(train_loader):
        data = to_device(data, device)
        data = extract_valid(data)

        # forward
        pred = model.model(data)

        # calculate loss
        loss, cur_loss_dict = loss_func.calcul_trainloss(pred, data)

        # backward
        model.optimizer.zero_grad()
        loss.backward()

        # optimize
        model.optimizer.step()
        if model.scheduler is not None:
            model.scheduler.batch_step()

        loss_batch = loss.detach() #/ batchsize
        print('epoch: %d/%d, batch: %d/%d, loss: %.6f' %(epoch, num_epoch, i, len_data, loss_batch), cur_loss_dict)
        train_loss += loss_batch

    return train_loss/len_data

def relation_test(model, loss_func, loader, device=torch.device('cpu')):

    print('-' * 10 + 'model testing' + '-' * 10)
    loss_all = 0.
    model.model.eval()
    with torch.no_grad():
        for i, data in enumerate(loader):
            batchsize = data['keypoints'].shape[0]
            data = to_device(data, device)
            data = extract_valid(data)

            # forward
            pred = model.model(data)

            # calculate loss
            loss, cur_loss_dict = loss_func.calcul_testloss(pred, data)

            if False:
                results = {}
                results.update(imgs=data['imgname'])
                results.update(pred_trans=pred['pred_cam_t'].detach().cpu().numpy().astype(np.float32))
                results.update(pred_pose=pred['pred_pose'].detach().cpu().numpy().astype(np.float32))
                results.update(pred_shape=pred['pred_shape'].detach().cpu().numpy().astype(np.float32))
                results.update(img_h=data['valid_img_h'].detach().cpu().numpy().astype(np.float32))
                results.update(img_w=data['valid_img_w'].detach().cpu().numpy().astype(np.float32))
                model.save_params(results, i, batchsize)


            if i < 1:
                results = {}
                results.update(imgs=data['ori_imgname'])
                results.update(pred_trans=pred['pred_cam_t'].detach().cpu().numpy().astype(np.float32))
                results.update(gt_trans=data['gt_cam_t'].detach().cpu().numpy().astype(np.float32))
                results.update(focal_length=data['valid_focal_length'].detach().cpu().numpy().astype(np.float32))
                results.update(valid=data['valid'].detach().cpu().numpy().astype(np.float32))

                if 'MPJPE_instance' in cur_loss_dict.keys() or 'MPJPE_H36M_instance' in cur_loss_dict.keys():
                    results.update(MPJPE=loss.detach().cpu().numpy().astype(np.float32))

                if 'pred_verts' not in pred.keys():
                    results.update(pred_joints=pred['pred_joints'].detach().cpu().numpy().astype(np.float32))
                    results.update(gt_joints=data['gt_joints'].detach().cpu().numpy().astype(np.float32))
                    model.save_joint_results(results, i, batchsize)
                else:
                    results.update(pred_verts=pred['pred_verts'].detach().cpu().numpy().astype(np.float32))
                    results.update(gt_verts=data['verts'].detach().cpu().numpy().astype(np.float32))
                    model.save_test_results(results, i, batchsize)

            loss_batch = loss.detach().mean() #/ batchsize
            print('batch: %d/%d, loss: %.6f ' %(i, len(loader), loss_batch), cur_loss_dict)
            loss_all += loss_batch
        loss_all = loss_all / len(loader)
        return loss_all

def relation_demo(model, loader, device=torch.device('cpu')):

    print('-' * 10 + 'model demo' + '-' * 10)
    model.model.eval()
    with torch.no_grad():
        for i, data in tqdm(enumerate(loader), total=len(loader)):
            batchsize = data['img'].shape[0]
            data = to_device(data, device)
            data = extract_valid_demo(data)

            # forward
            pred = model.model(data)

            results = {}
            results.update(imgs=data['imgname'])
            results.update(pred_trans=pred['pred_cam_t'].detach().cpu().numpy().astype(np.float32))
            results.update(focal_length=data['valid_focal_length'].detach().cpu().numpy().astype(np.float32))
            if 'pred_verts' not in pred.keys():
                results.update(pred_joints=pred['pred_joints'].detach().cpu().numpy().astype(np.float32))
                model.save_demo_joint_results(results, i, batchsize)
            else:
                results.update(pred_verts=pred['pred_verts'].detach().cpu().numpy().astype(np.float32))
                model.save_demo_results(results, i, batchsize)