-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathresnet50_hmr_h36m.py
186 lines (180 loc) · 6.23 KB
/
resnet50_hmr_h36m.py
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
_base_ = ['../_base_/default_runtime.py']
use_adversarial_train = True
# evaluate
evaluation = dict(interval=1, metric='joint_error')
# optimizer
optimizer = dict(
backbone=dict(type='Adam', lr=2.5e-4),
head=dict(type='Adam', lr=2.5e-4),
disc=dict(type='Adam', lr=1e-4))
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(policy='Fixed', by_epoch=False)
runner = dict(type='EpochBasedRunner', max_epochs=100)
log_config = dict(
interval=50,
hooks=[dict(type='TextLoggerHook'),
dict(type='TensorboardLoggerHook')])
img_res = 224
# model settings
model = dict(
type='ImageBodyModelEstimator',
backbone=dict(
type='ResNet',
depth=50,
out_indices=[3],
norm_eval=False,
# norm_cfg=dict(type='SyncBN', requires_grad=True),
init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
head=dict(
type='HMRHead',
feat_dim=2048,
smpl_mean_params='data/body_models/smpl_mean_params.npz'),
body_model_train=dict(
type='SMPL',
keypoint_src='smpl_54',
keypoint_dst='smpl_54',
model_path='data/body_models/smpl',
keypoint_approximate=True,
extra_joints_regressor='data/body_models/J_regressor_extra.npy'),
body_model_test=dict(
type='SMPL',
keypoint_src='h36m',
keypoint_dst='h36m',
model_path='data/body_models/smpl',
joints_regressor='data/body_models/J_regressor_h36m.npy'),
convention='smpl_54',
loss_keypoints3d=dict(type='SmoothL1Loss', loss_weight=100),
loss_keypoints2d=dict(type='SmoothL1Loss', loss_weight=10),
loss_vertex=dict(type='L1Loss', loss_weight=2),
loss_smpl_pose=dict(type='MSELoss', loss_weight=3),
loss_smpl_betas=dict(type='MSELoss', loss_weight=0.02),
loss_adv=dict(
type='GANLoss',
gan_type='lsgan',
real_label_val=1.0,
fake_label_val=0.0,
loss_weight=1),
disc=dict(type='SMPLDiscriminator'))
# dataset settings
dataset_type = 'HumanImageDataset'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
data_keys = [
'has_smpl', 'smpl_body_pose', 'smpl_global_orient', 'smpl_betas',
'smpl_transl', 'keypoints2d', 'keypoints3d', 'sample_idx'
]
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='RandomChannelNoise', noise_factor=0.4),
dict(type='RandomHorizontalFlip', flip_prob=0.5, convention='smpl_54'),
dict(type='GetRandomScaleRotation', rot_factor=30, scale_factor=0.25),
dict(type='RandomOcclusion', occlusion_prob=0.9),
dict(type='MeshAffine', img_res=224),
dict(type='Normalize', **img_norm_cfg),
dict(type='ImageToTensor', keys=['img']),
dict(type='ToTensor', keys=data_keys),
dict(
type='Collect',
keys=['img', *data_keys],
meta_keys=['image_path', 'center', 'scale', 'rotation'])
]
adv_data_keys = [
'smpl_body_pose', 'smpl_global_orient', 'smpl_betas', 'smpl_transl'
]
train_adv_pipeline = [dict(type='Collect', keys=adv_data_keys, meta_keys=[])]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='GetRandomScaleRotation', rot_factor=0, scale_factor=0),
dict(type='MeshAffine', img_res=224),
dict(type='Normalize', **img_norm_cfg),
dict(type='ImageToTensor', keys=['img']),
dict(type='ToTensor', keys=data_keys),
dict(
type='Collect',
keys=['img', *data_keys],
meta_keys=['image_path', 'center', 'scale', 'rotation'])
]
inference_pipeline = [
dict(type='MeshAffine', img_res=224),
dict(type='Normalize', **img_norm_cfg),
dict(type='ImageToTensor', keys=['img']),
dict(
type='Collect',
keys=['img', 'sample_idx'],
meta_keys=['image_path', 'center', 'scale', 'rotation'])
]
data = dict(
samples_per_gpu=32,
workers_per_gpu=1,
train=dict(
type='AdversarialDataset',
train_dataset=dict(
type='MixedDataset',
configs=[
dict(
type=dataset_type,
dataset_name='h36m',
data_prefix='data',
pipeline=train_pipeline,
convention='smpl_54',
ann_file='h36m_mosh_train.npz'),
# ann_file='h36m_valid_protocol2.npz'),
],
partition=[1.0],
),
adv_dataset=dict(
type='MeshDataset',
dataset_name='cmu_mosh',
data_prefix='data',
pipeline=train_adv_pipeline,
ann_file='cmu_mosh.npz')),
# val=dict(
# type=dataset_type,
# body_model=dict(
# type='GenderedSMPL',
# keypoint_src='h36m',
# keypoint_dst='h36m',
# model_path='data/body_models/smpl',
# joints_regressor='data/body_models/J_regressor_h36m.npy'),
# dataset_name='pw3d',
# data_prefix='data',
# pipeline=test_pipeline,
# ann_file='pw3d_test.npz'),
# test=dict(
# type=dataset_type,
# body_model=dict(
# type='GenderedSMPL',
# keypoint_src='h36m',
# keypoint_dst='h36m',
# model_path='data/body_models/smpl',
# joints_regressor='data/body_models/J_regressor_h36m.npy'),
# dataset_name='pw3d',
# data_prefix='data',
# pipeline=test_pipeline,
# ann_file='pw3d_test.npz'),
val=dict(
type=dataset_type,
body_model=dict(
type='GenderedSMPL',
keypoint_src='h36m',
keypoint_dst='h36m',
model_path='data/body_models/smpl',
joints_regressor='data/body_models/J_regressor_h36m.npy'),
dataset_name='h36m',
data_prefix='data',
pipeline=test_pipeline,
ann_file='h36m_valid2_sample.npz'),
test=dict(
type=dataset_type,
body_model=dict(
type='GenderedSMPL',
keypoint_src='h36m',
keypoint_dst='h36m',
model_path='data/body_models/smpl',
joints_regressor='data/body_models/J_regressor_h36m.npy'),
dataset_name='h36m',
data_prefix='data',
pipeline=test_pipeline,
ann_file='data/preprocessed_datasets/h36m_valid_protocol2.npz'),
)