def resnet152(pretrained=False, **kwargs):
"""Constructs a ResNet-152 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
if pretrained:
print("resnet_low_level: Loading image net weights...")
utils_train.transfer_partial_weights(model_zoo.load_url(model_urls['resnet152']), model)
print("resnet_low_level: Done loading image net weights...")
#model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
return model
def resnet101(pretrained=False, **kwargs):
"""Constructs a ResNet-101 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNetTwoStream(Bottleneck, [3, 4, 23, 1], **kwargs)
if pretrained:
print("Loading image net weights...")
utils_train.transfer_partial_weights(
model_zoo.load_url(model_urls['resnet101']), model)
print("Done image net weights...")
# model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
return model
def load_network(config_dict):
output_types= config_dict['output_types']
use_billinear_upsampling = 'upsampling_bilinear' in config_dict.keys() and config_dict['upsampling_bilinear']
lower_billinear = 'upsampling_bilinear' in config_dict.keys() and config_dict['upsampling_bilinear'] == 'half'
upper_billinear = 'upsampling_bilinear' in config_dict.keys() and config_dict['upsampling_bilinear'] == 'upper'
from_latent_hidden_layers = config_dict.get('from_latent_hidden_layers', 0)
num_encoding_layers = config_dict.get('num_encoding_layers', 4)
num_cameras = 4
if config_dict['active_cameras']: # for H36M it is set to False
num_cameras = len(config_dict['active_cameras'])
if lower_billinear:
use_billinear_upsampling = False
network_single = unet_encode3D.unet(dimension_bg=config_dict['latent_bg'],
dimension_fg=config_dict['latent_fg'],
dimension_3d=config_dict['latent_3d'],
feature_scale=config_dict['feature_scale'],
shuffle_fg=config_dict['shuffle_fg'],
shuffle_3d=config_dict['shuffle_3d'],
latent_dropout=config_dict['latent_dropout'],
in_resolution=config_dict['inputDimension'],
encoderType=config_dict['encoderType'],
is_deconv=not use_billinear_upsampling,
upper_billinear=upper_billinear,
lower_billinear=lower_billinear,
from_latent_hidden_layers=from_latent_hidden_layers,
n_hidden_to3Dpose=config_dict['n_hidden_to3Dpose'],
num_encoding_layers=num_encoding_layers,
output_types=output_types,
subbatch_size=config_dict['useCamBatches'],
implicit_rotation=config_dict['implicit_rotation'],
skip_background=config_dict['skip_background'],
num_cameras=num_cameras,
)
if 'pretrained_network_path' in config_dict.keys(): # automatic
if config_dict['pretrained_network_path'] == 'MPII2Dpose':
pretrained_network_path = '/cvlabdata1/home/rhodin/code/humanposeannotation/output_save/CVPR18_H36M/TransferLearning2DNetwork/h36m_23d_crop_relative_s1_s5_aug_from2D_2017-08-22_15-52_3d_resnet/models/network_000000.pth'
print("Loading weights from MPII2Dpose")
pretrained_states = torch.load(pretrained_network_path)
utils_train.transfer_partial_weights(pretrained_states, network_single, submodule=0, add_prefix='encoder.') # last argument is to remove "network.single" prefix in saved network
else:
print("Loading weights from config_dict['pretrained_network_path']")
pretrained_network_path = config_dict['pretrained_network_path']
pretrained_states = torch.load(pretrained_network_path)
utils_train.transfer_partial_weights(pretrained_states, network_single, submodule=0) # last argument is to remove "network.single" prefix in saved network
print("Done loading weights from config_dict['pretrained_network_path']")
if 'pretrained_posenet_network_path' in config_dict.keys(): # automatic
print("Loading weights from config_dict['pretrained_posenet_network_path']")
pretrained_network_path = config_dict['pretrained_posenet_network_path']
pretrained_states = torch.load(pretrained_network_path)
utils_train.transfer_partial_weights(pretrained_states, network_single.to_pose, submodule=0) # last argument is to remove "network.single" prefix in saved network
print("Done loading weights from config_dict['pretrained_posenet_network_path']")
return network_single
def load_network(self, config_dict):
output_types = config_dict['output_types']
use_billinear_upsampling = config_dict.get('upsampling_bilinear',
False)
lower_billinear = 'upsampling_bilinear' in config_dict.keys(
) and config_dict['upsampling_bilinear'] == 'half'
upper_billinear = 'upsampling_bilinear' in config_dict.keys(
) and config_dict['upsampling_bilinear'] == 'upper'
if lower_billinear:
use_billinear_upsampling = False
network_single = detect_encode_decode.detect_encode_decode(
dimension_bg=config_dict['latent_bg'],
dimension_fg=config_dict['latent_fg'],
dimension_3d=config_dict['latent_3d'],
feature_scale=config_dict['feature_scale'],
shuffle_fg=config_dict['shuffle_fg'],
shuffle_3d=config_dict['shuffle_3d'],
latent_dropout=config_dict['latent_dropout'],
in_resolution=config_dict['inputDimension'],
is_deconv=not use_billinear_upsampling,
upper_billinear=upper_billinear,
lower_billinear=lower_billinear,
num_encoding_layers=config_dict.get('num_encoding_layers', 4),
output_types=output_types,
subbatch_size=config_dict['useCamBatches'],
implicit_rotation=config_dict['implicit_rotation'],
mode=config_dict['training_mode'],
spatial_transformer=config_dict.get('spatial_transformer', False),
ST_size=config_dict.get('spatial_transformer_num', 1),
spatial_transformer_bounds=config_dict.get(
'spatial_transformer_bounds', {
'border_factor': 1,
'min_size': 0.1,
'max_size': 1
}),
masked_blending=config_dict.get('masked_blending', True),
scale_mask_max_to_1=config_dict.get('scale_mask_max_to_1', True),
predict_transformer_depth=config_dict.get(
'predict_transformer_depth', False),
normalize_mask_density=config_dict.get('normalize_mask_density',
False),
match_crops=config_dict.get('match_crops', False),
offset_crop=config_dict.get('offset_crop', False),
similarity_bandwidth=config_dict.get('similarity_bandwidth', 10),
disable_detector=config_dict.get('disable_detector', False),
)
if 'pretrained_network_path' in config_dict.keys(): # automatic
print(
"Loading weights from config_dict['pretrained_network_path']")
pretrained_network_path = config_dict['pretrained_network_path']
pretrained_states = torch.load(pretrained_network_path,
map_location=device)
utils_train.transfer_partial_weights(
pretrained_states, network_single, submodule=0
) # last argument is to remove "network.single" prefix in saved network
print(
"Done loading weights from config_dict['pretrained_network_path']"
)
if 'pretrained_detector_path' in config_dict.keys(): # automatic
print(
"Loading weights from config_dict['pretrained_detector_path']")
pretrained_network_path = config_dict['pretrained_detector_path']
pretrained_states = torch.load(pretrained_network_path,
map_location=device)
utils_train.transfer_partial_weights(
pretrained_states,
network_single,
submodule=0,
prefix='detector'
) # last argument is to remove "network.single" prefix in saved network
print(
"Done loading weights from config_dict['pretrained_detector_path']"
)
return network_single
