def ResNet34(num_classes, pretrained=None):
"""
ResNet34 for classification
Args:
num_classes (int): the number of classes
criterion: loss function
pretrained (str | None):
:return: model
"""
model = ResNet_CLS(num_classes, depth=34)
if pretrained is not None:
model_dict = model.state_dict()
pretrained_dict = torch.load(pretrained)
matched_dict, unmatched = validate_ckpt(model_dict, pretrained_dict)
if len(unmatched) > 0:
print('unmatched key ({}): '.format(len(unmatched)))
for i in range(len(unmatched)):
unmatched_key = unmatched[i]
if unmatched_key in model_dict.keys():
print(
f'model_dict[{unmatched_key}].shape={model_dict[unmatched_key].shape} vs '
f'pretrained[{unmatched_key}].shape={pretrained_dict[unmatched_key].shape}'
)
else:
print(f'key \'{unmatched_key}\' is not in model_dict')
model_dict.update(matched_dict)
model.load_state_dict(model_dict)
return model
def ResNeXt101_32x8d(num_classes, pretrained=None):
model = ResNeXt_CLS(num_classes, depth=101, groups=32, width_per_group=8)
if pretrained is not None:
model_dict = model.state_dict()
pretrained_dict = torch.load(pretrained)
matched_dict, unmatched = validate_ckpt(model_dict, pretrained_dict)
if len(unmatched) > 0:
print('unmatched key ({}): '.format(len(unmatched)))
for i in range(len(unmatched)):
unmatched_key = unmatched[i]
if unmatched_key in model_dict.keys():
print(
f'model_dict[{unmatched_key}].shape={model_dict[unmatched_key].shape} vs '
f'pretrained[{unmatched_key}].shape={pretrained_dict[unmatched_key].shape}'
)
else:
print(f'key \'{unmatched_key}\' is not in model_dict')
model_dict.update(matched_dict)
model.load_state_dict(model_dict)
return model