def get_resnet(type, load_ckpt=True):
"""
loading ResNet of trained on places365 as basenet
"""
if type == 18:
model_file = 'Place365/wideresnet18_places365.pth.tar'
model_res_net = wideresnet.resnet18(num_classes=365)
elif type == 34:
model_file = 'Place365/resnet34-333f7ec4.pth' # Pre-trained on ImageNet
model_res_net = wideresnet.resnet34()
elif type == 50:
model_file = 'Place365/resnet50_places365.pth.tar'
model_res_net = wideresnet.resnet50(num_classes=365)
else:
raise Exception('Unknown ResNet Type')
if load_ckpt:
try:
checkpoint = torch.load(model_file,
map_location=lambda storage, loc: storage)
except IOError:
print('Fail to load checkpoint, because no such file:', model_file)
else:
if type == 34:
model_res_net.load_state_dict(checkpoint)
else:
state_dict = {
str.replace(k, 'module.', ''): v
for k, v in checkpoint['state_dict'].items()
}
model_res_net.load_state_dict(state_dict)
return model_res_net
def baseResNet(type=50):
from Place365 import wideresnet
if type == 18:
model_res_net = wideresnet.resnet18(num_classes=365)
elif type == 34:
model_res_net = wideresnet.resnet34()
elif type == 50:
model_res_net = wideresnet.resnet50(num_classes=365)
else:
raise Exception('Unknown ResNet Type')
layers = list(model_res_net.children())[:-2] # children()只包括了第一代儿子模块，get rid of the last two layers: avepool & fc
return layers
def get_resnet(type):
"""
loading ResNet-18/34/50 of trained on places365 (ImageNet) as basenet
"""
from Place365 import wideresnet
if type == 18:
model_file = 'Place365/wideresnet18_places365.pth.tar'
model_res_net = wideresnet.resnet18(num_classes=365)
elif type == 34:
model_file = 'Place365/resnet34-333f7ec4.pth' # Pre-trained on ImageNet
model_res_net = wideresnet.resnet34()
elif type == 50:
model_file = 'Place365/resnet50_places365.pth.tar'
model_res_net = wideresnet.resnet50(num_classes=365)
else:
raise Exception('Unknown ResNet Type')
return model_res_net, model_file
def loadSceneRecognitionModel(trainedNetVLADLayers):
# this model has a last conv feature map as 14x14
model_file = 'Place365/wideresnet18_places365.pth.tar'
model = wideresnet.resnet18(num_classes=365)
state_dict = getPretrainedParams(model_file)
model.load_state_dict(state_dict)
model.eval()
# hook the feature extractor
features_names = ['layer4', 'avgpool'] # this is the last conv layer of the resnet
for name in features_names:
model._modules.get(name).register_forward_hook(hook_feature)
# hook the layer for DataSet
features_candidates = ['layer4', 'layer3', 'layer2', 'layer1', 'relu1']
model._modules.get(features_candidates[trainedNetVLADLayers]).register_forward_hook(hook_layer)
return model
encoder_dim = 512
encoder = models.vgg16(pretrained=pretrained)
# capture only feature part and remove last relu and maxpool
layers = list(encoder.features.children())[:-2]
if pretrained:
# if using pretrained then only train conv5_1, conv5_2, and conv5_3
for l in layers[:-5]:
for p in l.parameters():
p.requires_grad = False
elif opt.arch.lower() == 'resnet18':
encoder_dim = 512
# loading resnet18 of trained on places365 as basenet
from Place365 import wideresnet
model_file = 'Place365/wideresnet18_places365.pth.tar'
modelResNet = wideresnet.resnet18(num_classes=365)
checkpoint = torch.load(model_file,
map_location=lambda storage, loc: storage)
state_dict = {
str.replace(k, 'module.', ''): v
for k, v in checkpoint['state_dict'].items()
}
modelResNet.load_state_dict(state_dict)
# get rid of the last two layers: avepool & fc
layers = list(modelResNet.children())[:-2]
# 让最后1\2个block参与netVLAD训练
for l in layers[:-2]:
for p in l.parameters():
p.requires_grad = False
if opt.mode.lower() == 'cluster':