def APGeM(self, checkpoint):
GeM = nets.create_model("resnet101_rmac",
pretrained=checkpoint,
without_fc=False)
GeM = GeM.to(self.device)
GeM.eval()
return GeM
def load_model(path, iscuda):
checkpoint = common.load_checkpoint(path, iscuda)
net = nets.create_model(pretrained="", **checkpoint['model_options'])
net = common.switch_model_to_cuda(net, iscuda, checkpoint)
net.load_state_dict(checkpoint['state_dict'])
net.preprocess = checkpoint.get('preprocess', net.preprocess)
if 'pca' in checkpoint: net.pca = checkpoint.get('pca')
return net
def load_model(path, iscuda, whiten=None):
checkpoint = common.load_checkpoint(path, iscuda)
net = nets.create_model(pretrained="", **checkpoint['model_options'])
net = common.switch_model_to_cuda(net, iscuda, checkpoint)
net.load_state_dict(checkpoint['state_dict'])
net.preprocess = checkpoint.get('preprocess', net.preprocess)
if whiten is not None and 'pca' in checkpoint:
if whiten in checkpoint['pca']:
net.pca = checkpoint['pca'][whiten]
return net
def load_model(path=None, iscuda=''):
checkpoint = common.load_checkpoint('./pretrained_model/Resnet50-AP-GeM.pt', iscuda)
net = nets.create_model(pretrained="", **model_options)#**checkpoint['model_options']
net = common.switch_model_to_cuda(net, iscuda, checkpoint)
if path:
checkpoint_2 = common.load_checkpoint(path, iscuda)
checkpoint_state_dict = checkpoint_2['state_dict']
start_epoch = checkpoint_2['epoch']
else:
checkpoint_state_dict = checkpoint['state_dict']
start_epoch = 0
# net.load_state_dict(checkpoint_state_dict,False)
load_param(net,checkpoint_state_dict)
net.preprocess = checkpoint.get('preprocess', net.preprocess)
if 'pca' in checkpoint:
net.pca = checkpoint.get('pca')
return net,start_epoch
args.iscuda = common.torch_set_gpu(args.gpu)
train_set = datasets.create('Landmarks_clean')
val_set = datasets.create('RParis6K')
model_options = {
'arch': args.arch,
'out_dim': args.out_dim,
'pooling': args.pooling,
'gemp': args.gemp
}
start_epoch = 0
if os.path.isfile(args.resume):
checkpoint = common.load_checkpoint(args.resume, args.iscuda)
net = nets.create_model(pretrained='', **model_options)
net = common.switch_model_to_cuda(net, args.iscuda, checkpoint)
net.load_state_dict(checkpoint['state_dict'])
start_epoch = int(
os.path.splitext(os.path.basename(args.resume))[0].split('_')[-1])
else:
net = nets.create_model(pretrained='imagenet', **model_options)
net = common.switch_model_to_cuda(net, iscuda=True)
optimizer = torch.optim.Adam(net.parameters(),
lr=args.lr,
weight_decay=1e-6)
criterion = APLoss(20, -1, 1)
best_map = 0
for epoch in range(start_epoch, 300):
