def load_net(dataset, net_type, nb_classes):
if net_type == 'bn_inception':
model = net.bn_inception(pretrained=True)
model.last_linear = nn.Linear(1024, nb_classes)
model.load_state_dict(
torch.load('net/finetuned_' + dataset + '_' + net_type + '.pth'))
elif net_type == 'densenet121':
model = net.densenet121(pretrained=True)
model.classifier = nn.Linear(1024, nb_classes)
model.load_state_dict(
torch.load('net/finetuned_' + dataset + '_' + net_type + '.pth'))
elif net_type == 'densenet161':
model = net.densenet161(pretrained=True)
model.classifier = nn.Linear(2208, nb_classes)
model.load_state_dict(
torch.load('net/finetuned_' + dataset + '_' + net_type + '.pth'))
elif net_type == 'densenet169':
model = net.densenet169(pretrained=True)
model.classifier = nn.Linear(1664, nb_classes)
model.load_state_dict(
torch.load('net/finetuned_' + dataset + '_' + net_type + '.pth'))
elif net_type == 'densenet201':
model = net.densenet201(pretrained=True)
model.classifier = nn.Linear(1920, nb_classes)
model.load_state_dict(
torch.load('net/finetuned_' + dataset + '_' + net_type + '.pth'))
return model
# --------------------------------
# Load pre-trained model & weights
# --------------------------------
# Backbone Model
trained_dataset = 'inshop'
model_name = 'bn_inception'
length = 6400 # CUB: 2500 | Cars: 4900 | SOP: 8100 | In-shop: 6400
if model_name == 'googlenet':
model = googlenet(embedding_size=args.sz_embedding, pretrained=True, is_norm=args.l2_norm, bn_freeze = args.bn_freeze)
elif model_name == 'bn_inception':
model = bn_inception(embedding_size=args.sz_embedding, pretrained=True, is_norm=args.l2_norm, bn_freeze = args.bn_freeze)
elif model_name == 'resnet50':
model = Resnet50(embedding_size=args.sz_embedding, pretrained=True, is_norm=args.l2_norm, bn_freeze = args.bn_freeze)
model = model.cuda()
if args.gpu_id == -1:
model = nn.DataParallel(model)
# Load pre-trained weights
if trained_dataset == 'cub':
print("CUB!")
path = 'path/to/cub_weights.pth'
elif trained_dataset == 'cars':
print("Cars!")
path = 'path/to/cars_weights.pth'
num_workers=args.nb_workers,
drop_last=True,
pin_memory=True)
dl_ev = torch.utils.data.DataLoader(dataset.Data.factory(
type=type,
root=args.root_folder,
labels=list(range(args.nb_classes, 2 * args.nb_classes)),
is_extracted=args.is_extracted,
transform=dataset.utils.make_transform(is_train=False)),
batch_size=args.sz_batch,
shuffle=False,
num_workers=args.nb_workers,
pin_memory=True)
model = net.bn_inception(pretrained=True)
net.embed(model, sz_embedding=args.sz_embedding)
model = model.cuda()
criterion = proxynca.ProxyNCA(args.sz_embedding, args.nb_classes,
args.sz_batch).cuda()
opt = torch.optim.Adam(
[
{ # embedding parameters
'params': model.embedding_layer.parameters(),
'lr': args.lr_embedding
},
{ # proxy nca parameters
'params': criterion.parameters(),
'lr': args.lr_proxynca
def prepare_model(self, base_model_name, num_classes):
"""
base_model: string contains the model name
this function is used to modify the last layer (fully connected layer)
for a given architecture to suit our dataset number of actions.
"""
#add other architectures later
if 'resnet' in base_model_name:
#Load pretrained model
self.base_model = getattr(torchvision.models, base_model_name)(pretrained=True)
self.last_layer_name = 'fc'
#set the input size for the model
self.input_size = 224
self.input_mean = [0.485, 0.456, 0.406]
self.input_std = [0.229, 0.224, 0.225]
#There's no point of substarct means from RGBDiff frames
if self.modality == 'RGBDiff':
#[0.485, 0.456, 0.406 , 0, 0, 0, 0, 0,.....]
self.input_mean = [0.485, 0.456, 0.406] + [0] * 3 * self.new_length
#Expand the list with the average 0.452
self.input_std = self.input_std + [np.mean(self.input_std) * 2] * 3 * self.new_length
#BNInception doesn't exist in torchvision models, so we have to get it from net folder
elif base_model_name == 'BNInception':
import net
self.base_model = net.bn_inception(pretrained = True)
self.last_layer_name = 'last_linear'
self.input_size = 224
self.input_mean = [104, 117, 128]
self.input_std = [1]
#BNInception takes input in range of (0~255), so only mean subtraction should be used in preprocessing.
#This is different from ResNet models, which take input in the range of (0~1).
if self.modality == 'RGBDiff':
self.input_mean = self.input_mean * (1+self.new_length)
else:
raise ValueError('Unknown base model: {}'.format(base_model_name))
#Get the input size for the last layer of CNN
features_dim = getattr(self.base_model, self.last_layer_name).in_features
#In case of no dropout,Only nn.Linear will be added
if self.dropout == 0:
setattr(self.base_model, self.last_layer_name, nn.Linear(features_dim, num_classes))
self.new_fc = None
print('The modified linear layer is :', getattr(self.base_model, self.last_layer_name))
#In case of dropout, only nn.Dropout will be added and nn.Linear will be prepared to be added later
else:
setattr(self.base_model, self.last_layer_name, nn.Dropout(self.dropout))
self.new_fc = nn.Linear(features_dim, num_classes)
print('Dropout Layer added and The modified linear layer is :',self.new_fc)
#Modify Wighets of newly created Linear layer
std=0.001
if self.new_fc == None:
normal_(getattr(self.base_model, self.last_layer_name).weight,0,std)
constant_(getattr(self.base_model, self.last_layer_name).bias,0)
else:
normal_(self.new_fc.weight, 0, std)
constant_(self.new_fc.bias,0)
def setup_model(args=args):
model = net.bn_inception(pretrained=True)
net.embed(model, sz_embedding=args.sz_embedding)
model = model.cuda()
return model
parser.add_argument('--net_type',
default='densenet121',
type=str,
help='type of net we want to use')
parser.add_argument('--embed',
default=False,
type=bool,
help='Number of iterations we want to do for GTG')
args = parser.parse_args()
batch_size = 32
# torch.cuda.set_device(args.gpu_id)
if args.net_type == 'bn_inception':
model = net.bn_inception(pretrained=True, nb_classes=args.nb_classes)
model.last_linear = nn.Linear(1024, args.nb_classes)
if args.embed:
model = net.Inception_embed(model, 1024, args.sz_embedding,
args.nb_classes)
elif args.net_type == 'resnet18':
model = net.resnet18(pretrained=True)
model.fc = nn.Linear(512, args.nb_classes)
elif args.net_type == 'resnet34':
model = net.resnet34(pretrained=True)
model.fc = nn.Linear(512, args.nb_classes)
elif args.net_type == 'resnet50':
model = net.resnet50(pretrained=True)
model.fc = nn.Linear(2048, args.nb_classes)
elif args.net_type == 'resnet101':
)
dl_ev = torch.utils.data.DataLoader(
dataset.Birds(
root = args.cub_root,
labels = list(range(args.nb_classes, 2 * args.nb_classes)),
is_extracted = args.cub_is_extracted,
transform = dataset.utils.make_transform(is_train = False)
),
batch_size = args.sz_batch,
shuffle = False,
num_workers = args.nb_workers,
pin_memory = True
)
model = net.bn_inception(pretrained = True)
net.embed(model, sz_embedding=args.sz_embedding)
model = model.cuda()
criterion = proxynca.ProxyNCA(args.sz_embedding, args.nb_classes,
args.sz_batch).cuda()
opt = torch.optim.Adam(
[
{ # embedding parameters
'params': model.embedding_layer.parameters(),
'lr' : args.lr_embedding
},
{ # proxy nca parameters
'params': criterion.parameters(),
'lr': args.lr_proxynca