def __init__(self):
super(coarse_extractor,self).__init__()
self.model_fc = model_no.Resnet50Fc()
self.bottleneck = nn.Linear(2048, 256)
self.bottleneck.weight.data.normal_(0, 0.005)
self.bottleneck.bias.data.fill_(0.1)
self.bottleneck_layer = nn.Sequential(self.bottleneck, nn.ReLU(), nn.Dropout(0.5))
def __init__(self, feature_len):
super(fine_net,self).__init__()
self.model_fc = model_no.Resnet50Fc()
self.bottleneck_0 = nn.Linear(feature_len, 256)
self.bottleneck_0.weight.data.normal_(0, 0.005)
self.bottleneck_0.bias.data.fill_(0.1)
self.bottleneck_layer = nn.Sequential(self.bottleneck_0, nn.ReLU(), nn.Dropout(0.5))
self.classifier_layer = predictor(256, cate_all[0])
def __init__(self):
super(coarse_net, self).__init__()
self.model_fc = model_no.Resnet50Fc()
self.bottleneck = nn.Linear(2048, 256)
self.bottleneck.weight.data.normal_(0, 0.005)
self.bottleneck.bias.data.fill_(0.1)
self.bottleneck_layer = nn.Sequential(self.bottleneck, nn.ReLU(),
nn.Dropout(0.5))
self.classifier_layer_1 = nn.Linear(256, cate_all[1])
self.classifier_layer_1.weight.data.normal_(0, 0.01)
self.classifier_layer_1.bias.data.fill_(0.0)
self.classifier_layer_2 = nn.Linear(256, cate_all[2])
self.classifier_layer_2.weight.data.normal_(0, 0.01)
self.classifier_layer_2.bias.data.fill_(0.0)
self.classifier_layer_3 = nn.Linear(256, cate_all[3])
self.classifier_layer_3.weight.data.normal_(0, 0.01)
self.classifier_layer_3.bias.data.fill_(0.0)
batch_size=4,
shuffle=False,
num_workers=args.num_workers)
prep_dict_test = prep.image_test_10crop(resize_size=256, crop_size=224)
for i in range(10):
dataset_list = ImageList(open(dataset_test).readlines(),
transform=prep_dict_test["val" + str(i)])
dataset_loaders["val" + str(i)] = torch.utils.data.DataLoader(
dataset_list, batch_size=4, shuffle=False, num_workers=6)
# network construction
feature_len = 2048
# fine-grained feature extractor + fine-grained label predictor
devices = list(range(torch.cuda.device_count()))
model_fc = model_no.Resnet50Fc()
model_fc = model_fc.to(device)
model_fc = nn.DataParallel(model_fc, device_ids=devices)
my_fine_net = fine_net(feature_len)
my_fine_net = my_fine_net.to(device)
my_fine_net.train(True)
model_fc.train(True)
# criterion and optimizer
criterion = {
"classifier": nn.CrossEntropyLoss(),
"kl_loss": nn.KLDivLoss(size_average=False),
"adversarial": nn.BCELoss()
}
optimizer_dict = [{