def get_labels():
batch_size = 4
test_transform = baseline.test_transform
testset = per_class_dataset(csv_file=r'/valid.csv',
root_dir=present_dir_path,
transform=test_transform,
RGB=True)
testloader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=1)
all_labels = np.array([])
for data in testloader:
images, labels, weights = data
all_labels = np.append(all_labels, labels.numpy())
return all_labels
def make_Difficulty():
load_model = True
batch_size = 8
if load_model == True:
load_path = "./save/densenet169/mura3cropfocal_1_acc.pkl"
train_transform = transforms.Compose([
transforms.Resize((320, 320)),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
net = torchvision.models.densenet169(pretrained=False)
net.classifier = nn.Linear(1664, 1)
net.load_state_dict(torch.load(load_path))
if torch.cuda.is_available():
net = net.cuda()
present_dir_path = os.getcwd()
trainset = per_class_dataset(csv_file=r'/train.csv',
root_dir=present_dir_path,
transform=train_transform)
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=False,
num_workers=4)
tot_loss = np.array([])
net.eval()
print("Start to evaluate")
for (i, data) in enumerate(trainloader, 0):
print("step %d" % i)
images, labels, weights = data
images = images.cuda()
labels = labels.cuda()
weights = weights.cuda().float()
outputs = torch.sigmoid(net(images))
outputs = outputs.select(1, 0)
outputs = torch.clamp(outputs, min=1e-7, max=1 - 1e-7)
loss = -(labels.float() * outputs.log() + (1 - labels.float()) *
(1 - outputs).log())
loss *= weights
tot_loss = np.append(tot_loss, loss.cpu().detach().numpy())
Difficulty = np.argsort(tot_loss)
np.save("./result/easy_to_hard.npy", Difficulty)
def get_from_densenet201(load_path):
batch_size = 4
test_transform = densenet201.test_transform
testset = per_class_dataset(csv_file=r'/valid.csv',
root_dir=present_dir_path,
transform=test_transform,
RGB=True)
testloader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
shuffle=False,
num_workers=1)
net = torchvision.models.densenet201(pretrained=False)
net.classifier = nn.Linear(1920, 1)
net.load_state_dict(torch.load(load_path))
# if torch.cuda.is_available():
# net = net.cuda()
predicts = get_predicts(net, testloader)
print("finish model: %s" % load_path)
return predicts
# to be modifed when testing on different model
load_path = './save/densenet169/b8_lr1e-4_d0_logloss.pkl'
root_dir = os.getcwd()
net = torchvision.models.densenet201(pretrained=False)
net.classifier = nn.Linear(1920, 1)
total = 0
result = 0.0
test_transform = densenet.test_transform
print(test_transform)
for i in study_type:
for j in pn:
testset[i][j] = per_class_dataset(root_dir,
'/traincsv/%s_%s.csv' % (i, j),
transform=test_transform,
RGB=True)
total += len(testset[i][j])
testloader[i][j] = torch.utils.data.DataLoader(testset[i][j],
batch_size=8,
shuffle=False,
num_workers=2)
if torch.cuda.device_count() > 1:
print("Let's use GPUS!")
# net = nn.DataParallel(net,device_ids=[0,2,3,4,5])
net.load_state_dict(torch.load(load_path))
if torch.cuda.is_available():
net = net.cuda()
learning_rate = args.learning_rate
drop_rate = args.drop_rate
load_model = args.load_model
if load_model:
load_path = args.load_path
net_name = args.net_name
save_path = args.save_path
if os.path.exists(save_path) == False:
os.makedirs(save_path)
save_path += net_name + '.pkl'
loss_type = args.loss_type
present_dir_path = os.getcwd()
trainset = per_class_dataset(csv_file=r'/train.csv',
root_dir=present_dir_path,
transform=train_transform,
RGB=True)
# print ("trainset 0")
# print (trainset[0])
# print ("trainset 1")
# print (trainset[1])
trainloader = torch.utils.data.DataLoader(trainset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
testset = per_class_dataset(csv_file='/valid.csv',
root_dir=present_dir_path,
transform=test_transform,
RGB=True)
testloader = torch.utils.data.DataLoader(testset,
batch_size=batch_size,
