lr=max_lr,
momentum=momentum,
weight_decay=weight_decay)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, scheduler_step, min_lr)
# Load data
train_id = fold_train[idx]
val_id = fold_valid[idx]
X_train, y_train = trainImageFetch(train_id)
X_val, y_val = trainImageFetch(val_id)
train_data = DataSource1(X_train,
mode='train',
mask_list=y_train,
fine_size=fine_size,
pad_left=pad_left,
pad_right=pad_right)
train_loader = DataLoader(train_data,
shuffle=RandomSampler(train_data),
batch_size=batch_size,
num_workers=8,
pin_memory=True)
val_data = DataSource1(X_val,
mode='val',
mask_list=y_val,
fine_size=fine_size,
pad_left=pad_left,
pad_right=pad_left)
val_loader = DataLoader(val_data,
for i in range(5):
# Load test data
image_test, _ = trainImageFetch(test_id[i])
overall_pred = np.zeros((len(test_id[i]), 202, 202), dtype=np.float32)
# Get model
salt = EncNet(1)
salt = salt.to(device)
pred_null = []
pred_flip = []
# Load weight
param = torch.load(save_weight + weight_name +str(i)+ '.pth')
salt.load_state_dict(param)
# Create DataLoader
test_data = DataSource1(image_test, mode='test', fine_size=fine_size, pad_left=pad_left, pad_right=pad_right)
test_loader = DataLoader(
test_data,
shuffle=False,
batch_size=batch_size,
num_workers=8,
pin_memory=True)
# Prediction with no TTA test data
salt.eval()
print(len(test_loader))
for images in tqdm(test_loader, total=len(test_loader)):
images = images.to(device)
with torch.set_grad_enabled(False):
pred = salt(images)[0]
# pred = salt(images)
pred = F.sigmoid(pred).squeeze(1).cpu().numpy()