inference.py

test_transform=transforms.Compose([
    transforms.Resize(IMAGE_SIZE),
    transforms.CenterCrop(IMAGE_SIZE),
    transforms.ToTensor(),
    transforms.Normalize(IMG_MEAN,IMG_STD)
])


test_dataset = CollectionsDatasetTest(csv_file='../input/sample_submission.csv',
                                  root_dir='../input/test/',
                                  image_size=IMAGE_SIZE,
                                  transform=test_transform)

test_dataset_loader = torch.utils.data.DataLoader(test_dataset,
                                             batch_size=TEST_BATCH_SIZE,
                                             shuffle=False,
                                             num_workers=4)


model_ft.load_state_dict(torch.load("model.bin"))
model_ft = model_ft.to(device)

for param in model_ft.parameters():
    param.requires_grad = False

model_ft.eval()
test_preds = np.zeros((len(test_dataset), NUM_CLASSES))
tk0 = tqdm(test_dataset_loader)
for i, x_batch in enumerate(tk0):
    x_batch = x_batch["image"]
    pred = model_ft(x_batch.to(device))
    test_preds[i * TEST_BATCH_SIZE:(i + 1) * TEST_BATCH_SIZE, :] = pred.detach().cpu().squeeze().numpy()

test_preds = torch.from_numpy(test_preds).float().to(device).sigmoid()
test_preds = test_preds.detach().cpu().squeeze().numpy()