AssertionError for 2 input channel image and label

[mahootiha-maryam]

Hi, I wanted to do the segmentation on kidneys dataset. For preprocessing data I tried to use cropforground and use labels so I could extract two kidneys. then for giving good dataset to training model I extract the left kidney and right kidney separably and I did these steps for labels. So I had left kidney images, left kidney labels, right kidney images and right kidney labels. I made two datasets with monai left kidney that contains left images and labels, right kidney that contains right images and labels. Because kidneys are two before making dataloader I tried to make a custom dataset with pytorch that gets left kidney dataset and right kidney dataset. Then concatenate the left image with right image and left label with right label. Finally it returns image and label which both of them has 2 channels. This is my custome dataset:

class Data_Single(Dataset):

    def __init__(self, im_dataset_l, im_dataset_r):
        self.im_dataset_l = im_dataset_l
        self.im_dataset_r = im_dataset_r

    def __len__(self):
        return len(self.im_dataset_l)

    def __getitem__(self, index):
        if type(index) is not int:
            if type(index) is not np.int64:
                raise ValueError(f"Need `index` to be `int`. Got {type(index)}.")
        if type(index) is np.int64:
            index = index.item()
            print(index)
                    
        img1 = self.im_dataset_l[index]['image']
        img2 = self.im_dataset_r[index]['image']
        
        
        lbl1 = self.im_dataset_l[index]['label']
        lbl2 = self.im_dataset_r[index]['label']
        
        img = torch.cat((img1,img2),0)
        lbl = torch.cat((lbl1,lbl2),0)

Then I chose Unet network that has 2 input channels and 2 output channels.

model = UNet(
    spatial_dims=3,
    in_channels=2,
    out_channels=2,
    channels=(16, 32, 64, 128, 256),
    strides=(2, 2, 2, 2),
    num_res_units=2,
    norm=Norm.BATCH,
).to(device)

I tried to extract image and label from batch data:

   for batch_data in train_loader:
        step += 1
        inputs, labels = (
            batch_data[0].to(device),
            batch_data[1].to(device)
                    )

And the output of batch data for image and label is:

torch.Size([1, 2, 128, 128, 128])
torch.Size([1, 2, 128, 128, 128])

But when I try to train model with this method:

outputs = model(inputs)
loss = loss_function(outputs, labels)
loss.backward()

I get this error on the second line of previous code:

raise AssertionError("labels should have a channel with length equal to one.")

I can not understand where is problem, can you please help me?

[wyli]

I think the issue is with loss_function, you can set to_onehot_y=False, sigmoid=True

[mahootiha-maryam]

Thanks for your answer. It solved problem. But do you know what is the usage of to_onehot_y in the loss function and when I set it to false what happens? I wanted to understand the logic of the error I got.

[JohnMasoner]

Hey @mahootiha-maryam ,

maybe the docs is helpful for you. https://docs.monai.io/en/stable/losses.html#diceloss

Sure, this is a loss function demo.

to_onehot_y (bool) – whether to convert y into the one-hot format. Defaults to False.

Example:

not_one_hot data:[1]
one_hot data [1, 0]

the one_hot encoding is more convenient to calculate the loss function.

Hope this is useful to you.