RuntimeError: Expected 3D (unbatched) or 4D (batched) input to conv2d, but got input of size: [1, 1, 374, 402, 3]

1

image
image897×757 38.7 KB

I am getting error in this line. I can’t understand why I am getting this.

2

image
image1016×409 23 KB

Had to add in the comment section as new users can’t upload more than one media.

3

nn.Conv2d expects an input in the shape [batch_size, channels, height, width] while you are passing a 5-dimensional input as given in the error message.
Remove the unnecessary dimension and permute the input to have the aforementioned channels-first memory layout.

PS: you can post code snippets by wrapping them into three backticks ```.

2 Likes
5

@jrTanvirHasan27
is your problem solved or not?
i am facing the same error?

6

You can either reduce the dimension as @ptrblck says, or if available for your use case use the conv3d instead of the 2d counterpart.

7

You can use the .reshape() method to flatten the data. Specifically, if you have a tensor with dimensions [batch_size, num_frames, channels, height, width], you can use .reshape() to turn it into a tensor with dimensions [batch_size * num_frames, channels, height, width]. This means all batch and frame sizes will be stacked together to form a single large batch. Result should be : ([1, 374, 402, 3])

8

I have the same problem but I’m using 3D MRI images which are in nifti format
the original dimension is [170, 195, 160], in which 170 is the number of slices.

I wrote a custom dataset class and a custom transform.

also I tracked the dimension they changes from ([ 1, 1, 170, 195, 160]) to ([1, 1, 1, 170, 195, 160]) before entering the model

how to solve this problem
(I’m using 3D model with conv3d)

9

Hey there,
nn.Conv3d expects the input in the format (N,Cin,D,H,W) → 5 dim. Yours is a 6 dim input.

  • you could reduce a dimension with torch.squeeze() the appropriate dim (1 or 2)
  • Also, to double check if your inputs match with the expected input format for nn.Conv3d, otherwise permute it accordingly
    For more information: Conv3d — PyTorch 2.4 documentation
10

tahnks for quick replay, but I still have a problem.

I wrote this transform class

class ToTensorNIfTI:
def call(self, img):
img = np.array(img.get_fdata(), dtype=np.float32) if hasattr(img, ‘get_fdata’) else np.array(img, dtype=np.float32)

    #print(f"Original NIfTI image shape: {img.shape}")
    
    img -= img.min()
    img /= img.max()
    img = torch.tensor(img, dtype=torch.float32)
    #print(f"Tensor shape after ToTensorNIfTI: {img.shape}")

    # Add channel dimension
    if img.ndimension() == 3:  # For 3D images
        img = img.unsqueeze(0)  # Add channel dimension

    
    return img

so the error was as I said
Expected 4D (unbatched) or 5D (batched) input to conv3d, but got input of size: [1, 1, 1, 160, 195, 170]

and when I removed this lines

Add channel dimension

    if img.ndimension() == 3:  # For 3D images
        img = img.unsqueeze(0)  # Add channel dimension

the error became: expected 5D input (got 4D input)

Then I rewrote these lines.

Add channel dimension

if img.ndimension() == 3:  # For 3D images
    img = img.unsqueeze(0)  # Add channel dimension

and used reshape()
def train(model, dataloader, optimizer, criterion, device):
model.train()
running_loss = 0.0
corrects = 0
total = 0

for inputs, labels in dataloader:
   # Print input shape before reshaping
   #print(f"Input shape before reshaping: {inputs.shape}")

   # Reshape the input tensor to [batch_size, channels, depth, height, width]
    # Remove the extra dimension (in this case, the dimension of size 1)
 if inputs.ndimension() == 6 and inputs.size(2) == 1:
  inputs = inputs.squeeze(2)  # Remove the dimension with size

but the error still
expected 5D input (got 4D input)

so what should I do?

11

It seems you cannot narrow down where exactly the image dimensions increase so could you post a minimal, formatted, and executable code snippet reproducing the issue?

1 Like