[SOLVED] Getting wrong input error in pytorch model

Getting wrong input error in pytorch model

I am getting the following error when I am trying to execute a model in pytorch.

Given groups=1, weight of size [64, 3, 4, 4], expected input[1, 4, 512, 512] to have 3 channels, but got 4 channels instead

I understand that the input image I am giving to the model has 4 channels and it requires to have 3 instead. But, can someone tell me what is the meaning of each number in the parenthesis and from where can I start to debug this problem? Because I am giving 2 types of images and labels for the images as input to the model.

If you require any more information, please comment. I will happy to provide.

Solution

The kernel group is of size [64,3,4,4] -

64 filters or kernels in this layer
3 channels per kernel (should match channels per image)
kernel size (height and width) of 4 and 4, respectively

And in your image, as state in comments - [1,4,512,512] = [batch size, channels, height, width].

To address this issue, you can either remove a channel from your image, or else add an additional channel to dimension 1 of the kernel, something like this:

model.conv1.shape #  [64,3,4,4]

new_conv_layer = torch.nn.Conv2d(4, 64, kernel_size=4)

# initialize first 3 channels with values from old layer
new_conv_layer.weight.data[:,:3,:,:] = model.conv1.weight.data.clone()

# initialize new channel however you want
new_conv_layer.weight.data[:,3,:,:] =  ... 

# assign replacement layer 
model.conv1 = new_conv_layer

model.conv1.shape  # [64,4,4,4]