Setting value for hidden_channels in FullyConnectedNet model

[Kiarashdnsh]

Hello,

I'm a novice in AI and image segmentation.
I have a dataset of [256, 256, 256]-sized images and labels. The input channel is 1, and the output channel is 3 (label IDs are 0, 1, and 2, respectively). I'm trying to segment the images using FullyConnectedNet with the MONAI framework. In this model's parameters, I don't know how to set the hidden_channels; however, I used a list of [10] as a placeholder since this is a required parameter.
For this model, I have used the following code:

model = FCN(
        in_channels=1,
        out_channels=3,
        hidden_channels=[10],
        dropout=0.20,
)

and used this tutorial here for running the trainer.

While the trainer is running, on the outputs = model(inputs) line I get this error:

RuntimeError: mat1 and mat2 shapes cannot be multiplied (4x884736 and 1x10)

To avoid this issue, I changed the in_channels to 884736 so that the two matrices could be multiplied. After this fix, on the next run, I get another error on the loss = loss_function(outputs, labels) line:

AssertionError: ground truth has different shape (torch.Size([4, 2, 96, 96, 96])) from input (torch.Size([4, 2]))

Now I have two related questions regarding the situation:

1. How do I know what value to set in the hidden_channels?
2. What I've done to solve the first issue doesn't seem to be rational, since I have only 1 channel in my images. How do I solve that? I, think this issue is related to the value of hidden_channels.

Thanks in advance,
Kiarash

[Kiarashdnsh]

I've read the source code for this model here and still couldn't figure out how to set values for hidden_channels.

[KumoLiu]

Hi @Kiarashdnsh, thanks for your interest here.

I guess you want to use the 'fully convolutional network' (FCN) model to segment your images instead of the fully-connected network (FullyConnectedNet) which is used to do classification. FullyConnectedNet should accept one-dimensional data, shape like Batch_size*Channel.

MONAI/monai/networks/nets/fullyconnectednet.py

Lines 33 to 35 in 4addc5d

	Simple full-connected layer neural network composed of a sequence of linear layers with PReLU activation and
	dropout. The network accepts input with `in_channels` channels, has output with `out_channels` channels, and
	hidden layer output channels given in `hidden_channels`. If `bias` is True then linear units have a bias term.

As a first step, I recommend that you try 'UNet' (a popular variant of FCN); here is a tutorial of a basic UNet training example. (https://github.com/Project-MONAI/tutorials/blob/main/3d_segmentation/torch/unet_training_array.py)

Hope it helps, thanks!

[Kiarashdnsh]

Thanks for your answer. ✔