Converting feedforward NN to CNN

vision
1

Hi everyone,

I’m training a basic feedforward neural network to make amino acid predictions based on protein structure information.

Here is how the NN architecture is set up:

class NeuralNetwork(nn.Module):
    def __init__(self, input_nodes):
        super(NeuralNetwork, self).__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(input_nodes, 64),             # input 
            nn.ReLU(),
            nn.Dropout(p=0.5),

            nn.Linear(64, 64),                      # hidden layer 1 
            nn.ReLU(),
            nn.Dropout(p=0.5),

            nn.Linear(64, 64),                      # hidden layer 2 
            nn.ReLU(),
            nn.Dropout(p=0.5),

            nn.Linear(64, 64),                      # hidden layer 3 
            nn.ReLU(),
            nn.Dropout(p=0.5),

            nn.Linear(64, 20)                       # output
        )

I’m wondering if its possible to make modifications to the set-up above in order to convert it to a convolutional NN?
I’m using a non-image dataset, so I’m wondering if a CNN would even be possible?

2

Yes, you could e.g. replace the nn.Linear layers with nn.ConvXd.

Yes, it would be possible as long as the inputs are in the expected shape, e.g. [batch_size, channels, height, width] for nn.Conv2d. It depends on your use case and how you are handling the data to claim it “makes sense”. If you think your input data has a correlation between neighboring “pixels” it sounds like a valid idea to try out a CNN on this data.