Feeding the CNN with Specific Features

vision
1

Hi,

I want to feed a CNN with vectors or matrices based on the result of some feature extraction methods like GLCM, Haralick, etc. How can I configure it?

2

Have a look at this post.

Briefly, as a mini-example:

import torch
conv = torch.nn.Conv2d(3, 16, 1)
desired_weight = np.random.randn(16, 3, 1, 1)  # initiating with something random for illustration
conv.weight = torch.nn.Parameter(torch.Tensor(desired_weight))

This will set the weights (but not the biases :slight_smile: exercise left to the reader) to your custom desired_weight.

If you’d like to freeze these layers during subsequent training (for example you want to use these precise features but you do want to train a downstream classifier) you can do this:

for param in conv.parameters():
    param.requires_grad = False

Hope this helps!
Andrei