Fine-tuning convolutional filters

vision
1

Hi, I want to fine-tune some of the convolutional filters in ResNet-50. To achieve this, I need to freeze the other convolutional filters. As far as I know, I can only freeze convolutional “layers”, but not convolutional “filters”.
I already tried the following code, but it seems that it is impossible to make selected convolutional filters require gradient.

import torch
import torchvision

net = torchvision.models.resnet50(pretrained=True)
for layer_num, (name, param) in enumerate(net.named_parameters()):
    if 'conv' in name:
        param.requires_grad = False
        param[0].requires_grad = True
        print(param.requires_grad, param[0].requires_grad) # expected False True, but False False

Is there anyway to perform this kind of fine-tuning?

2

When you say, “filters” and “layers”, do you mean weights and biases? If that is the case, here is one way you might go about it:

import torchvision
import torch.nn as nn

net = torchvision.models.resnet50(pretrained=True)
with torch.no_grad():
    for layer in net.modules():
        if isinstance(layer, nn.Conv2d):
            print(layer)
            print(layer.weight.requires_grad)
            layer.weight.requires_grad = False
            print(layer.weight.requires_grad) 
            if not isinstance(layer.bias, type(None)): #Check if layer has no bias, otherwise you'll get an error that NoneType object has no requires_grad
                print(layer.bias.requires_grad)

As a side note, the Conv2d layers in the ResNet-50 do not use bias because they use BatchNorm2d, and a bias would be redundant. To target those, just use:

if isinstance(layer, nn.BatchNorm2d):
    ...