How to reset parameters of layer

1

Hello everyone,
How to reset the parameters of layer4 in resnet18, using the Module.apply? Here my code but it doesn’t work.
Thank you.

def reset_weights(m):
    for name, layer in m.named_children():
        if name=='layer4' and hasattr(layer, 'reset_parameters'):
                print(f'Reset trainable parameters of layer = {layer}')
                layer.reset_parameters()

model.appy(reset_weight)
2

layer4 is an nn.Sequential module, which doesn’t have the reset_parameters method, so your conditions won’t be met.
You could iterate all modules of layer4 and check for this attribute:

def reset_weights(m):
    for name, layer in m.named_children():
        if name=='layer4':
            print(name)
            for n, l in layer.named_modules():
                print(n)
                if hasattr(l, 'reset_parameters'):
                    print(f'Reset trainable parameters of layer = {l}')
                    l.reset_parameters()


model = models.resnet18()
model.apply(reset_weights)
2 Likes
3

It works. I owe you one.

4

How can I reset parameters of layers 3 and higher layers of Resnet18 so that I can train for those layers?