How to change the params in one conv1d layer

1 

import torch
import torch.nn as nn
import pdb
import torch.nn.init as init
pdb.set_trace()
# Third convolutional layer
conv3 = nn.Conv1d(64, 128, kernel_size=3)

ab = conv3.weight
print("ab",ab[:,1:3,:])

new_params = torch.ones(128, 2, 3)  # Example: generate new parameters

ab.data[:, 1:3, :] = new_params
cd = conv3.weight
print("cd",cd[:,1:3,:])
print("ab",ab[:,1:3,:])
print(sum(cd - ab))

For the above code, I think ab is the previous params of conv3, and cd is the new replaced weights and they are different. But why they are the same value. Thanks for your reply.

2

Don’t use the deprecated .data attribute, but copy the new values directly to the weight in a no_grad context:

with torch.no_grad():
    conv3.weight[:, 1:3].copy_(new_params)