Revision history [back]

In PyTorch, weights can be duplicated and altered by accessing the parameters of a neural network model. Here is an example:

import torch.nn as nn

# Create a neural network model
model = nn.Sequential(
    nn.Linear(10, 5),
    nn.ReLU(),
    nn.Linear(5, 2),
    nn.Sigmoid()
)

# Duplicate and alter the weights of the first linear layer
new_layer = nn.Linear(10, 5)
new_layer.weight.data = model[0].weight.data.clone()  # Duplicate weights
new_layer.bias.data = model[0].bias.data.clone()  # Duplicate biases
new_layer.weight.data *= 2  # Alter weights by multiplying by 2
new_layer.bias.data += 1  # Alter biases by adding 1

# Replace the first linear layer with the new one
model[0] = new_layer

In this example, we first create a neural network model using nn.Sequential(). We then duplicate and alter the weights of the first linear layer by creating a new nn.Linear() layer, cloning the weights and biases from the original layer, and then modifying them as desired. Finally, we replace the first linear layer of the model with the new one.

Note that altering weights and biases directly can be unstable and may cause the model to perform poorly. It is generally better to use PyTorch's optimization algorithms to train the model and update the weights and biases automatically.