Remove certain weights from a trained NN

1

Hello! I have a trained feed forward NN with a given number of inputs and I want to remove all the weights associated to one of the inputs (including the input node itself). Basically I want to keep everything, but have an input 1 dimension lower than before. What is the most effective way to do so? Thank you!

2

You could slice the weight parameter and reassign it to the .weight attribute.
E.g. if you would like to keep the first 9 input features out of 10:

lin = nn.Linear(10, 5)
with torch.no_grad():
    lin.weight = nn.Parameter(lin.weight[:, :9])
    
x = torch.randn(2, 9)
output = lin(x)