PyTorch specify model parameters

Question

I am trying to create a convolutional model in PyTorch where

one layer is fixed (initialized to prescribed values)
another layer is learned (but initial guess taken from prescribed values).

Here is a sample code for model definition:

import torch.nn as nn

class Net(nn.Module):
    def __init__(self, weights_fixed, weights_guess):
        super(Net, self).__init__()
        self.convL1 = nn.Conv1d(1, 3, 3, bias=False)
        self.convL1.weight = weights_fixed # I want to keep these weights fixed

        self.convL2 = nn.Conv1d(3, 1, 1, bias=False)
        self.convL1.weight = weights_guess # I want to learn these weights

    def forward(self, inp_batch):
        out1 = self.convL1(inp_batch)
        out2 = self.convL2(out1)

        return out2

and the sample use:

weights_fixed = ...
weights_guess = ...

model = Net(weights_fixed, weights_guess)

loss_fn = nn.CrossEntropyLoss()
optim = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)

train_dataset = ... #define training set here

for (X, y) in train_dataset:
    optim.zero_grad()
    out = model(X)
    loss = loss_fn(out, y)
    loss.backward()
    optim.step()

How can I make the weights weights_fixed - fixed and weights_guess - learnable?

My guess would be weights_fixed = nn.Parameter(W1,requires_grad=False) weights_guess = nn.Parameter(W2,requires_grad=True) where for the sake of completeness import numpy as np import torch

krnl = np.zeros((5,order+1))
krnl[:,0] = [ 0. , 1., 0. ]
krnl[:,1] = [-0.5, 0., 0.5]
krnl[:,2] = [ 1. ,-2., 1. ]
W1 = torch.tensor(krnl)

a = np.array((1.,2.,3.))
W2 = torch.tensor(a)

But I am utterly confused. Any suggestions or references would be greatly appreciated. Of course I went over PyTorch docs, but it did not add clarity to my understanding.

Sergii Dymchenko · Accepted Answer · 2019-03-21 20:25:05Z

3

Just wrap the learnable parameter with nn.Parameter (requires_grad=True is the default, no need to specify this), and have the fixed weight as a Tensor without nn.Parameter wrapper.

All nn.Parameter weights are automatically added to net.parameters(), so when you do training like optimizer = optim.SGD(net.parameters(), lr=0.01), the fixed weight will not be changed.

So basically this:

weights_fixed = W1
weights_guess = nn.Parameter(W2)

edited Mar 21, 2019 at 20:25

answered Mar 21, 2019 at 20:14

Sergii Dymchenko

7,2671 gold badge25 silver badges49 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

Anton Ganichev · Accepted Answer · 2020-01-09 09:17:32Z

1

You can pass to optimizer only parameters that you want to learn:

optim = torch.optim.SGD(model.convL2.parameters(), lr=0.1, momentum=0.9)
# Now optimizer bypass parameters from convL1

If you model have more layers, you must convert parameters to list:

params_to_update = list(model.convL2.parameters()) + list(model.convL3.parameters())
optim = torch.optim.SGD(params_to_update, lr=0.1, momentum=0.9)

as described here: https://discuss.pytorch.org/t/giving-multiple-parameters-in-optimizer/869

answered Jan 9, 2020 at 9:17

Anton Ganichev

2,6021 gold badge20 silver badges19 bronze badges

Comments

akshayk07 · Accepted Answer · 2019-03-21 12:47:46Z

0

You can do this :

# this will be inside your class mostly
self.conv1.weight.requires_grad = False

And this will be where you are defining the optimizer :

optimizer = optim.SGD(filter(lambda p: p.requires_grad, net.parameters()), lr=0.1)

So, the optimizer will only use the parameters which have gradients enabled.

answered Mar 21, 2019 at 12:47

akshayk07

2,2201 gold badge25 silver badges35 bronze badges

Comments

eljiwo · Accepted Answer · 2021-02-18 20:30:40Z

0

Modify your model definition to be:

import torch.nn as nn

class Net(nn.Module):
    def __init__(self, weights_fixed, weights_guess):
        super(Net, self).__init__()
        self.convL1 = nn.Conv1d(1, 3, 3, bias=False)
        self.convL1.weight = weights_fixed # I want to keep these weights fixed
        self.convL1.requires_grad = False

        self.convL2 = nn.Conv1d(3, 1, 1, bias=False)
        self.convL1.weight = weights_guess # I want to learn these weights

    def forward(self, inp_batch):
        out1 = self.convL1(inp_batch)
        out2 = self.convL2(out1)

        return out2

answered Feb 18, 2021 at 20:30

eljiwo

8562 gold badges11 silver badges40 bronze badges

Collectives™ on Stack Overflow

PyTorch specify model parameters

4 Answers 4

Comments

Comments

Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

4 Answers 4

Comments

Comments

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related