How to check error/loss between layers?

1

Considering a multi-layer network L1 -> L2 -> L3 -> …

and using “loss.backward()” function to update the weights.

Is there a way to compute the loss between each layers as it is propagated? L1 <-(loss here?) L2

2

Usually you don’t use a loss between intermediate layers.
Some model architectures like Inception use an auxiliary loss.
Would you rather like to see the gradients at a specific layer?
If so, you could just print it using:

loss.backward()
print(model.some_layer.weights.grad)
3

Actually I was interested in the loss itself and not the gradient.

To be specific about my use case I wanted to use two models in sequence and propagate the loss from the 2nd model to the 1st.

Any suggestions on how this can be accomplished?

4

If you want to use two models sequentially, you can just pass the output of one model to the other:

output1 = model1(x)
output2 = model2(output1)

As long as you don’t detach the tensors, the loss will be backpropagated to model1.

1 Like
5

I did try that but the weights of the 2nd model are not updated, while the weights of the first model are updated.

criterion = nn.CrossEntropyLoss()
optimizer.zero_grad()

output1 = model1(x)
output2 = model2(output1)

loss = criterion( output2 , labels )

a = copy.deepcopy(model2.linear1.weight.data)
loss.backward()
optimizer.step()
b = copy.deepcopy(model2.linear1.weight.data)

print(torch.equal(a, b))

The above code prints “True” when run.
However if you check the “model1” instead of the “model2” weights it prints “False”.

Can you please help me find out where I am going wrong. Thanks in advance.

6

So sorry my bad. I had defined the optimizer only on the first model. That is why it was only being updated.

Thank you for your insights.

1 Like