How to get the partial derivative of the output data with the middle layer data

[wxwx (wx)]

In a MLP neural network, I want to get partial derivative of the output data with the middle layer data. such as, as shown below: ,i need the partial derivative of Etotal with neto1. i find register_backward_hook can get this value, but i fail to use. below is my code. i do not know the error

[richard]

Does that not work for you? What happens?

[wxwx (wx)]

first, thanks for reply. This code is work, but i don’t know whether the obtained values of model.layer3.register_backward_hook(hook) are partial derivative of Etotal and neto. Second, I want to use nn.MSELoss() replace nn.CrossEntropyLoss(), so, the criterion= nn.CrossEntropyLoss() change to criterion= nn.MSELoss(size_average = False).

this code proceed with error below:

then, i add line code:

but, there also has error:

for this error, i really don’t know how to solve this problem or how to use MSELoss in my program
I know the ‘’target‘’ of MSELoss is ：

[jpeg729]

Presumably label contains the labels as class identifier numbers, not as a 1-hot encoding.

Convert int into one-hot format

Hi, You can use the scatter_ method to achieve this. I would also advise to create the y_onehot tensor once and then just fill it: import torch batch_size = 5 nb_digits = 10 # Dummy input that HAS to be 2D for the scatter (you can use view(-1,1) if needed) y = torch.LongTensor(batch_size,1).random_() % nb_digits # One hot encoding buffer that you create out of the loop and just keep reusing y_onehot = torch.FloatTensor(batch_size, nb_digits) # In your for loop y_onehot.zero_() y_onehot.scat…

[wxwx (wx)]

thanks for replay. according above contents, I change my code ,as below:

but there is a error, do you know how to solve?

[jpeg729]

It looks like label is a Variable, not a Tensor. Try this…

label_onehot.scatter_(1, label.data, 1.0)

[wxwx (wx)]

there is a new error.

[jpeg729]

We need to check the size of label. What does print(label.size()) show?

[wxwx (wx)]

this is the result. I really appreciate your help.

[jpeg729]

I assume 32 is your batch size.

Try this…

label_onehot.scatter_(1, label.data.unsqueeze(dim=1), 1.0)

[wxwx (wx)]

there is another problem,sorry.

As you say, batch_size=32

[jpeg729]

The one_hot labels are Tensors not Variables, you need to wrap them in a Variable.
e.g. just after the line we worked on previously.

return Variable(label_onehot)

[wxwx (wx)]

thanks , this is work ,but my code have another error. i think i need myself to do first, if i can’t solve, i will ask you, thanks, thanks,thanks

[wxwx (wx)]

my program is work, thank you very much