How to get gradients of the output w.r.t. to a convolutional layer?

[Sandareka Wickramanayake]

Hi all, 

I'm trying to implement this paper - https://arxiv.org/abs/1610.02391 using python. For that I want to get gradient of the output of a specific class with respect to the last convolutional layer. I came across the following usage of backward() function. 

label = np.zeros((1, 6))
label[0, interested_class] = 1

net.backward(**{net.output[0]: label}) 

Assuming I have size classes in my network. 

However it gives gradient w.r.t to the input layer.

I tried to use the following usage but it is not giving the desirable output.

label = np.zeros((1,6))

label[0,interested_class] = 1

net.backward(end=conv, **{net.output[0]:label}) 

Precisely, I want to get the gradient of output layer w.r.t conv layer values. 

Any help is highly appreciated!

[zhuwei]

Hi Sandareka,

I am working on the gradient visualization recently. Did you solve this problem?  

Any suggestions or advice will be greatly appreciated!

在 2017年8月4日星期五 UTC-4下午12:40:41，Sandareka Wickramanayake写道：

[Sandareka Wickramanayake]

Hi Zhuwei, 

Please check if this post helps you.