Use backward twice, and apply retain_graph=True for the first backward, but the evaluation result is always the same

1

Use backward twice, and apply retain_graph=True for the first backward, but the evaluation result is always the same.

image
image564×594 39.6 KB

2

Hi,

Could you give more information please about your code and what do these prints correspond to?

3

Of course. In one epoch, I want to have the two backward of the network, In the first backward, I can get the gradient of every layer which i want to use to update the parameters. And I have the second background with a different cost function.

4

Could you give a small code sample (30-40 lines) that reproduces what you see? So that I can test it on my machine?