Regarding custom layer

1

Hi,
Can anyone help me how to define addition of two vectors (adding element one by one) with user defined forward and backward propagation using custom layer.
Thanks

2

Hi,

This section of the doc should contain all the informations you need :slight_smile:

3

Actually this kind of example is confusing. Is there any simplified examples like addition of two vectors ?
Thanks

4

There is an example in that tutorial that presents multiplication by a constant (called MulConstant). Isn’t that close enough to adding two vectors?

5

I tried to run the custom c++ extension code as mentioned in the post (https://pytorch.org/tutorials/advanced/cpp_extension.html) but the time taken for python and cpp versions are entirely different for backward propagation especially.
(test1) gputest@gputest-OptiPlex-745:~/PycharmProjects/test1/extension-cpp-master$ python benchmark.py py
Forward: 524.282/560.827 us | Backward 844.955/890.770 us
(test1) gputest@gputest-OptiPlex-745:~/PycharmProjects/test1/extension-cpp-master$ python benchmark.py cpp
Forward: 473.022/497.065 us | Backward 1243.830/1299.400 us

Because he mentioned that backward propagation time should be similar for both. But here I got different values will that would be an issue?
Thanks

6

Hi,

A difference of less than 0.4ms for a single run is not really significant I think. This can be caused by any other program running at the same time and competing for the same ressource.