mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Xingjian Shi <notificati...@github.com>
Subject Re: [apache/incubator-mxnet] [Development] MXNet 2.0 Update (#18931)
Date Wed, 19 Aug 2020 18:46:21 GMT
@StevenJokes In addition, for the DCGAN issue that is related to D2L, a simple way to verify
that you are correct (and also convince the others), is to write a test case that checks whether
these two networks are **equivalent**.

For example, you have a network A implemented in MXNet and a network B implemented in PyTorch.
There are several checks that you can do:

- Just try to see if these two networks have the same number of parameters
- Do a forward pass of both networks and check whether the outputs are the same. 
- Do a forward + backward and match the gradient.

Usually, you will need to do more to convince the others that certain issues exist. There
are some examples:

- Here, the minimal reproducible example related to Autograd helps us locate the problem:
https://github.com/apache/incubator-mxnet/issues/17989
- A minimal example that captures a potential issue of the GELU implementation in MKLDNN https://github.com/apache/incubator-mxnet/issues/18826

It will be a good practice if you can write such test cases and tell D2L people.


-- 
You are receiving this because you were mentioned.
Reply to this email directly or view it on GitHub:
https://github.com/apache/incubator-mxnet/issues/18931#issuecomment-676597008
Mime
  • Unnamed multipart/alternative (inline, 7-Bit, 0 bytes)
View raw message