Model output is not the same after speedup.

[gmiesw]

Environment:

• NNI version: commit f1b8cd2
• NNI mode (local|remote|pai): remote
• Client OS: Ubuntu 16.04.5 LTS (Xenial Xerus)
• Python version: Python 3.6.4
• PyTorch/TensorFlow version: PyTorch 1.5.0
• Is conda/virtualenv/venv used?: pyenv
• Is running in Docker?: No

Log message:

• nnimanager.log: INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.0.conv1 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.0.conv2 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.2.conv3,layer3.1.conv3,layer3.0.downsample.0,layer3.5.conv3,layer3.0.conv3,layer3.3.conv3,layer3.4.conv3 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.1.conv1 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.1.conv2 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.2.conv1 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.2.conv2 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.3.conv1 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.3.conv2 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.4.conv1 INFO:nni.compression.torch.pruning.one_shot:Pruning the dependent layers: layer3.4.conv2

What issue meet, what's expected?:

I have different outputs before and after model speedup:

It's expected that outputs would be roughly the same.

How to reproduce it?:

Please, run the code from the following gist: https://gist.github.com/nikitakaraevv/1eaa735d5323f833356152f6d41ef8e4

Additional information:

I'd like to understand why it happens first of all and how it can be avoided.

[gmiesw]

Example comment ...