Fleshing this task out for community members to pick up.

This is a task in which you will be modifying the neural network layer nn.AvgPool3d to have implicit zero padding.

To visually explain, this is a layer that does Average downsampling, like this:
https://gist.github.com/anonymous/26ebfabaded1c3ee8a7cf0c63e305737

Now, while nn.AvgPool2d has support for padding, nn.AvgPool3d does not.
You will be adding this feature.

You will look at the nn.AvgPool2d code and understand how padding is introduced in that code, and replicate the same behavior in 3d.
Here are the code pointers:

Python entry point: https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/pooling.py#L553
Calls here: https://github.com/pytorch/pytorch/blob/master/torch/nn/functional.py#L203
Calls here: https://github.com/pytorch/pytorch/blob/master/torch/nn/_functions/thnn/pooling.py#L340
Now we are entering C
The CPU version: https://github.com/pytorch/pytorch/blob/master/torch/lib/THNN/generic/VolumetricAveragePooling.c
The GPU version: https://github.com/pytorch/pytorch/blob/master/torch/lib/THCUNN/generic/VolumetricAveragePooling.cu
Both CPU and GPU version have to be modified.

Start with the CPU version and write unit tests in test_nn.py (search for existing tests).
Once you finish the CPU version, modify the GPU version VolumetricAveragePooling.cu . If you do not have a NVIDIA CUDA based GPU I will take over the PR from here to test the GPU side of things.

You will have more questions, and we shall resolve them via comments here or on slack.

Lastly, look at the CONTRIBUTING document for tips on locally developing pytorch.

padding for nn.AvgPool3d? #915

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions