Vision: Feature Pyramid Network code bug

[cl2227619761]

https://github.com/pytorch/vision/blob/c2e8a00885e68ae1200eb6440f540e181d9125de/torchvision/ops/feature_pyramid_network.py#L60

this line does not work, I think it should be modified as self.modules() instead of self.children()

[pmeier]

Hi @cl2227619761, what do you mean by

this line does not work

?

Please post a minimal example where you encounter an issue with this line.

[fmassa]

Hi @cl2227619761

That is a very good catch!

@pmeier the issue is that self.children() returns the immediate modules, so in this case it would be inner_blocks and layer_blocks, none of which are nn.Conv2d.
The fix is what @cl2227619761 mentioned: replace self.children() with self.modules().
But this brings the question on the impact of the initialization. If I remember correctly, we needed the custom initialization in order to get best results, so I'm surprised that this was left out.

A PR fixing this would be great, but it would also be great if we could measure the impact of this incorrect initialization in the model performance.

@mthrok if you find the time, it would be great to assess how much this mistake affects performance.

[pmeier]

My bad, I misunderstood the comment above the loop:

https://github.com/pytorch/vision/blob/c2e8a00885e68ae1200eb6440f540e181d9125de/torchvision/ops/feature_pyramid_network.py#L59-L63

[gan3sh500]

Can I help with this? If you point me to what model/scripts you used to train I can train with the correct init and the current one.

[pmeier]

@gan3sh500 You can find the training scripts here.

[gan3sh500]

I’ll have to train on single 2080ti. I’ve seen smaller batch size create worse convergence but is it fine if I compare with both the kaiming init and otherwise trained with same conditions?

[fmassa]

@gan3sh500 sorry for the delay in replying.

If you change the number of GPUs, you'll need to adapt the learning rate to follow the linear scaling rule -- if you divide the global batch size by 8x, you should divide the learning rate by 8x