Adapted SqueezeNet for Variable Size Input Images

[raunakdoesdev]

Replaced final AveragePooling layer with an AdaptiveAveragePooling layer to allow for no code changes when switching from 255x255 images to other image sizes.

[fmassa]

Thanks for the PR.
According to the discussion in #155 , it might be preferable to add a constructor argument to switch between both behaviors. What do you think?

[raunakdoesdev]

The way I see it is, in the current code base any input image of a different size (bigger or smaller) causes the network to crash when run. When given an adaptive pool layer rather than fixed average pooling the network still runs as intended with 255x255 images for classification but also supports other image sizes with the same resultant behavior. I think in the case of other jobs such as segmentation (fully dense outputs), it should be up to the end user to make the necessary modifications since there are too many use cases to create arguments in the constructor for each case.

The change to Adaptive Pooling makes sense to me since it maintains the current functionality of the network, adds new functionality and doesn't introduce unneeded complexity for the user.

[fmassa]

I'm ok to merge this as is. Do you have an idea on how the network performs for different image sizes?

[apaszke]

If we're going to change this network shouldn't we change all other ones too? I'm not sure if it's a good idea. Replacing a single layer should be quite straightforward and is more explicit

[raunakdoesdev]

Yeah I agree we should be replacing the other ones as well. If no one else wants to I can take it on, but it may take me a few days since I'm still pretty new to Pytorch and Deep Neural Nets. It should be a good learning experience to go through and look at the structures of each of these nets though.

…

On Sat, May 13, 2017, 3:49 AM Adam Paszke ***@***.***> wrote: If we're going to change this network shouldn't we change all other ones too? I'm not sure if it's a good idea, replacing a single layer should be quite straightforward and is more explicit — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#166 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AKchR_zRVepIzYlROVtbGM-tBDVF8BPFks5r5YrFgaJpZM4NTQTd> .

[chao1224]

Just happend to face same problem. I think at least it would be reasonable to add some comments in the code.

[fmassa]

This has been implemented in #643.

Thanks for the PR!