add pytorch/torchvision porting and comparison script

[CarloLucibello]

No description provided.

[theabhirath]

This script and an older version that @darsnack had posted as a gist both work, but I find that loadmodel! is taking an insane amount of memory for even small models like VGG-11 or ResNet-18. However I have only tried this on my PC and cannot be sure...can someone else confirm if this is an issue?

[darsnack]

Here is also a script for porting weights.

[theabhirath]

Here is also a script for porting weights.

Yep, that's the one I mentioned. Here too loadmodel! seems to eat up a ton of memory for me on my PC.

[ToucheSir]

Can you take a quick allocation profile with and show the flamegraph here using PProf or the VSCode profile view?

[theabhirath]

Sure. Got this with PProf. It's a screenshot of the flame graph, but if you want the raw results file I can try and upload that as well to somewhere else (I don't think GitHub supports sharing that).

[theabhirath]

Ah, @ToucheSir, I neglected to mention something which on a re-run I find might be the cause of this. I was using @darsnack's script to port weights but I added an fmapstructure(identity, weights) in the end before saving because IIRC we had discussed this would avoid saving information like the RNGs and allow us to save and load weights independent of Julia version (as would be ideal). When I remove that line, the memory requirement goes down to something manageable again. Is loadmodel! somehow very bad for loading models stored in this format? Can this be fixed?