I have it trained to perform 80% correct In a matter of minutes, instead of hours. plus it runs on the GPU. Thanks all!

https://github.com/sjhalayka/pytorch_cats_vs_dogs

So, I made a sliding window for handling the training data. this way it doesn’t require a whole lot of GPU RAM. I’m also able to use 256x256 sized images instead of 32x32. Preliminary result is 90% correctness. Now that I can be satisfied with.

So it seems, in the end, that both the competitive networks and a regular network produce similar results in similar amounts of time. No magic bullet.