This is a repo [no more wip ;)] on implementing Deep Convolutional GAN Architecture on the Pokemon Dataset

This fun little endeavor of mine is essentially built upon concepts from Radford et al. (2015)

Currently, I am not getting clear results (hence why it's WIP), Hopefully after some fine tuning I can get it up and running.

Also facing issues with computation power (oh well). Hopefully, all will be resolved soon.

[UPDATE] 24/4/24: Alright, I changed the architecture for the generator a bit and it looks more coherent now, although I suspect I definitely need more epochs

[UPDATE] 5/5/24: Getting more or less desirable results, changed the Network architecture by a lot (Kaggle), ran the code on a kaggle notebook using a T4 GPU X2 for about 500 epochs. Planning on modifying my older architecture to match this performance now that the computational power issue has been resolved. (Output isn't visible in the notebook because I suppose it doesn't support the generated GIF with the video player buttons)

[UPDATE] 9/5/24: Got it working on my local Gtx 1650, gonna tinker around a bit

[UPDATE] 11/5/24: too much noise on local gpu, 1000 epochs fail to generate anything even remotely comprehensible. I think the previous results were the best ones yet. As you can see down below the Discriminator loss is high => It might be overpowering the generator due to which the generator fails over and over.

Also, the Discriminator loss abnormally spikes at epochs 274 & 275, which basically dooms the remaining cycles for the Generator.

[UPDATE] 12/5/24: turns out I had mixed two datasets, surprised I still got any results, anyways the final output video was too large so I uploaded it here. I'm going to consider this as a W (although there was noise in the 900s of epochs so 800 is the sweet spot for epochs on this)

Generator

Discriminator

Output:

Final.PKMN.VID.mp4

Some that I think make great potential candidates:

NOTE: The Alt-branch of this code contains implementation of the same on another pokemon images dataset (larger essentially), you can check that out too.