Hi guys and gals...
I was bored this week, to I'm trying to build a GAN (Generative Adversial Network) for tbamud.
So for now I'm training a model for generate mobs and a discriminator to identify if the mob generated is valid. The GAN project supports all of lib/world files. I included two encoders:
- one that converts all text in a file as a single encoded text. So all mobs in the file 30.mob is stored as a single text, 31.mob on another text. All in the same encoded mob file.
- The other encoder splits each item at the #. Since the zon files doesn't follows the same pattern, it doesn't work with this encoder.
The code have four parts: 2 kinds of encoder, 1 for loading and training the Gan model and 1 for testing the generator (It can considered a sample code for usage)
The code was written in python. Wanted to code it in C, but will write something to convert the torch tensor to ONNX for use with C when I'm satisfied with the model generator. The script support all types of tbamud world files (mobs, obj, wld, shp, trg, qst), but we have a small sample of qst files actually. But it can work with your on mud.
If you want to try it, it's available at
github.com/Forneck/gan.tbamud
It's included a fake.pt file with generated samples by Google Gemma model for training. Since generating fake samples take a lot of time, it's included.
For the memory usage, it's use DataSets and DataLoader to loads parts of the tensors and reduce the memory load so I can run the training for on my phone when idle. Any suggestions is welcome. I hope you all enjoy. Oh, btw, the code supports saving the trained model locally and on cloud on HuggingFace (needs an auth token -
not included)
Since I don't know if it's a Development, Snippet or Building topic (it may include all of these or none at all), I'm posting it here.
