A structured Python package (multi_gan) for exploring Generative Adversarial Networks across distinct environmental profiles including synthesized multi-modal clusters and complex grid imagery (e.g. mnist, cifar10).
This repository utilizes PyTorch to run modular experiments comparing standalone and multiple Generator/Discriminator couplings (Multi-GAN). Deep network variations like GeneratorDCGAN28 and DiscriminatorResNet32 are exposed dynamically via Sacred experiments depending on the dataset provided.
This project was initially coupled tightly to Google Colab environments via hardcoded outputs (e.g. /content/drive/...). It has been explicitly decoupled to establish dynamic local resolution, protecting execution flows across standalone PCs and generalized Cloud Virtual Machines.
- Removed
Run.ipynb: The Google environment orchestrator has been purged. - Added
run_pipeline.ipynb: Integrates the library organically (viapip install -e .) and fires the Sacred orchestration experimentrun_training.pysequentially without reliance on specialized magic commands. Large artifacts will neatly dump to a generateddata/output/multi_gan/structure securely nestled within your filesystem.
Ensure your Python environment meets the dependencies defined in setup.py and standard deep learning requisites:
- Torch
- Torchvision
- Sacred
- Joblib
To register multi_gan as a module and fulfill setup.py:
pip install -e .Run the default pipeline via the new Orchestrator notebook (run_pipeline.ipynb) or invoke the Python script directly via CLI:
python run_training.pyrun_training.py provides several @exp.named_config decorators (e.g., synthetic(), cifar(), mnist()). You can override the default parameters at runtime via Sacred CLI conventions:
python run_training.py with cifar batch_size=64 G_lr=1e-4