---
tags:
- MuseMachine
- pytorch
- generative-adversarial-network
- variational-autoencoder
---

# MuseMachine

A modular AE-GAN pipeline for portrait generation.

## Architecture

| Component | Description |
|---|---|
| Encoder | VAE encoder producing latent representations |
| Generator | Decodes latent vectors into coarse images |
| Refiner | Iteratively refines generations over 4 steps |
| Discriminator | PatchGAN-style discriminator |

## Hyperparameters

```json
{
  "image_size": 256,
  "latent_dim": 256,
  "batch_size": 4,
  "epochs": 300,
  "lr_g": 0.0002,
  "lr_d": 5e-05,
  "alpha_recon": 100.0,
  "beta_adv": 1.0,
  "beta_kl": 0.05,
  "noise_std": 0.15,
  "refinement_steps": 4,
  "refiner_step_size": 0.08,
  "refine_decay": 0.8,
  "dataset_name": "ajehsmihba/aesthetic-female-portraits",
  "outputs_dir": "outputs",
  "models_dir": "models",
  "save_every": 50,
  "push_to_hf": true,
  "hf_repo_id": "ajehsmihba/MuseMachine",
  "grad_clip": 1.0,
  "vis_every": 10
}
```

## Best Metrics

```json
{
  "D": 1.6232355684041977,
  "EG": 55.429039001464844,
  "KL": 142.61621220906576,
  "Recon": 0.12780248870452246,
  "Coarse_Recon": 0.12812133505940437
}
```

## Resume Training

```python
from huggingface_hub import hf_hub_download
import torch

ckpt_path = hf_hub_download(repo_id="ajehsmihba/MuseMachine", filename="best_checkpoint.pth")
ckpt = torch.load(ckpt_path, map_location="cpu")
```