---
language:
- en
license: other
tags:
- pytorch
- vision
- siren
- implicit-neural-representation
- touhou
- bad-apple
---

# BadApple × NN: SIREN Checkpoint

This repository contains the pre-trained weights for the **BadApple × NN** project. It is a 512-wide, 5-layer SineMLP (Implicit Neural Representation) that has mathematically compressed and memorized the 6,505 frames of the "Bad Apple!!" music video into ~1.05M parameters.

## Available Checkpoints

| File | Epoch | Loss | Notes |
|------|-------|------|-------|
| `epoch_590_loss_0.2248_(best).pt` | 590 | 0.2248 | **Best checkpoint** — lowest loss achieved |
| `epoch_600_loss_0.2255_(final).pt` | 600 | 0.2255 | Final checkpoint at end of training |
| `epoch_380_loss_0.2434_(best edges).pt` | 380 | 0.2434 | Previous best before bug fixes |
| `epoch_200_loss_0.0945_previous_best.pt` | 200 | 0.0945 | Best from the original 256-wide model (not comparable — different output format) |
| `epoch_287_loss_0.3438_latest.pt` | 287 | 0.3438 | Intermediate snapshot from the 512-wide model |

Training completed at 600 epochs. For the full training history see [TRAINING.md](https://github.com/Pratik-3706/BadApple_nn/blob/main/TRAINING.md) in the main repo.

## Usage
These weights are designed to be run directly with the real-time PyTorch streaming inference engine found in the main GitHub repository.

**GitHub Repository:** [Pratik-3706/BadApple_nn](https://github.com/Pratik-3706/BadApple_nn)

## License & Copyright Disclaimer
- **The Code / Architecture**: The structural code and training implementation are licensed under MIT.
- **The Model Weights**: The neural network weights provided here are a derivative mathematical representation of the original video. The "Bad Apple!!" video, audio, and underlying intellectual property belong to **ZUN / Team Shanghai Alice** and **Alstroemeria Records**. 
- These weights are provided strictly for **educational, non-commercial, and transformative AI research purposes** under Fair Use. You may not use this model for commercial purposes.