GenClimb

Try it out at genclimb.com)

GenClimb is a generative AI model designed to create climbing routes for Standardized Interactive Climbing Training Boards (SICTBs). Utilizing a seq2seq transformer architecture, GenClimb generates climbs conditioned on specific board layouts and difficulty levels.

Currently, only Kilter Boards are supported.

Inferencing

GenClimb performs inferencing directly in your browser, using WebAssembly. This approach offers several benefits:

• Privacy: Your data stays on your device, with no information sent to external servers.
• Performance: Utilizes ONNX Runtime Web for efficient, on-device inference.
• Edge Computing: Brings AI directly to your device, eliminating the need for server-side processing.

Model Details: GenClimb-Tiny-Quantized

The model is quantized to int8 for improved efficiency.

Architecture Parameters

• Dimension: 256
• Number of Heads: 4
• Number of Layers: 2
• Feed-Forward Dimension: 256
• Dropout: 0.3
• Activation Function: GELU
• Layer Normalization Epsilon: 1e-5

Training Configuration

• Device: CUDA
• Learning Rate: 1e-4
• Epochs: 8
• Weight Decay: 0.015
• Batch Size: 64
• Train/Test Split: 90/10

Resources

• GenClimb-Large-Quantized Model (Hugging Face)
• Curated Kilter Board Dataset

Acknowledgements

Special thanks to lemeryfertitta for the BoardLib utility, which was instrumental for getting the data.

Future Improvements

1. Integration with Kilter Account to save climbs.
2. Improved inference efficiency through the use of KV caching and encoder caching.

Run Locally

1. Clone this repository:

git clone https://github.com/mstafam/GenClimb.git

2. Navigate to the project directory:

cd GenClimb

3. Install dependencies:

npm install

4. Start the application:

npm start

The app should launch at http://localhost:3000.