KAN-GPT

The PyTorch implementation of Generative Pre-trained Transformers (GPTs) using Kolmogorov-Arnold Networks (KANs) for language modeling

Install it from PyPI

pip install kan_gpt

Usage

Refer to the KAN_GPT.ipynb and kan_gpt/prompt.py for usage examples. The following is an outine of how to use the model:

from kan_gpt.model import GPT
from transformers import GPT2Tokenizer

model_config = GPT.get_default_config()
model_config.model_type = "gpt2"
model_config.vocab_size = 50257
model_config.block_size = 1024
model = GPT(model_config)

tokenizer = GPT2Tokenizer.from_pretrained('gpt2')

prompt = "Bangalore is often described as the "

prompt_encoded = tokenizer.encode(
  text=prompt, add_special_tokens=False
)

x = torch.tensor(prompt_encoded).unsqueeze(0)

model.eval()
y = model.generate(x, 50)  # sample 50 tokens

result = tokenizer.decode(y)

print(result)

# Bangalore is often described as the Silicon Valley of India.
# The city has witnessed rapid growth in the past two decades.....

Setup for Development

# Download Repo
git clone https://github.com/AdityaNG/kan-gpt
cd kan-gpt
git pull

# Download Dataset
./scripts/download_webtext.sh
./scripts/download_tinyshakespeare.sh

# Install dependencies for development
pip install -r requirements.txt
pip install -e .

Train

Use the following dummy script to make sure everything is working as expected

WANDB_MODE=offline CUDA_VISIBLE_DEVICE="" python3 -m kan_gpt.train --architecture MLP --batch_size 1 --dummy_dataset --device cpu --max_iters 200
WANDB_MODE=offline CUDA_VISIBLE_DEVICE="" python3 -m kan_gpt.train --architecture KAN --batch_size 1 --dummy_dataset --device cpu --max_iters 200

Then make use of the training script

python -m kan_gpt.train

Prompt

You can prompt the model to produce text as follows

python -m kan_gpt.prompt --prompt "Bangalore is often described as the " --model_path (checkpoint)

Results

We train and compare KAN-GPT with an equivalent MLP-GPT model on the Tiny Shakespeare dataset. We observe that the KAN-GPT performs slightly better than the MLP-GPT. We are looking into further experiments to dive deeper. The results are shown below:

Metrics

TODOs

• Integrate minGPT and pykan
• Dataset downloading script for WebText
• PyTorch Dataset parser for WebText
• PyTorch Dataset parser for tinyshakespeare
• Mini training POC for KAN-GPT
  • Integrate KAN training logic from KAN.train_kan
  • Train a dummy batch w/o any memory issues
• Mini training POC for MLP-GPT
• Train MLP-GPT on the webtext dataset as a baseline
• Train KAN-GPT on the webtext dataset as a baseline
• Metrics comparing KAN-GPT and MLP-GPT
• Auto Save checkpoints
• Auto Save checkpoints to W&B
• Auto Download model weights from git / huggingface
• W&B hyperparam sweep script
• Script to load checkpoint in interactive mode
• Reduce requrements.txt constraints
• Define pydantic model for training and sweep args
• Pruning the package, get rid of unused code
• Training script to PyTorch Lighting
• Documentation: mkdocs gh-deploy
• Integrate with efficient-kan
• Test Cases
  • KAN: Forward-Backward test
  • GPT: Forward-Backward test
  • KAN_GPT: Forward-Backward test
  • EFFICIENT_KAN: Forward-Backward test

Development

Read the CONTRIBUTING.md file.

References

• minGPT
• pykan
• webtext
• tinyshakespeare