cognet-generative 0.1.0

COGNET V1: Concept Routing Graph Neural Network

COGNET V1

COGNET is a Concept Routing Graph Neural Network designed for complex reasoning over text.

Installation

You can install COGNET as a Python package. From the root directory:

pip install -e .

Requirements

• Python >= 3.10
• PyTorch >= 2.0.0
• PyTorch Geometric >= 2.3.0
• Transformers, spaCy, datasets

You may need to download the spaCy model:

python -m spacy download en_core_web_sm

Running the Pipelines

1. Minimal Working Pipeline (Training Sanity Check)

To test that the full forward and backward pass works on a dummy dataset:

python -m cognet.train --dummy

2. Full Training

To run the full training loop with checkpointing:

python -m cognet.train --train_size 1000 --val_size 200 --epochs 5 --batch_size 16

3. Ablation Experiment

To verify that SCR routing adds value compared to standard pooling:

python -m cognet.experiment --dummy

or on the real dataset:

python -m cognet.experiment --train_size 500 --val_size 100 --epochs 5

Inference API Usage

You can use the generate method for simple yes/no text classification.

from cognet import CognetModel

# Load a trained model
# model = CognetModel.load("./checkpoints/cognet_best.pt")

# Or create a new one
model = CognetModel()

answer = model.generate("Explain gravity")
print(f"Prediction: {answer}")

Lightweight Deployment (FastAPI)

You can serve the model over an HTTP API using the provided FastAPI server.

1. Set the environment variable for your model checkpoint (optional):

export COGNET_MODEL_PATH="./checkpoints/cognet_best.pt"

2. Start the server:

uvicorn deployment.server:app --host 0.0.0.0 --port 8000

3. Make a request:

curl -X POST "http://127.0.0.1:8000/generate" \
     -H "Content-Type: application/json" \
     -d '{"text": "Is the sky blue?"}'