m-decoster / DeepHand-PyTorch

PyTorch implementation of Deep Hand (inference only)

DeepHand-PyTorch

PyTorch implementation of Deep Hand (inference only). Based on the TensorFlow implementation.

Requirements

• Python 3.8
• PyTorch 1.8
• NumPy 1.20

Running inference

You can run inference by using the evaluate_test.py script. It requires a single argument, which is the path to the test folder of the One Million Hands dataset.

1. Download the One Million Hands test set
2. Run python evaluate_test.py /path/to/test/

Conversion code

I've also included the original NumPy weights for the TensorFlow model and the script I used to convert the weights to PyTorch. This is in the convert_weights.py script.

Because this is only intended for inference, the auxiliary loss weights loss1, loss2, are not converted to PyTorch.

This script saves the model as a .pth PyTorch checkpoint file.