models

Models

Some checked-in models models, created with zkp-gravity/BTHOWeN-0g.

To update the models, follow the steps in the BTHOWeN-0g readme, then copy the models over:

cp ../BTHOWeN/software_model/models/MNIST/*.hdf5 models/

"MNIST-Tiny": model_28input_256entry_1hash_1bpi (k = 14)

A very small toy model, used in the tests and the benchmark. Accuracy on the MNIST test set is 83.06%.

"MNIST-Small": model_28input_1024entry_2hash_2bpi (k = 15)

This model is comparable to the "MNIST-Small" model in the BTHOWeN paper (see Table III). It has the same parameters, but uses the "MishMash" hash function ((x^3 % p) % 2^l).

Accuracy on the MNIST test set is 92.81% (compares to 93.4% in the paper).

"MNIST-Medium": model_28input_2048entry_2hash_3bpi (k = 15)

This model has the same parameters as the "MNIST-Medium" model in the BTHOWeN paper.

Accuracy on the MNIST test set is 93.95% (compares to 94.3% in the paper).

"MNIST-Large": model_49input_8192entry_4hash_6bpi (k = 17)

This model has the same parameters as the "MNIST-Large" model in the BTHOWeN paper.

Accuracy on the MNIST test set is 95.10% (compares to 95.2% in the paper).