Sea-NN is a fully on-chain convolutional neural network written in @seahorse_lang.
It was trained off-chain using the MNIST dataset, and then the model was uploaded on-chain. The model is quite large and takes over 200 transactions to upload (though this process could likely be optimized).
The model is largely based on the one described here In order to optimize the model for on-chain execution, several optimizations had to be performed 😛
- Every pixel is either 0 or 1, so a single bit can be used to represent it.
- All 512 possible 3x3 convolutions are precomputed and stored in a lookup table.
- Rather than representing weights as floats, signed integers in the range of [-2^15, 2^15] are used.
- All of the layers are optimized into a single loop, allowing the output to be computed in-place.
The entire contract is a single 73-line python file, enabled by Seahorse's brevity. Click the "View source" button to see it 😄
Created by @wireless_anon Repo is available here