Open beenotung opened 3 months ago
@beenotung, well, this package was created quite a long time ago. It probably makes sense to switch to Tensorflow.js. It will be easier to expand it with different primitives as conv, lstm layers and also implement VAE
It seems better if we wrap the serialization and deserialization methods instead of directly exposing those from the encoder and decoder, since we also need to store the min/max range and other options.
@zemlyansky Thanks for the response, I actually switched from tensorflow to this library because tensorflow is too heavy.
I also used brain.js before but its dependency gl failed to be installed in recent environment.
When the computation task is not too heavy, I prefer the simplicity of node.js (without C/Python/GPU bindings)
I guess we need to dig into the methods in adnn.
For some background, I'm writing typescript wrapper on top of autoencoder [ref 1] and adnn [ref 2]
ref 1: https://github.com/beenotung/auto-encoder.ts ref 2: WIP