dmlc
The interface to training model
Currently we can only use the models dumped from mxnet main project. Can you please provide an interface to training or improve the existing model in Javascript project? Then we can start to do a lot of interesting works. Any documents or code would be appreciated.
Currently this is not on the roadmap, mainly because training API was broader and harder to expose, do you have anything particular in mind about what was wanted?
If training interface is missing from mxnetjs, we have to write in another language to improve the model and re-dump it.
For example, this is my avatar icon: https://avatars1.githubusercontent.com/u/16156781?v=3&s=460 The top class is joystick: Top-1: n03602883 joystick, value=0.4819868206977844, time-cost=7.963secs
The result is wrong and I want to adjust it. What I can do, is to download mxnet main project, training, re-dump the model and restart mxnetjs project. This is not convenient and hard to automate the process (even you can do it, it will be very slow). So I want to directly call an interface to improve the model till "Doraemon" is on the top.
I understand the training API is hard to expose. So we really don't need the training API and if an interface of Javascript can be made to access training API, the problem can be solved.
It will be appreciated even document can be provided so coding work can be done by others.