rkmtlab / liquidML

Machine learning server for interaction design

liquidML

Machine learning server for interaction design. LSTM by torch has been implemented as an example. back_server.py will response predicted parameters to POST requests. A POST request can be either of a teacher data or prediction requests.

Overview

Requirement

python >= 3.5.3
torch (0.2.0.post3)
urllib3 (1.22)
numpy (1.13.3)

Quick run

python back_server.py
run front_sample/front_sample.pde

Protocol (For current LSTM implementation)

A naive protocol between front and back has been set (pull requests welcomed).

x_t1,y_t1:x_t2,y_t2

means "Learn a model that predict (x_t2, y_t2) from (x_t1, y_t1)". For example:

0.23,0.54:0.12,0.54

The server stores teacher data to 'teacher.txt'. 'brain.py' start training when the teacher data reaches to enough data size.

The server is multithreaded so to response predictions while training the model.

Result

Screen capture of Processing. Predicting 5 steps (green) ahead of rotating point (red). 1000 teacher data were fed to LSTM Cell with 100 layers trained with 10 iterations.