Optimization of Serverless Platform with Theano

Current Version

1. Torch to Theano Converter
   
2. no implementation for Tensorflow and other framework

Converter Flow architecture

Example

1. Converter

from ml_inference.modeling import *
import numpy as np
import torch
hooking_dummy = torch.Tensor(np.random.rand(3,64,64))
weight_parser(dnn_model, 'torch', hooking_dummy)

2.restore model.

from ml_inference.modeling import *
model = NeuralNet('weights.h5')

Summary

1. The light package using theano with Scikit-Learn can upload to AWS Lambda.
2. It is slow than Pytorch as Theano need not setup G++ environment.
3. This library can't support Tensorflow or MXNet.

Contributor

Hyunjune Kim - email is '[email protected]' , You can call me Jey!
Kyungyong Lee - my professor is him, an assistant professor in KOOKMIN University.

Bigdata Lab in Kookmin University