I think this post is supposed to be written in English since it’s not a literature review.
I’ve now been working at NLPR, CASIA for more than 1 month and a half, working on a particular topic about Medical Data Processing. Here’s a summary of all DL frameworks and particular modules I’ve been tried. I will keep updating this post.
The programming platform I’ve used are MATLAB, Lua and Python. Currently I’m working using the last one.
Modules working on MATLAB
LSTM with Batching Learning Applied. This module is an implementation of LSTM with Batch Learning method to accelerate its learning process. If you roughly read the code, you will find its sophisticated support on big data due to Batch Learning and GPU. However, without GPU acceleration it will be relatively slow. Besides, its source code is a lil-bit hard to understand…
Modules working on Lua
NNgraph from Torch. This one is the one used for Oxford's Machine Learning course. However, due to the fact that its developing community is quite small and it’s not that flexible, I abandoned this one very quickly - It’s actually a very good framework to teach students but not a good choice in research excepted that you’re good at Lua and C and wanna work on back-end.
Modules working on Python
Theano-LSTM. This one is a nano-sized Theano module concerning RNN and LSTM. I’ve read the source code and impressed by its solid connectivity and completeness. However, I’m actually now a good Theano user - Unlike Caffe, Theano is essentially not a DL framework but a Scientific Computation one with many self-defined computational methods. Besides, I’m now working my project on Linux, which, although has good programming environments, is a lil-bit hard to handle and configure Theano-related bugs. Therefore, I decided to used other framework which is easier to handle and program on the front-end.
Scikit-Learn. I think this package is more familiar to Python programmers - I would say this is the current ‘chosen one’ in Machine Learning with Python implementations. It’s still maintained and developed by INRIA and Google along with other stars in the community. I plan to use this to implement a solution, integrated with the LSTM framework.
Current Viewer
Keras. Seems like a good one. It’s built as a front-end of Theano and Tensorflow. Working on it.
Wepe's ML. Helping me understanding Keras as well as other ML algorithms. Forked in my Github.
[Keep Updating]