mutual-ai / mxnet-tutorials-in-r

mxnet examples for people with intermediate R and deeplearning knowledge.

note that I do not intend to find the best model for a given problem, the purpose of this project is simply to give you an easy access to mxnet and its functionality.

dependencies:

• darch
• mxnet
• ggplot2
• reshape2

to use the code:

1. run preprocessing/get_mnist.R
2. run preprocessing/conversion.R
3. run models/

note that you have to edit the working directory in all files.

model overview

Model	Time to execute*	Type	Peak accuracy
01	roughly 1.3 min	very simple neural network with 3 layers	0.9784 @ 50 epochs
02	roughly 43.5 min	4 neural nets with 4 layers + dropout (benchmark)	best model yields 0.9793 @ 50 epochs
03	roughly 13.5 min	CNN with 3 conv/pooling + 3 dense layers	0.9862 accuracy @ 20 epochs

* models were executed on mainstream CPU with 4 cores/4 threads @ 3.9 GHZ

I'm planing to add more models in the future (autoencoder, lstms, segmentation models)

model 01 results:

model 02 results:

model 03 results: