Machine Learning Experiments

This repository is a simple way to train an LSTM model on some input data. Right now, the only supported dataset is a set of Shakespeare plays. The next version will be able to support more flexible input formats and other datasets.

Setup

You'll need to install Python 3 and TensorFlow. Here is one way to do so using Anaconda:

$ wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
$ chmod a+x Miniconda3-latest-Linux-x86_64.sh
$ ./Miniconda3-latest-Linux-x86_64.sh
* Answer yes to everything.
$ export PATH="~/miniconda3/bin:$PATH"
$ conda create -n tensorflow python=3.5
$ source activate tensorflow
$ conda install pytest
$ pip install https://github.com/lakshayg/tensorflow-build/raw/master/tensorflow-1.4.0-cp35-cp35m-linux_x86_64.whl
* Note, your system may need a different build. See https://github.com/lakshayg/tensorflow-build for more info.

Now clone the repo and get the data:

$ git clone https://github.com/xinxinw1/ml-experiments.git
$ cd ml-experiments
$ cd data
$ wget http://cs.stanford.edu/people/karpathy/char-rnn/shakespeare_input.txt
$ cd ..

Training

Shakespeare

Training on this dataset takes around 8 hours on my computer.

$ ./train/shakespeare.py baseline

Test model

This trains a model on arrays where the contents are n 1's followed by n 0's where n is a random integer.

$ ./train/count-1-0.py baseline

Generate text

Shakespeare

$ python
> from models import lstm
> model = lstm.LSTMModelFromFile("shakespeare", tag="baseline")
> model.sample("T", 500)

Example output:

Three milks ill off, the whore of all our statues
Upon the world musicians stronger than a
man,--mether with all man's pies and springs: 'tis wonderful
hearing, latter it; that thou art valued withal.

STINUS:
Go seek him hither; now and ecstasy
Where it may stand like pinchad. Wither grandam
Of violent hand hath sent me out: 'tis done,
Who may win brain of citch me on those shot
Whilst you with a prenew music is it so?

BASTARD:
If thou be plainle, I'll plugued thee in a
clean with sign and thin

Note: This example was generated using a model with higher parameters than the ones in the repository by default, so your output may look a bit different.

Test model

$ python
> from models import lstm
> model = lstm.LSTMModelFromFile("count-1-0", tag="baseline")
> model.sample()

Example outputs:

[1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[1, 1, 0, 0]
[1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]
[1, 1, 1, 1, 0, 0, 0, 0]
[]

Run unit and integration tests

$ python -m pytest tests