To clone:
git clone https://github.com/neerbek/taboo-core.git
Runs python3(!)
Dependencies: numpy, theano, nltk, click, flask, matplotlib, tkinter
pip3 (or conda or apt-get or ...) install
Ex:
sudo apt-get install python3-pip
sudo apt-get install python3-tk # needed for tkinter
sudo -H pip3 install -r requirements.txt
Download glove word embeddings: (not necessary for running tests) (requires curl to be installed)
./download_glove.sh
This installs glove word embedding files in ../code/glove/
run: nltk.download(), load package punkt
ex:
python3 -m nltk.downloader punkt
To run tests: from taboo-core, say:
./run_tests.sh
OMP_NUM_THREADS=2 ipython3 functionality/train_model.py -- -traintrees 201/train_custom250_random.txt -validtrees 201/dev_custom250_random.txt -testtrees 201/test_custom250_random.txt -nx 50 -nh 100 -lr 0.01 -n_epochs 10 -glove_path ../code/glove/
OMP_NUM_THREADS=2 sets the maximum numbers of threads used. We are memory intensive so best performance is with lower number of threads.
ipython3 ipython3 or python3
functionality/train_model.py main code file to run (i.e. training)
-- ipython syntax for "rest of arguments are passed on to main routine"
-traintrees 201/train_custom250_random.txt a set of labeled parse-trees for training
-validtrees 201/dev_custom250_random.txt a set of labeled parse-trees for validation (dev set)
-testtrees 201/test_custom250_random.txt a set of labeled parse-trees for test
-nx 50 size of word embeddings
-nh 100 size of hidden state
-lr 0.01 learning rate
-n_epochs 10 number of iterations before stopping (-1 for infinity)
-glove_path ../code/glove/ path to location of word embeddings
There are many more options, run train_model.py without arguments to get the list
export TREES=path-to-zip-file-with-data/trees0.zip\$test.txt
export MODEL=path-to-trained-model
export GLOVEPATH=path-to-glove
OMP_NUM_THREADS=2 ipython3 functionality/run_model.py -- -inputtrees $TREES -inputmodel $MODEL -nx 100 -nh 100 -L1_reg 0 -L2_reg 0.0001 -retain_probabilities 0.9 -batch_size 1000 -glove_path $GLOVEPATH -random_seed 1234
To get a list of all embeddings generated in the final layer of a particular model
export TREES=path-to-zip-file-with-data/trees0.zip\$test.txt
export MODEL=path-to-trained-model
export GLOVEPATH=path-to-glove
OMP_NUM_THREADS=2 ipython3 functionality/run_model_verbose.py -- -inputtrees $TREES -inputmodel $MODEL -nx 100 -nh 100 -L1_reg 0 -L2_reg 0.0001 -retain_probability 0.9 -batch_size 1000 -glove_path $GLOVEPATH -random_seed 1234 -output_embeddings > output.txt