Install

dependencies

$ git clone https://gitlab.com/antoinehonore/hq.git
$ git clone -b api https://gitlab.com/antoinehonore/gm_hmm.git

Interpreters

In hq/ and gm_hmm/ folders Run:

$ cd [FOLDER]
$ virtualenv -p python3 pyenv
$ . pyenv/bin/activate
$ python [FOLDER]/setup.py develop

The last line adds the folder containing hq/ and gm_hmm/ to PYTHONPATH.

Data

Place the train and test pickle files corresponding to 61 classes and 39 features and clean test with the names: train.feat0.pkl and test.feat0.pkl under exp/split\_c61f39clean/data. Follow the same naming notations for the rest of the possible datasets.

feat0 is not used here but might be necessary if we were to compute different sets of features.

Test

From hq/, run:

$ make model=gmmhmm splits=_c61f39clean feats=0

Advanced test

Once the previous test runs and gives results, we can try more advanced calls:

$ ./gmmhmm_submodels
$ ./gmmhmm_submodels.sh print
Models:  gmmhmm
Number of states (ns):  3 6 9
Number of iterations (niter):  2 10 20
Number of mixtures (nmix):  2 4 6 8 10 12
gmmhmm-ns\{3,6,9\}-niter\{2,10,20\}-nmix\{2,4,6,8,10,12\}

Copy the last line and use it in the make call

$ make gmmhmm-ns\{3,6,9\}-niter\{2,10,20\}-nmix\{2,4,6,8,10,12\} splits=_c61f39clean feats=0 -j 5

This allows to train all combinations of hyper parameters for gmmhmms on the data called c61f39clean. Once you have set up more split_ folders, you can run things like:

$ make gmmhmm-ns\{3,6,9\}-niter\{2,10,20\}-nmix\{2,4,6,8,10,12\} splits=_c61f\{39,13\}clean feats=0 -j 5

Acknowledge

This repos is forked from hq