Skip to content

Latest commit

 

History

History
51 lines (41 loc) · 2.38 KB

Adding a New Backend Model.md

File metadata and controls

51 lines (41 loc) · 2.38 KB
Raw

CMU Linguistic Annotation Backend -- How to add a new BackEnd Model

In this example we will add a new back-end model. We will use the VAD model silence (which, if you notice, is not readily available).

The first step is to create a subdirectory under annotator/BackendModels where the model's code will be stored:

mkdir -p annotator/BackendModels/silence
# add your code here -- (if you cloned the repo you should already have the code here)
# Make sure to add an __init__.py file
touch annotator/BackendModels/silence/__init__.py

With the code all set up and ready to be used, go to annotator/BackendModels/MLModels.py and create a model class for your new model:

class SilenceModel(MLModel):
    def __init__(self):
    	# Give a descriptive name to your model
        self.name = "simple VAD model" 
        # This will be useful when we have multiple versions of the same model
        self.trained_on_date = "August 2019" 
        # This will store the output of the model for a given segment
        self.output = '' 

    def get_results(self, input_file):
    	# You'll have to implement this
    	# Make sure to store the output in self.output
    	<your code here>
    	self.output =

We provide the simple implementation of this in annotator/BackendModels/MLModels-silence_added.py

With this implemented (do cp annotator/BackendModels/MLModels-silence_added.py annotator/BackendModels/MLModels.py) you can programmatically call it through a client. For example, take a look at example-clients/cmulab_elan.py which uses the added VAD model.

e.g. running the following should produce two new .eaf files under example-clients/output, which should have a new VAD tier.

cd example-clients
python cmulab_elan.py --input_dir Chatino/ --input_tiers en --output_dir output --output_tier VAD --model_name vad

If you navigate to http://localhost:8000/annotator/corpus/, you should see a new corpus named test_corpus with two segments. Each segment should include an AudioAnnotation with a pointer to the wavfile, SpanTextAnnotations that match the en tier of the original ELAN files, as well as the new TextAnnotation (with field_name="vad") with the VAD borders.

You can also add an entry for this model in the populate.py script, e.g. by adding

model4=Mlmodel(name="VAD_model", modelTrainingSpec="silence", status='ready', tags=Mlmodel.VAD)
model4.save()