MLRAM

Auger model

1. Import dataset and Train to get model

2. Make sure review section is correct in config.yml

3. Deploy model. It will add model to review section in Auger.ai

4. Predict and send actuals. See actuals API

Monitored model

See application example: https://github.com/augerai/mlram_apps

1. Create A2ML application with external provider:

$ a2ml new test_app -p external

2. Specify the following parameters in config.yml:

target: the feature which is the target
model_type: Can be regression or classification

    experiment:
      metric: <metric to calculate using actuals>

    review:
    roi:
        <See configuration section>
            alert:
                    <See configuration section>

3. Deploy model without model id:

   ctx = Context()
   a2ml = A2ML(ctx)
   result = a2ml.deploy(model_id=None, name="My Monitored model.", algorithm="RandomForest", score=0.76)
   model_id = result['data']['model_id']
   

4. Send actuals:

   ctx = Context()
   actual_records = [['predicted_value_1', 'actual_value_1'], ['predicted_value_2', 'actual_value_2']]
   columns = [target, 'actual']
   
   A2ML(ctx, "external").actuals('external_model_id', data=actual_records,columns=columns)
   

To review distribution chart , send training features with target and actuals:

ctx = Context()
actual_records = [['predicted_value_1', 'actual_value_1', 'value1', 'value2'], ['predicted_value_2', 'actual_value_2', 'value3', 'value4']]
columns = [target, 'actual', 'feature1', 'feature2']

A2ML(ctx, "external").actuals('external_model_id', data=actual_records,columns=columns)

5. Call review to check if model retrain is required:

   # If call just after actuals, wait some time till server process the data
   time.sleep(30)
   
   ctx = Context()
   result = A2ML(ctx).review(model_id='external_model_id')
   if result['data']['status'] == 'retrain':
           #Train new model using updated data
           a2ml.deploy(model_id=None, name="My Monitored model.", algorithm="RandomForest", score=0.77)