Examples for custom prediction in API Designer

Thanks for the reply, but doesn't seem to help. There isn't any useful code examples anywhere. The person in that thread also wasn't able to figure it out.

I'm not even trying to use a custom model. I'm attempting to use an existing model trained in my flow and create a custom api that has the exact same behavior as the "Prediction model" api.

Is this even possible with dataiku?

This thread is the thread I was referencing in my original post and getting the same error.

Thanks for bearing with me. Maybe I'm blind but I cant seem to find where it explains "how to make a Visual ML model available in an API service". It seems like the only way to use Visual ML model is through a Prediction endpoint which does not allow custom code.

On the concepts page:

The API node supports 7 kinds of endpoints:

• The Prediction or Clustering endpoint to predict or cluster using models created with the DSS Visual Machine Learning component.
• The Python prediction endpoint to perform predictions using a custom model developed in Python
• The MLflow Prediction endpoint to predict using imported MLflow models
• The R prediction endpoint to perform predictions using a custom model developed in R
• The Python function endpoint to call specific functions developed in Python
• The R function endpoint to call specific functions developed in R
• The SQL query endpoint to perform parametrized SQL queries
• The Dataset lookup endpoint to perform data lookups in one or more DSS datasets

Doesn't look like there is an endpoint for using DSS Visual ML in Python.

It sounds like I will have to recreate the Visual ML model in my own python environment and upload the model to a managed folder and then maybe I'll be able to get that model through the custom python prediction endpoint.

I see. Maybe my use case is wrong for the application. Basically, before using dataiku, I have manually trained a RandomForest classifier on a previous months of data. I have also spun a REST api that loads that model and inferences a field. Instead of a singular prediction, we send back the highest 3 probable predictions. This is the custom code I want to add on.

But my impression was that dataiku could automate the fetching, training, and api deployment all in one flow. Which would be a huge value add since I could automate the training of a different month's data or create multiple models for different groups.

So it sounds like using a Designer node for the api is not a good pattern. Sorry, I'm still new newbie. But the model created from the training recipe looks pretty good.

Thank you! Sounds like a good suggestion, I will give it a try. Translating to English for others.

def predict(self, features_df):        

  project = dataiku.Project("My Project")        

  model = dataiku.Model("My Model")        

  predictor = model.get_predictor(features_df, with_input_cols=True, with_probas=True)        

  predictions = predictor.predict(features_df)        

return predictions

The above code will not work properly.

The method I tried may not be what you want, but I will share it.

1. Create an API with a model called "My Model" that appears to be a model created in dataiku as a prediction model.
2. Request a query to the end point deployed in def predict. Since this is only a single inference, you may need to repeat it with a for loop.
   Then, you will be able to get the predictions you want.

The above code is an example of implementation.

from dataiku.apinode import utils




def predict(self, features_df): 

  for feature in features_df : # df를 한줄씩 처리

        client = utils.get_self_client()


        result = client.predict_record("My Model_endpoint", {"feature1" : "value1", "feature2" : 42})


        predictions.append(result)

  predictions

return predictions

The code above may not be accurate, but if you use this concept, you will be able to receive the results you want.

from dataiku.apinode import utils 를 사용해서 했지만 직접 end-point url 에 쿼리요청하여 결과를 받아 보실 수도있으 실 겁니다.

import requestsurl = "https://앤드인트url"data = {"name": "John", "age": 30}response = requests.post(url, data=data)