Added mlflow model server #774
Conversation
|
Using Seldon storage instead of MLFLow storage as MLFLow storage doesn't support anonymous access to google buckets. |
|
Documentation has been added |
|
Samples have been added (also to the google bucket), but can only be tested properly until operator changes are added. |
|
Operator PR has been added via SeldonIO/seldon-operator#49 - waiting for that PR to be landed, and for the GRPC in this PR to be tested in order to land |
…ldon-core into mlflow_model_server
|
Added a full example using the MLFlow server which aligns to the blog post currently being written: https://github.com/axsaucedo/seldon-core/tree/mlflow_model_server/examples/models/mlflow_server_ab_test_ambassador. Looks so clean! |
| # df = pd.DataFrame(data=X, columns=feature_names) | ||
| #else: | ||
| # df = pd.DataFrame(data=X) | ||
| result = self._model.predict(X) |
There was a problem hiding this comment.
Is this contracting to the statement that
* The input to the model is set to be pandas by default, so the numpy array passed will be converted into a pandas dataframe
in doc/source/servers/mlflow.md?
There was a problem hiding this comment.
The results should also be converted back to numpy array.
| # from the actual python wrapper. Raise exception instead | ||
| #raise requests.HTTPError("Model not loaded yet") | ||
| #if not feature_names is None and len(feature_names)>0: | ||
| # df = pd.DataFrame(data=X, columns=feature_names) |
There was a problem hiding this comment.
Is there a reason this pandas conversion was commented out?
There was a problem hiding this comment.
@cliveseldon is there any other to send a dataframe into a served MLFlow model? because this sort of breaks their expected API. maybe i'm missing something though..
|
@phsiao @lennon310 the image that was pushed to DockerHub doesn't have the Pandas Dataframe commented out so it would still work as expected in the documentation, which is why I must have missed it - thanks for the heads up! We'll submit a PR right away 👍 |
This PR contains an MLFlow model server. Key things that needs to be confirmed before landing include:
Current issues