Support kwargs in predict() and predict_proba() #113
Conversation
| @@ -704,6 +702,7 @@ def predict( | |||
| instance_count: int = 1, | |||
| custom_image_uri: Optional[str] = None, | |||
| wait: bool = True, | |||
| inference_kwargs: Optional[Dict[str, Any]] = None, | |||
There was a problem hiding this comment.
Let's add the new parameter in the doctring
| @@ -265,6 +267,10 @@ def run( | |||
| else: | |||
| model_cls = AutoGluonNonRepackInferenceModel | |||
| logger.log(20, "Creating inference model...") | |||
| inference_kwargs_str = json.dumps(inference_kwargs) if inference_kwargs is not None else None | |||
There was a problem hiding this comment.
Can we support complex payload in the predict args such as known_covariates from TimeSeriesPredictor
There was a problem hiding this comment.
We might need to handle DataFrame specifically. I can add that to the PR.
| cols.insert(0, cols.pop(id_index)) | ||
| if target is not None: | ||
| target_index = cols.index(target) | ||
| cols.append(cols.pop(target_index)) | ||
| data = data[cols] | ||
|
|
||
| if static_features is not None: |
There was a problem hiding this comment.
Making id_column an optional argument might break the pd.merge in line 45.
I think that keeping the id_column, timestamp_column and target as part of the TimeSeriesSagemakerBackend API is fine since this class is not user-facing. In the public API of the TimeSeriesCloudPredictor these arguments are optional.
| id_column: str = "item_id", | ||
| timestamp_column: str = "timestamp", | ||
| id_column: Optional[str] = None, | ||
| timestamp_column: Optional[str] = None, |
There was a problem hiding this comment.
What is our motivation for changing the defaults here?
There was a problem hiding this comment.
I was trying to make the .predict() and .predict_real_time() API align with what we have in the Chronos tutorial. I see that we might not have id_column and timestamp_column in the train_data, but please correct me if I misunderstood the example.
| @@ -175,13 +178,18 @@ def predict_real_time( | |||
| Pandas.DataFrame | |||
| Predict results in DataFrame | |||
| """ | |||
| self.id_column = id_column or self.id_column | |||
| self.timestamp_column = timestamp_column or self.timestamp_column | |||
| self.target_column = target or self.target_column | |||
There was a problem hiding this comment.
What will happen if both self.target_column is None and target is None?
There was a problem hiding this comment.
I think the TabularPredictor API does not require target at the moment, that's why I made it optional. See this tutorial, e.g.
predictor = TimeSeriesPredictor(prediction_length=14).fit(train_data)
It has been handled https://github.com/autogluon/autogluon/blob/bda6174f4a1fb8398aef4f375d9eacfd29bb46d9/timeseries/src/autogluon/timeseries/predictor.py#L179
| @@ -805,6 +805,7 @@ def predict_proba( | |||
| instance_count: int = 1, | |||
| custom_image_uri: Optional[str] = None, | |||
| wait: bool = True, | |||
| inference_kwargs: Optional[Dict[str, Any]] = None, | |||
| @@ -556,6 +556,7 @@ def predict( | |||
| custom_image_uri: Optional[str] = None, | |||
| wait: bool = True, | |||
| backend_kwargs: Optional[Dict] = None, | |||
| **kwargs, | |||
There was a problem hiding this comment.
Add example code should be added to tutorials that showcase specifying kwargs. Otherwise it will be hard for users to realize how to do this.
There was a problem hiding this comment.
Good idea. I will add some tutorials with this PR.
| @@ -648,6 +650,7 @@ def predict_proba( | |||
| custom_image_uri: Optional[str] = None, | |||
| wait: bool = True, | |||
| backend_kwargs: Optional[Dict] = None, | |||
| **kwargs, | |||
| @@ -199,6 +210,7 @@ def predict( | |||
| custom_image_uri: Optional[str] = None, | |||
| wait: bool = True, | |||
| backend_kwargs: Optional[Dict] = None, | |||
| **kwargs, | |||
Issue #, if available:
Description of changes:
By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.