feat(JAQPOT-391): support two onnx models (#47)

* feat: add preprocessor sklearn_post_handler in predict_onnx * feat: support preprocessing through onnx * feat: support doa and preprocessing * fix: access doa data * refactor: sklearn_post_handler * refactor: predict_onnx
ntua-unit-of-control-and-informatics · Nov 4, 2024 · 7cafd3a · 7cafd3a
1 parent fa2c538
commit 7cafd3a
Show file tree

Hide file tree

Showing 2 changed files with 67 additions and 68 deletions.
diff --git a/src/handlers/predict_sklearn.py b/src/handlers/predict_sklearn.py
@@ -7,30 +7,32 @@
 
 def sklearn_post_handler(request: PredictionRequest) -> PredictionResponse:
     model = model_decoder.decode(request.model.raw_model)
+    preprocessor = (
+        model_decoder.decode(request.model.raw_preprocessor)
+        if request.model.raw_preprocessor
+        else None
+    )
     data_entry_all, jaqpot_row_ids = json_to_predreq.decode(request)
-    prediction, doa_predictions = predict_onnx(model, data_entry_all, request)
-    task = request.model.task.lower()
-    if task == "binary_classification" or task == "multiclass_classification":
-        probabilities = predict_proba_onnx(model, data_entry_all, request)
-    else:
-        probabilities = [None for _ in range(len(prediction))]
+    predicted_values, probabilities, doa_predictions = predict_onnx(
+        model, preprocessor, data_entry_all, request
+    )
 
     predictions = []
     for jaqpot_row_id in jaqpot_row_ids:
         if len(request.model.dependent_features) == 1:
-            prediction = prediction.reshape(-1, 1)
+            predicted_values = predicted_values.reshape(-1, 1)
         jaqpot_row_id = int(jaqpot_row_id)
         results = {
-            feature.key: int(prediction[jaqpot_row_id, i])
+            feature.key: int(predicted_values[jaqpot_row_id, i])
             if isinstance(
-                prediction[jaqpot_row_id, i],
+                predicted_values[jaqpot_row_id, i],
                 (np.int16, np.int32, np.int64, np.longlong),
             )
-            else float(prediction[jaqpot_row_id, i])
+            else float(predicted_values[jaqpot_row_id, i])
             if isinstance(
-                prediction[jaqpot_row_id, i], (np.float16, np.float32, np.float64)
+                predicted_values[jaqpot_row_id, i], (np.float16, np.float32, np.float64)
             )
-            else prediction[jaqpot_row_id, i]
+            else predicted_values[jaqpot_row_id, i]
             for i, feature in enumerate(request.model.dependent_features)
         }
         results["jaqpotMetadata"] = {

diff --git a/src/helpers/predict_methods.py b/src/helpers/predict_methods.py
@@ -26,14 +26,14 @@ def calculate_doas(input_feed, request):
         for doa_data in request.model.doas:
             if doa_data.method == "LEVERAGE":
                 doa_method = Leverage()
-                doa_method.h_star = doa_data.data.hStar
-                doa_method.doa_matrix = doa_data.data.doaMatrix
+                doa_method.h_star = doa_data.data["hStar"]
+                doa_method.doa_matrix = doa_data.data["doaMatrix"]
             elif doa_data.method == "BOUNDING_BOX":
                 doa_method = BoundingBox()
-                doa_method.bounding_box = doa_data.data.boundingBox
+                doa_method.bounding_box = doa_data.data["boundingBox"]
             elif doa_data.method == "MEAN_VAR":
                 doa_method = MeanVar()
-                doa_method.bounds = doa_data.data.bounds
+                doa_method.bounds = doa_data.data["bounds"]
             doa_instance_prediction[doa_method.__name__] = doa_method.predict(
                 pd.DataFrame(data_instance.values.reshape(1, -1))
             )[0]
@@ -51,7 +51,7 @@ def calculate_doas(input_feed, request):
     return doas_results
 
 
-def predict_onnx(model, dataset: JaqpotpyDataset, request):
+def predict_onnx(model, preprocessor, dataset: JaqpotpyDataset, request):
     """
     Perform prediction using an ONNX model.
     Parameters:
@@ -73,27 +73,44 @@ def predict_onnx(model, dataset: JaqpotpyDataset, request):
     - The request dictionary should contain the necessary configuration for preprocessors and DOAS.
     """
 
-    sess = InferenceSession(model.SerializeToString())
+    if preprocessor:
+        onnx_graph = preprocessor
+    else:
+        onnx_graph = model
+
+    # prepare initial types for preprocessing
     input_feed = {}
-    for independent_feature in model.graph.input:
+    for independent_feature in onnx_graph.graph.input:
         np_dtype = onnx.helper.tensor_dtype_to_np_dtype(
             independent_feature.type.tensor_type.elem_type
         )
-
-        if len(model.graph.input) == 1:
+        if len(onnx_graph.graph.input) == 1:
             input_feed[independent_feature.name] = dataset.X.values.astype(np_dtype)
         else:
             input_feed[independent_feature.name] = (
                 dataset.X[independent_feature.name]
                 .values.astype(np_dtype)
                 .reshape(-1, 1)
             )
+    if preprocessor:
+        preprocessor_session = InferenceSession(preprocessor.SerializeToString())
+        input_feed = {"input": preprocessor_session.run(None, input_feed)[0]}
+
     if request.model.doas:
         doas_results = calculate_doas(input_feed, request)
     else:
         doas_results = None
 
-    onnx_prediction = sess.run(None, input_feed)[0]
+    model_session = InferenceSession(model.SerializeToString())
+    for independent_feature in model.graph.input:
+        np_dtype = onnx.helper.tensor_dtype_to_np_dtype(
+            independent_feature.type.tensor_type.elem_type
+        )
+
+    input_feed = {
+        model_session.get_inputs()[0].name: input_feed["input"].astype(np_dtype)
+    }
+    onnx_prediction = model_session.run(None, input_feed)
 
     if request.model.extra_config["preprocessors"]:
         for i in reversed(range(len(request.model.extra_config["preprocessors"]))):
@@ -108,56 +125,36 @@ def predict_onnx(model, dataset: JaqpotpyDataset, request):
                 len(request.model.dependent_features) == 1
                 and preprocessor_name != "LabelEncoder"
             ):
-                onnx_prediction = preprocessor_recreated.inverse_transform(
-                    onnx_prediction.reshape(-1, 1)
+                onnx_prediction[0] = preprocessor_recreated.inverse_transform(
+                    onnx_prediction[0].reshape(-1, 1)
                 )
-            onnx_prediction = preprocessor_recreated.inverse_transform(onnx_prediction)
+            onnx_prediction[0] = preprocessor_recreated.inverse_transform(
+                onnx_prediction[0]
+            )
 
     if len(request.model.dependent_features) == 1:
-        onnx_prediction = onnx_prediction.flatten()
-
-    return onnx_prediction, doas_results
-
-
-def predict_proba_onnx(model, dataset: JaqpotpyDataset, request):
-    """
-    Predict the probability estimates for a given dataset using an ONNX model.
-    Parameters:
-    model (onnx.ModelProto): The ONNX model used for prediction.
-    dataset (JaqpotpyDataset): The dataset containing the features for prediction.
-    request (dict): A dictionary containing additional request information, including model configuration.
-    Returns:
-    list: A list of dictionaries where each dictionary contains the predicted probabilities for each class,
-          with class labels as keys and rounded probability values as values.
-    """
+        onnx_prediction[0] = onnx_prediction[0].flatten()
 
-    sess = InferenceSession(model.SerializeToString())
-    input_feed = {}
-    for independent_feature in model.graph.input:
-        np_dtype = onnx.helper.tensor_dtype_to_np_dtype(
-            independent_feature.type.tensor_type.elem_type
-        )
-        if len(model.graph.input) == 1:
-            input_feed[independent_feature.name] = dataset.X.values.astype(np_dtype)
-        else:
-            input_feed[independent_feature.name] = (
-                dataset.X[independent_feature.name]
-                .values.astype(np_dtype)
-                .reshape(-1, 1)
-            )
-    onnx_probs = sess.run(None, input_feed)
+    # Probabilities estimation
     probs_list = []
-    for instance in onnx_probs[1]:
-        rounded_instance = {k: round(v, 3) for k, v in instance.items()}
-        if (
-            request.model.extra_config["preprocessors"]
-            and request.model.extra_config["preprocessors"][0]["name"] == "LabelEncoder"
-        ):
-            labels = request.model.extra_config["preprocessors"][0]["config"][
-                "classes_"
-            ]
-            rounded_instance = {labels[k]: v for k, v in rounded_instance.items()}
+    if request.model.task.lower() in [
+        "binary_classification",
+        "multiclass_classification",
+    ]:
+        for instance in onnx_prediction[1]:
+            rounded_instance = {k: round(v, 3) for k, v in instance.items()}
+            if (
+                request.model.extra_config["preprocessors"]
+                and request.model.extra_config["preprocessors"][0]["name"]
+                == "LabelEncoder"
+            ):
+                labels = request.model.extra_config["preprocessors"][0]["config"][
+                    "classes_"
+                ]
+                rounded_instance = {labels[k]: v for k, v in rounded_instance.items()}
 
-        probs_list.append(rounded_instance)
+            probs_list.append(rounded_instance)
+    else:
+        probs_list = [None for _ in range(len(onnx_prediction[0]))]
 
-    return probs_list
+    return onnx_prediction[0], probs_list, doas_results