squash commit

Author: leeclemnet

.github/workflows/integration_tests_workflows_x86.yml

@@ -63,5 +63,7 @@ jobs:
           python -m pip install --upgrade pip
           pip install --upgrade setuptools
           pip install --extra-index-url https://download.pytorch.org/whl/cpu -r requirements/_requirements.txt -r requirements/requirements.sam.txt -r requirements/requirements.cpu.txt -r requirements/requirements.http.txt -r requirements/requirements.test.unit.txt -r requirements/requirements.doctr.txt -r requirements/requirements.yolo_world.txt -r requirements/requirements.transformers.txt -r requirements/requirements.sdk.http.txt -r requirements/requirements.easyocr.txt
+          # Test against source build of inference_models, not pinned PyPI version
+          pip install --force-reinstall --no-deps ./inference_models
       - name: 🧪 Integration Tests of Workflows
         run: USE_INFERENCE_MODELS=${{ matrix.use-inference-models }} ROBOFLOW_API_KEY=${{ secrets.API_KEY }} MODAL_TOKEN_ID=${{ secrets.MODAL_TOKEN_ID }} MODAL_TOKEN_SECRET=${{ secrets.MODAL_TOKEN_SECRET }} MODAL_ALLOW_ANONYMOUS_EXECUTION=True MODAL_ANONYMOUS_WORKSPACE_NAME=github-test MODAL_SKIP_FLORENCE2_TEST=FALSE LOAD_ENTERPRISE_BLOCKS=TRUE python -m pytest tests/workflows/integration_tests

inference/core/entities/requests/inference.py

@@ -146,9 +146,14 @@ class ObjectDetectionInferenceRequest(CVInferenceRequest):
         description="If provided, only predictions for the listed classes will be returned",
     )
     confidence: Optional[float] = Field(
-        default=0.4,
+        default=None,
         examples=[0.5],
-        description="The confidence threshold used to filter out predictions",
+        description=(
+            "The confidence threshold used to filter out predictions. If omitted, "
+            "the server uses the model's F1-optimal threshold from model evaluation "
+            "when available, otherwise falls back to 0.4. Pass an explicit value to "
+            "override both."
+        ),
     )
     fix_batch_size: Optional[bool] = Field(
         default=False,

inference/core/models/classification_base.py

@@ -1,6 +1,6 @@
 from io import BytesIO
 from time import perf_counter
-from typing import Any, List, Tuple, Union
+from typing import Any, List, Optional, Tuple, Union
 
 import numpy as np
 
@@ -299,7 +299,7 @@ def make_response(
         self,
         predictions,
         img_dims,
-        confidence: float = 0.5,
+        confidence: Optional[float] = 0.5,
         **kwargs,
     ) -> Union[ClassificationInferenceResponse, List[ClassificationInferenceResponse]]:
         """
@@ -320,6 +320,8 @@ def make_response(
             - Predictions below the confidence threshold are filtered out.
         """
         responses = []
+        if confidence is None:
+            confidence = 0.5
         confidence_threshold = float(confidence)
         for ind, prediction in enumerate(predictions):
             if self.multiclass:

inference/core/models/inference_models_adapters.py

@@ -677,25 +677,29 @@ def postprocess(
         List[ClassificationInferenceResponse],
     ]:
         mapped_kwargs = self.map_inference_kwargs(kwargs)
-        post_processed_predictions = self._model.post_process(
-            predictions, **mapped_kwargs
-        )
-        if isinstance(post_processed_predictions, list):
-            # multi-label classification
-            return prepare_multi_label_classification_response(
-                post_processed_predictions,
-                image_sizes=returned_metadata,
-                class_names=self.class_names,
-                confidence_threshold=kwargs.get("confidence", 0.5),
+        if isinstance(self._model, MultiLabelClassificationModel):
+            post_processed_predictions = self._model.post_process(
+                predictions, **mapped_kwargs
             )
-        else:
-            # single-label classification
-            return prepare_classification_response(
+            return prepare_multi_label_classification_response(
                 post_processed_predictions,
                 image_sizes=returned_metadata,
                 class_names=self.class_names,
-                confidence_threshold=kwargs.get("confidence", 0.5),
             )
+        # Single-label classification: top-1 always wins regardless of
+        # confidence, so per-class refinement isn't meaningful here. The base
+        # class deliberately opts out of recommendedParameters entirely. The
+        # response builder still uses kwargs.get("confidence", 0.5) for the
+        # cutoff that decides which alternative classes show up.
+        post_processed_predictions = self._model.post_process(
+            predictions, **mapped_kwargs
+        )
+        return prepare_classification_response(
+            post_processed_predictions,
+            image_sizes=returned_metadata,
+            class_names=self.class_names,
+            confidence_threshold=kwargs.get("confidence", 0.5),
+        )
 
     def clear_cache(self, delete_from_disk: bool = True) -> None:
         """Clears any cache if necessary. TODO: Implement this to delete the cache from the experimental model.
@@ -747,20 +751,27 @@ def prepare_multi_label_classification_response(
     post_processed_predictions: List[MultiLabelClassificationPrediction],
     image_sizes: List[Tuple[int, int]],
     class_names: List[str],
-    confidence_threshold: float,
 ) -> List[MultiLabelClassificationInferenceResponse]:
+    """Build the API response from a model's post-processed predictions.
+
+    `prediction.class_ids` is the authoritative list of "passed" classes —
+    the model's `post_process` already applied the
+    full priority chain (user → per-class → global → default), so the
+    response builder doesn't re-threshold here. The full per-class score
+    vector is still emitted in `image_predictions_dict` for UI display.
+    """
     results = []
     for prediction, image_size in zip(post_processed_predictions, image_sizes):
-        image_predictions_dict = dict()
-        predicted_classes = []
-        for class_id, confidence in enumerate(prediction.confidence.cpu().tolist()):
-            cls_name = class_names[class_id]
-            image_predictions_dict[cls_name] = {
+        image_predictions_dict = {
+            class_names[class_id]: {
                 "confidence": confidence,
                 "class_id": class_id,
             }
-            if confidence > confidence_threshold:
-                predicted_classes.append(cls_name)
+            for class_id, confidence in enumerate(prediction.confidence.cpu().tolist())
+        }
+        predicted_classes = [
+            class_names[class_id] for class_id in prediction.class_ids.tolist()
+        ]
         results.append(
             MultiLabelClassificationInferenceResponse(
                 predictions=image_predictions_dict,

inference/core/models/instance_segmentation_base.py

@@ -1,4 +1,4 @@
-from typing import Any, List, Tuple, Union
+from typing import Any, List, Optional, Tuple, Union
 
 import numpy as np
 
@@ -49,7 +49,7 @@ def infer(
         self,
         image: Any,
         class_agnostic_nms: bool = False,
-        confidence: float = DEFAULT_CONFIDENCE,
+        confidence: Optional[float] = DEFAULT_CONFIDENCE,
         disable_preproc_auto_orient: bool = False,
         disable_preproc_contrast: bool = False,
         disable_preproc_grayscale: bool = False,
@@ -94,6 +94,8 @@ def infer(
             - Applies non-maximum suppression to the predictions.
             - Decodes the masks according to the specified mode.
         """
+        if confidence is None:
+            confidence = DEFAULT_CONFIDENCE
         return super().infer(
             image,
             class_agnostic_nms=class_agnostic_nms,

inference/core/models/object_detection_base.py

@@ -43,7 +43,7 @@ def infer(
         self,
         image: Any,
         class_agnostic_nms: bool = DEFAULT_CLASS_AGNOSTIC_NMS,
-        confidence: float = DEFAULT_CONFIDENCE,
+        confidence: Optional[float] = DEFAULT_CONFIDENCE,
         disable_preproc_auto_orient: bool = False,
         disable_preproc_contrast: bool = False,
         disable_preproc_grayscale: bool = False,
@@ -81,6 +81,8 @@ def infer(
         Raises:
             ValueError: If batching is not enabled for the model and more than one image is passed for processing.
         """
+        if confidence is None:
+            confidence = DEFAULT_CONFIDENCE
         return super().infer(
             image,
             class_agnostic_nms=class_agnostic_nms,

inference/core/workflows/core_steps/models/roboflow/instance_segmentation/v2.py

@@ -85,11 +85,15 @@ class BlockManifest(WorkflowBlockManifest):
     images: Selector(kind=[IMAGE_KIND]) = ImageInputField
     model_id: Union[Selector(kind=[ROBOFLOW_MODEL_ID_KIND]), str] = RoboflowModelField
     confidence: Union[
-        FloatZeroToOne,
+        Optional[FloatZeroToOne],
         Selector(kind=[FLOAT_ZERO_TO_ONE_KIND]),
     ] = Field(
-        default=0.4,
-        description="Confidence threshold for predictions.",
+        default=None,
+        description=(
+            "Confidence threshold for predictions. If omitted, the inference "
+            "server uses the model's F1-optimal threshold from model evaluation "
+            "when available, otherwise falls back to 0.4."
+        ),
         examples=[0.3, "$inputs.confidence_threshold"],
     )
     class_filter: Union[Optional[List[str]], Selector(kind=[LIST_OF_VALUES_KIND])] = (

inference/core/workflows/core_steps/models/roboflow/keypoint_detection/v2.py

@@ -84,11 +84,15 @@ class BlockManifest(WorkflowBlockManifest):
     images: Selector(kind=[IMAGE_KIND]) = ImageInputField
     model_id: Union[Selector(kind=[ROBOFLOW_MODEL_ID_KIND]), str] = RoboflowModelField
     confidence: Union[
-        FloatZeroToOne,
+        Optional[FloatZeroToOne],
         Selector(kind=[FLOAT_ZERO_TO_ONE_KIND]),
     ] = Field(
-        default=0.4,
-        description="Confidence threshold for predictions.",
+        default=None,
+        description=(
+            "Confidence threshold for predictions. If omitted, the inference "
+            "server uses the model's F1-optimal threshold from model evaluation "
+            "when available, otherwise falls back to 0.4."
+        ),
         examples=[0.3, "$inputs.confidence_threshold"],
     )
     keypoint_confidence: Union[

inference/core/workflows/core_steps/models/roboflow/multi_class_classification/v2.py

@@ -78,11 +78,15 @@ class BlockManifest(WorkflowBlockManifest):
     images: Selector(kind=[IMAGE_KIND]) = ImageInputField
     model_id: Union[Selector(kind=[ROBOFLOW_MODEL_ID_KIND]), str] = RoboflowModelField
     confidence: Union[
-        FloatZeroToOne,
+        Optional[FloatZeroToOne],
         Selector(kind=[FLOAT_ZERO_TO_ONE_KIND]),
     ] = Field(
-        default=0.4,
-        description="Confidence threshold for predictions.",
+        default=None,
+        description=(
+            "Confidence threshold for predictions. If omitted, the inference "
+            "server uses the model's F1-optimal threshold from model evaluation "
+            "when available, otherwise falls back to 0.5."
+        ),
         examples=[0.3, "$inputs.confidence_threshold"],
     )
     disable_active_learning: Union[bool, Selector(kind=[BOOLEAN_KIND])] = Field(

inference/core/workflows/core_steps/models/roboflow/multi_label_classification/v2.py

@@ -77,11 +77,15 @@ class BlockManifest(WorkflowBlockManifest):
     images: Selector(kind=[IMAGE_KIND]) = ImageInputField
     model_id: Union[Selector(kind=[ROBOFLOW_MODEL_ID_KIND]), str] = RoboflowModelField
     confidence: Union[
-        FloatZeroToOne,
+        Optional[FloatZeroToOne],
         Selector(kind=[FLOAT_ZERO_TO_ONE_KIND]),
     ] = Field(
-        default=0.4,
-        description="Confidence threshold for predictions.",
+        default=None,
+        description=(
+            "Confidence threshold for predictions. If omitted, the inference "
+            "server uses the model's F1-optimal threshold from model evaluation "
+            "when available, otherwise falls back to 0.5."
+        ),
         examples=[0.3, "$inputs.confidence_threshold"],
     )
     disable_active_learning: Union[bool, Selector(kind=[BOOLEAN_KIND])] = Field(