fix: don't keep numpy ndarray in debug.inputs dict

robotoff/prediction/category/neural/keras_category_classifier_3_0/__init__.py

@@ -60,7 +60,7 @@ def generate_image_embeddings(product: JSONType, stub) -> Optional[np.ndarray]:
         return np.frombuffer(
             response.raw_output_contents[0],
             dtype=np.float32,
-        ).reshape((len(images), -1))
+        ).reshape((len(non_null_images), -1))
     return None
 
 
@@ -103,11 +103,7 @@ def predict(
         threshold = 0.5
 
     inputs = generate_inputs_dict(product, ocr_texts, image_embeddings)
-    debug: JSONType = {
-        "model_name": model_name.value,
-        "threshold": threshold,
-        "inputs": inputs,
-    }
+    debug = generate_debug_dict(model_name, threshold, inputs)
     scores, labels = _predict(inputs, model_name)
     indices = np.argsort(-scores)
 
@@ -125,6 +121,39 @@ def predict(
     return category_predictions, debug
 
 
+def generate_debug_dict(
+    model_name: NeuralCategoryClassifierModel, threshold: float, inputs: JSONType
+) -> JSONType:
+    """Generate dict containing debug information.
+
+    :param model_name: name of the model used during prediction
+    :param threshold: detection threshold used
+    :param inputs: inputs dict used for inference
+    :return: the debug dict
+    """
+    debug = {
+        "model_name": model_name.value,
+        "threshold": threshold,
+        "inputs": {
+            k: v
+            for k, v in inputs.items()
+            # Don't keep numpy ndarray in debug.inputs dict
+            if k not in ("image_embeddings", "image_embeddings_mask")
+        },
+    }
+
+    if inputs["image_embeddings_mask"].sum() == 1:
+        # `image_embeddings_mask` always has at least one non-zero element,
+        # check whether there is an image or not by checking if
+        # `image_embeddings` is zero-filled
+        num_images = 0 if np.all(inputs["image_embeddings"][0] == 0) else 1
+    else:
+        num_images = int(inputs["image_embeddings_mask"].sum())
+
+    debug["inputs"]["num_images"] = num_images  # type: ignore
+    return debug
+
+
 # Parameters on how to prepare data for each model type, see `build_triton_request`
 model_input_flags: dict[NeuralCategoryClassifierModel, dict] = {
     NeuralCategoryClassifierModel.keras_image_embeddings_3_0: {},