Fix yolact

src/deepsparse/engine.py

@@ -29,10 +29,10 @@
 from deepsparse.utils import (
     generate_random_inputs,
     get_output_names,
+    join_engine_outputs,
     model_to_path,
     override_onnx_input_shapes,
     split_engine_inputs,
-    join_engine_outputs,
 )
 
 

src/deepsparse/utils/data.py

@@ -230,8 +230,6 @@ def split_engine_inputs(
                 [arr[section_idx : section_idx + batch_size] for arr in items]
             )
 
-    print(f"sliced up {total_batch_size} into {len(batches)} pieces")
-
     return batches, total_batch_size
 
 

src/deepsparse/yolact/pipelines.py

@@ -166,24 +166,6 @@ def process_inputs(
         )
         return [image_batch], postprocessing_kwargs
 
-    @staticmethod
-    def join_engine_outputs(
-        batch_outputs: List[List[numpy.ndarray]],
-    ) -> List[numpy.ndarray]:
-        boxes, confidence, masks, priors, protos = Pipeline.join_engine_outputs(
-            batch_outputs
-        )
-
-        # priors never has a batch dimension
-        # so the above step doesn't concat along a batch dimension
-        # reshape into a batch dimension
-        num_priors = boxes.shape[1]
-        batch_priors = numpy.reshape(priors, (-1, num_priors, 4))
-
-        # all the priors should be equal, so only use the first one
-        assert (batch_priors == batch_priors[0]).all()
-        return [boxes, confidence, masks, batch_priors[0], protos]
-
     def _preprocess_image(self, image) -> numpy.ndarray:
         if isinstance(image, str):
             image = cv2.imread(image)
@@ -209,6 +191,15 @@ def process_engine_outputs(
         priors = torch.from_numpy(priors).cpu()
         protos = torch.from_numpy(protos).cpu()
 
+        # priors never has a batch dimension
+        # so the above step doesn't concat along a batch dimension
+        # reshape into a batch dimension
+        # all the priors should be equal, so only use the first one
+        num_priors = boxes.shape[1]
+        batch_priors = numpy.reshape(priors, (-1, num_priors, 4))
+        assert (batch_priors == batch_priors[0]).all()
+        priors = batch_priors[0]
+
         batch_size, num_priors, _ = boxes.size()
 
         # Preprocess every image in the batch individually