Convert input for quantized YOLOv8 (#1521) (#1628)

* Added call to function that skips the quantization of the input if the model is quantized

* Removed bare exception

* Move input data type conversion upstream so it is properly saved and used as input to ort

* Style and quality fixes

---------

Co-authored-by: Alexandre Marques <[email protected]>
Co-authored-by: Michael Goin <[email protected]>

Author: 3 people

src/sparseml/yolov8/trainers.py

@@ -29,6 +29,7 @@
 
 from sparseml.optim.helpers import load_recipe_yaml_str
 from sparseml.pytorch.optim.manager import ScheduledModifierManager
+from sparseml.pytorch.sparsification.quantization import skip_onnx_input_quantize
 from sparseml.pytorch.utils import ModuleExporter
 from sparseml.pytorch.utils.helpers import download_framework_model_by_recipe_type
 from sparseml.pytorch.utils.logger import LoggerManager, PythonLogger, WANDBLogger
@@ -729,7 +730,13 @@ def export(self, **kwargs):
             else ["output0"],
         )
 
-        onnx.checker.check_model(os.path.join(save_dir, name))
+        complete_path = os.path.join(save_dir, name)
+        try:
+            skip_onnx_input_quantize(complete_path, complete_path)
+        except Exception:
+            pass
+
+        onnx.checker.check_model(complete_path)
         deployment_folder = exporter.create_deployment_folder(onnx_model_name=name)
         if args["export_samples"]:
             trainer_config = get_cfg(cfg=DEFAULT_SPARSEML_CONFIG_PATH)

src/sparseml/yolov8/utils/export_samples.py

@@ -106,10 +106,15 @@ def export_sample_inputs_outputs(
         preprocessed_batch = preprocess(batch=batch, device=device)
         image = preprocessed_batch["img"]
 
-        # Save inputs as numpy array
-        _export_inputs(image, sample_in_dir, file_idx, save_inputs_as_uint8)
         # Save torch outputs as numpy array
         _export_torch_outputs(image, model, sample_out_dir_torch, file_idx)
+
+        # Convert input data type if needed
+        if save_inputs_as_uint8:
+            image = (255 * image).to(dtype=torch.uint8)
+
+        # Save inputs as numpy array
+        _export_inputs(image, sample_in_dir, file_idx)
         # Save onnxruntime outputs as numpy array
         _export_ort_outputs(
             image.cpu().numpy(), ort_session, sample_out_dir_ort, file_idx
@@ -166,13 +171,9 @@ def _export_ort_outputs(
     numpy.savez(sample_output_filename, preds)
 
 
-def _export_inputs(
-    image: torch.Tensor, sample_in_dir: str, file_idx: str, save_inputs_as_uint8: bool
-):
+def _export_inputs(image: torch.Tensor, sample_in_dir: str, file_idx: str):
 
     sample_in = image.detach().to("cpu")
-    if save_inputs_as_uint8:
-        sample_in = (255 * sample_in).to(dtype=torch.uint8)
 
     sample_input_filename = os.path.join(sample_in_dir, f"inp-{file_idx}.npz")
     numpy.savez(sample_input_filename, sample_in)