add onnxslim intergration (#811)

* add onnxslim intergration

* fix onnxslim version

scripts/convert.py

@@ -14,6 +14,7 @@
 )
 
 import onnx
+import onnxslim
 from optimum.exporters.onnx import main_export, export_models
 from optimum.onnx.graph_transformations import check_and_save_model
 from optimum.exporters.tasks import TasksManager
@@ -224,6 +225,12 @@ class ConversionArguments:
             "that desire a finer-grained control on the export."
         }
     )
+    skip_onnxslim: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether or not to skip onnxslim."
+        }
+    )
 
 
 def get_operators(model: onnx.ModelProto) -> Set[str]:
@@ -623,6 +630,13 @@ def main():
 
     os.makedirs(os.path.join(output_model_folder, 'onnx'), exist_ok=True)
 
+    if not conv_args.skip_onnxslim:
+        onnx_models = [os.path.join(output_model_folder, x)
+                    for x in os.listdir(output_model_folder) if x.endswith('.onnx')]
+
+        for model in onnx_models:
+            onnxslim.slim(model, model)
+
     # Step 2. (optional, recommended) quantize the converted model for fast inference and to reduce model size.
     if conv_args.quantize:
         # Update quantize config with model specific defaults

scripts/requirements.txt

@@ -4,3 +4,4 @@ optimum==1.17.1
 onnx==1.15.0
 onnxconverter-common==1.14.0
 tqdm
+onnxslim==0.1.29.1