Add 'rtol', 'atol' arguments to control tolerance in model validation

convert_to_mixed.py

@@ -7,8 +7,6 @@ from onnxconverter_common import auto_mixed_precision_model_path
 import argparse
 
 PROVIDERS=[('TensorrtExecutionProvider', {'trt_fp16_enable':True,}), 'CUDAExecutionProvider', 'CPUExecutionProvider']
-RTOL=0.1
-ATOL=0.1
 
 def detect_model_input_size(model_path):
     model = onnx.load(model_path)
@@ -46,7 +44,7 @@ def main(args):
                                     input_feed=input_feed, 
                                     target_model_path=args.target_model_path,
                                     customized_validate_func=None, 
-                                    rtol=RTOL, atol=ATOL,
+                                    rtol=args.rtol, atol=args.atol,
                                     provider=PROVIDERS, 
                                     keep_io_types=True,
                                     verbose=True)
@@ -54,7 +52,7 @@ def main(args):
     original_result = infer(args.source_model_path, input_feed)
     converted_result = infer(args.target_model_path, input_feed)
     
-    is_close = np.allclose(original_result[0], converted_result[0], rtol=RTOL, atol=ATOL)
+    is_close = np.allclose(original_result[0], converted_result[0], rtol=args.rtol, atol=args.atol)
     print(f"Validation result: {'Success' if is_close else 'Failure'}")
 
 if __name__ == "__main__":
@@ -62,6 +60,8 @@ if __name__ == "__main__":
     parser.add_argument("source_model_path", type=str, help="Path to the source ONNX model.")
     parser.add_argument("target_model_path", type=str, help="Path where the mixed precision model will be saved.")
     parser.add_argument("test_image_path", type=str, help="Path to a test image for validating the model conversion.")
+    parser.add_argument('--rtol', type=float, default=0.01, help=' the relative tolerance to do validation')
+    parser.add_argument('--atol', type=float, default=0.001, help=' the absolute tolerance to do validation')
 
     args = parser.parse_args()