Revert "Make keepdims to its default value when adding ReduceMin/ReduceMax for quantization calibration (microsoft#6788)" (microsoft#6825)

This reverts commit 9b3171e.

Author: oliviajain

onnxruntime/python/tools/quantization/calibrate.py

@@ -10,7 +10,7 @@
 import numpy as np
 import onnx
 import onnxruntime
-from onnx import helper, TensorProto, ModelProto, shape_inference
+from onnx import helper, TensorProto, ModelProto
 from onnx import onnx_pb as onnx_proto
 from six import string_types
 from enum import Enum
@@ -52,9 +52,6 @@ def __init__(self, model, op_types_to_calibrate=[], augmented_model_path='augmen
         else:
             raise ValueError('model should be either model path or onnx.ModelProto.')
 
-        # Apply shape inference on the model
-        self.model = onnx.shape_inference.infer_shapes(self.model)
-
         self.op_types_to_calibrate = op_types_to_calibrate
         self.augmented_model_path = augmented_model_path
 
@@ -163,27 +160,26 @@ def augment_graph(self):
 
         added_nodes = []
         added_outputs = []
-        tensors, value_infos = self.select_tensors_to_calibrate(model) 
+        tensors, _ = self.select_tensors_to_calibrate(model) 
 
         for tensor in tensors:
-
-            # Get tensor's shape 
-            dim = len(value_infos[tensor].type.tensor_type.shape.dim)
-            shape = (1,) if dim == 1 else list(1 for i in range(dim))
-
             # Adding ReduceMin nodes
             reduce_min_name = tensor + '_ReduceMin'
-            reduce_min_node = onnx.helper.make_node('ReduceMin', [tensor], [tensor + '_ReduceMin'], reduce_min_name)
+            reduce_min_node = onnx.helper.make_node('ReduceMin', [tensor], [tensor + '_ReduceMin'],
+                                                    reduce_min_name,
+                                                    keepdims=0)
 
             added_nodes.append(reduce_min_node)
-            added_outputs.append(helper.make_tensor_value_info(reduce_min_node.output[0], TensorProto.FLOAT, shape))
+            added_outputs.append(helper.make_tensor_value_info(reduce_min_node.output[0], TensorProto.FLOAT, ()))
 
             # Adding ReduceMax nodes
             reduce_max_name = tensor + '_ReduceMax'
-            reduce_max_node = onnx.helper.make_node('ReduceMax', [tensor], [tensor + '_ReduceMax'], reduce_max_name)
+            reduce_max_node = onnx.helper.make_node('ReduceMax', [tensor], [tensor + '_ReduceMax'],
+                                                    reduce_max_name,
+                                                    keepdims=0)
 
             added_nodes.append(reduce_max_node)
-            added_outputs.append(helper.make_tensor_value_info(reduce_max_node.output[0], TensorProto.FLOAT, shape))
+            added_outputs.append(helper.make_tensor_value_info(reduce_max_node.output[0], TensorProto.FLOAT, ()))
 
         model.graph.node.extend(added_nodes)
         model.graph.output.extend(added_outputs)