saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-07 17:15:29 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Add option to allow quantize_input() use input_qtype for initializers. (#5721)

Browse source
This commit is contained in:
Zhang Lei 2020-11-06 09:33:24 -08:00 committed by GitHub
parent f666c3d7d7
commit 77b1eea9cf
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 8 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
onnxruntime/python/tools/quantization/onnx_quantizer.py
View file

@ -130,7 +130,7 @@ class ONNXQuantizer:
self.model.model.opset_import.remove(ai_onnx_domain[0])
self.model.model.opset_import.extend([onnx.helper.make_opsetid("", 11)])
opset_version = 11
self.fuse_dynamic_quant = True
return opset_version
@ -208,7 +208,7 @@ class ONNXQuantizer:
onnx.numpy_helper.to_array(initializer_scale)
]
#connect the previous and successive node input and output
# connect the previous and successive node input and output
for succ_node in succ_nodes:
succ_idx = get_elem_index(next_node.output[0], succ_node.input)
if succ_idx != -1:
@ -223,11 +223,11 @@ class ONNXQuantizer:
self.quantization_params = {}
self.quantization_params[param_name] = zp_and_scale
#remove fake-quantized nodes
# remove fake-quantized nodes
nodes_to_remove.extend([curr_node])
nodes_to_remove.extend([next_node])
#remove unused initializers in graph
# remove unused initializers in graph
initializers_to_remove.extend([initializer_scale])
initializers_to_remove.extend([initializer_zp])
@ -685,7 +685,7 @@ class ONNXQuantizer:
# Check if DequantizeLinear node needs to be added to graph.
if len(nodes_using_weight) != 0 and \
self.model.find_node_by_name(dequantize_linear_name,self.new_nodes,self.model.graph()) is None:
self.model.find_node_by_name(dequantize_linear_name, self.new_nodes, self.model.graph()) is None:
inputs = [weight.name + "_quantized", weight.name + "_scale", weight.name + "_zero_point"]
node = onnx.helper.make_node("DequantizeLinear", inputs, [output_name], dequantize_linear_name)
nodes_list.append(node)
@ -790,7 +790,7 @@ class ONNXQuantizer:
return quantized_bias_name
def quantize_inputs(self, node, indices):
def quantize_inputs(self, node, indices, initializer_use_weight_qType=True):
'''
Given a node, this function quantizes the inputs as follows:
- If input is an initializer, quantize the initializer data, replace old initializer
@ -899,4 +899,4 @@ class ONNXQuantizer:
for output in self.model.graph().output:
dequantize_node = self._dequantize_value(output.name)
if dequantize_node is not None:
self.new_nodes.append(dequantize_node)
self.new_nodes.append(dequantize_node)

2
onnxruntime/python/tools/quantization/operators/binary_op.py
View file

@ -17,7 +17,7 @@ class QLinearBinaryOp(QuantOperatorBase):
return super().quantize()
(quantized_input_names, zero_point_names, scale_names, nodes) = \
self.quantizer.quantize_inputs(node, [0, 1])
self.quantizer.quantize_inputs(node, [0, 1], initializer_use_weight_qType=False)
qlinear_binary_math_output = node.output[0] + "_quantized"
qlinear_binary_math_name = node.name + "_quant" if node.name != "" else ""