saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-06-01 23:30:35 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add symmetric quant in softmax (#14640)

Browse source 
### Description

https://github.com/microsoft/onnxruntime/issues/14626


### Motivation and Context

https://github.com/microsoft/onnxruntime/issues/14626
This commit is contained in:
Chen Fu 2023-02-10 08:36:04 -08:00 committed by GitHub
parent 9bd022b8be
commit 0de4bc7050
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 38 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
onnxruntime/python/tools/quantization/operators/softmax.py
View file

@ -80,6 +80,11 @@ class QDQSoftmax(QDQOperatorBase):
if self.quantizer.activation_qType == onnx.onnx_pb.TensorProto.UINT8:
out_scale = 1 / 256.0
out_zero_point = 0
elif self.quantizer.is_activation_symmetric:
# results are all greater or equal to 0, so we can only use
# half of the range
out_scale = 1 / 127.0
out_zero_point = 0
else:
out_scale = 1 / 256.0
out_zero_point = -128

41
onnxruntime/test/python/quantization/test_op_softmax.py
View file

@ -10,10 +10,11 @@ Softmax quantization test case
# --------------------------------------------------------------------------
import unittest
from pathlib import Path
import numpy as np
import onnx
from onnx import TensorProto, helper
from onnx import TensorProto, helper, numpy_helper
from op_test_utils import TestDataFeeds, check_model_correctness, check_op_type_count, check_qtype_by_node_type
from onnxruntime.quantization import QuantFormat, QuantType, quantize_static
@ -148,13 +149,33 @@ class TestOpSoftmax(unittest.TestCase):
weight_type=weight_type,
extra_options=extra_options,
)
qdqnode_counts = {
"Conv": 1,
"QuantizeLinear": 3,
"DequantizeLinear": 4,
"Softmax": 1,
}
check_op_type_count(self, model_q8_qdq_path, **qdqnode_counts)
result_model = onnx.load(Path(model_q8_qdq_path))
qnode_cnt = 0
dqnode_cnt = 0
softmax_cnt = 0
qnode_zeropoints = []
for node in result_model.graph.node:
if node.op_type == "QuantizeLinear":
qnode_cnt += 1
qnode_zeropoints.append(node.input[2])
elif node.op_type == "DequantizeLinear":
dqnode_cnt += 1
elif node.op_type == "Softmax":
softmax_cnt += 1
self.assertEqual(3, qnode_cnt, "Expected 3 QuantizeLinear nodes, found {}".format(qnode_cnt))
self.assertEqual(4, dqnode_cnt, "Expected 4 DequantizeLinear nodes, found {}".format(dqnode_cnt))
self.assertEqual(1, softmax_cnt, "Expected 1 Softmax node, found {}".format(softmax_cnt))
if extra_options.get("ActivationSymmetric", False):
for tensor in result_model.graph.initializer:
if tensor.name in qnode_zeropoints:
np_value = numpy_helper.to_array(tensor)
self.assertEqual(
0,
np_value,
"QuantizeLinear node zero point value must be 0, found {} instead!".format(np_value),
)
qnode_io_qtypes = {
"QuantizeLinear": [
["i", 2, activation_proto_qtype],
@ -169,6 +190,10 @@ class TestOpSoftmax(unittest.TestCase):
self.quantize_softmax_test(QuantType.QUInt8, QuantType.QUInt8)
def test_quantize_softmax_s8s8(self):
self.quantize_softmax_test(
QuantType.QInt8,
QuantType.QInt8,
)
self.quantize_softmax_test(
QuantType.QInt8,
QuantType.QInt8,