QDQ removal around Resize (mode=linear) causes wrong numeric values

[mgehre-amd]

Describe the issue

The optimization in

onnxruntime/onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc

Line 58 in 88336ff

{"Resize", {}}},

removes Q-DQ nodes around Resize and thus makes the Resize compute in int8.
This might be fine for nearest interpolation mode, but in linear interpolation mode it changes the numeric output values.

The mismatch does not happen when using sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL.

To reproduce

#!/usr/bin/env python3
import onnx
import onnx.reference
from numpy import array, float32
import numpy

model = """\
<
  ir_version: 8,
  opset_import: ["" : 19]
>
main_graph (float[4] Concat_output_0, float[1,1,1,2] Abs_output_0) 
=> (float[unk__0,unk__1,unk__2,unk__3] DequantizeLinear_2_output_0) 
{
  Cast_2_output_0 = Constant <value: tensor = int8 {0}> ()
  Cast_3_output_0 = Constant <value: tensor = float {1.0}> ()
  QuantizeLinear_1_output_0 = QuantizeLinear (Abs_output_0, Cast_3_output_0, Cast_2_output_0)
  DequantizeLinear_1_output_0 = DequantizeLinear (QuantizeLinear_1_output_0, Cast_3_output_0, Cast_2_output_0)
  Resize_output_0 = Resize <mode: string = "linear"> (DequantizeLinear_1_output_0, , Concat_output_0)
  Cast_4_output_0 = Constant <value: tensor = int8 {0}> ()
  Cast_5_output_0 = Constant <value: tensor = float {1.0}> ()
  QuantizeLinear_2_output_0 = QuantizeLinear (Resize_output_0, Cast_5_output_0, Cast_4_output_0)
  DequantizeLinear_2_output_0 = DequantizeLinear (QuantizeLinear_2_output_0, Cast_5_output_0, Cast_4_output_0)
}

"""

m = onnx.parser.parse_model(model)

inputs = {'Concat_output_0': array([1., 1., 1., 2.], dtype=float32), 
         'Abs_output_0': array([[[[0.0 , 1.0]]]],dtype=float32)}

out = onnx.reference.ReferenceEvaluator(m).run(None, inputs)[0]
print("onnx.reference", out)

import onnxruntime
sess_options = onnxruntime.SessionOptions()
#sess_options.log_severity_level = 0
# Uncommenting the next line fixes the numeric mismatches
# sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL
sess_options.optimized_model_filepath = "opt.onnx"
onnxref = onnxruntime.InferenceSession(m.SerializeToString(), sess_options).run(None, inputs)[0]

print("onnxruntime", onnxref)

numpy.testing.assert_allclose(out, onnxref, rtol=1e-3, atol=1e-3)

shows a numeric mismatch:

onnx.reference [[[[0. 0. 1. 1.]]]]
onnxruntime [[[[0. 0. 0. 1.]]]]
AssertionError: 
Not equal to tolerance rtol=0.001, atol=0.001

If you look at opt.onnx, you see that the model has been transformed into

which is invalid because the linear interpolation cannot be done correctly in int8.

Urgency

No response

Platform

Linux

OS Version

Ubuntu 22.04

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.1

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response