Failed to quantize ConvTranspose with per_channel=True

[mo-ja]

Describe the issue

ConvTranspose layer cannot be quantized with per_channel=True.

In the case of ConvTranspose, the axis of weight, axis=1, corresponds to the number of channels in the output, whereas QDQConv's per_channel quantization always quantizes on the axis of axis=0.
This seems to cause an error during the bias scale calculation in ConvTranspose because the shape of bias_scale does not match the shape of bias.
I have solved this issue by using different axis for Conv and ConvTranspose like here:

• mo-ja@9f42bf8

To reproduce

Here is the onnx file I used: onnx.zip

model generation with torch

import torch
import torch.nn as nn


class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.conv_transpose = nn.ConvTranspose2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1)

    def forward(self, x):
        x = self.conv_transpose(x)
        return x

model = MyModel()
input_data = torch.randn(1, 3, 32, 32)

torch.onnx.export(model, input_data, "model.onnx",opset_version=13)

quantization code with ONNX Runtime

import onnxruntime.quantization as quantization
import numpy as np
class SampleReader(quantization.CalibrationDataReader):
    def __init__(self):
       self.flg = False

    def get_next(self):
      if self.flg:
         return None
      self.flg = True
      return {"onnx::ConvTranspose_0":np.random.uniform(0, 1, [1, 3, 32, 32]).astype(np.float32)}
    
cdr = SampleReader()
quantization.quantize_static("./model.onnx", "quant.onnx", cdr, per_channel=True)

Raised Error

ValueError                                Traceback (most recent call last)
[/tmp/ipykernel_833326/1616630263.py](https://file+.vscode-resource.vscode-cdn.net/tmp/ipykernel_833326/1616630263.py) in 
     12 
     13 cdr = SampleReader()
---> 14 quantization.quantize_static("./model.onnx", "quant.onnx", cdr, per_channel=True)

[~/.local/lib/python3.8/site-packages/onnxruntime/quantization/quantize.py](https://file+.vscode-resource.vscode-cdn.net/home/<user name>/work/ort_report/001_ConvTranspose/torch/~/.local/lib/python3.8/site-packages/onnxruntime/quantization/quantize.py) in quantize_static(model_input, model_output, calibration_data_reader, quant_format, op_types_to_quantize, per_channel, reduce_range, activation_type, weight_type, nodes_to_quantize, nodes_to_exclude, use_external_data_format, calibrate_method, extra_options)
    535         )
    536 
--> 537     quantizer.quantize_model()
    538     quantizer.model.save_model_to_file(model_output, use_external_data_format)
    539     if not pre_processed:

[~/.local/lib/python3.8/site-packages/onnxruntime/quantization/qdq_quantizer.py](https://file+.vscode-resource.vscode-cdn.net/home/<user name>/work/ort_report/001_ConvTranspose/torch/~/.local/lib/python3.8/site-packages/onnxruntime/quantization/qdq_quantizer.py) in quantize_model(self)
    264         self._quantize_sharing_param_tensors()
    265         if self.quantize_bias:
--> 266             self._quantize_bias_tensors()
    267         self.remove_nodes()
    268         if not self.add_qdq_pair_to_weight:

[~/.local/lib/python3.8/site-packages/onnxruntime/quantization/qdq_quantizer.py](https://file+.vscode-resource.vscode-cdn.net/home/<user name>/work/ort_report/001_ConvTranspose/torch/~/.local/lib/python3.8/site-packages/onnxruntime/quantization/qdq_quantizer.py) in _quantize_bias_tensors(self)
    480                 continue
    481             # Quantize the input
--> 482             self.quantize_bias_static(bias_name, input_name, weight_name, beta)
    483             init = find_by_name(bias_name, self.model.initializer())
    484             self.model.remove_initializer(init)

[~/.local/lib/python3.8/site-packages/onnxruntime/quantization/onnx_quantizer.py](https://file+.vscode-resource.vscode-cdn.net/home/<user name>/work/ort_report/001_ConvTranspose/torch/~/.local/lib/python3.8/site-packages/onnxruntime/quantization/onnx_quantizer.py) in quantize_bias_static(self, bias_name, input_name, weight_name, beta)
    917             bias_scale = input_scale * weight_scale * beta
    918 
--> 919             quantized_data = (np.asarray(bias_data) [/](https://file+.vscode-resource.vscode-cdn.net/) bias_scale).round().astype(np.int32)
    920 
    921             # update bias initializer

ValueError: operands could not be broadcast together with shapes (64,) (3,)

Urgency

No response

Platform

Linux

OS Version

Ubuntu 20.04.5 LTS

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.17.1

ONNX Runtime API

Python

Architecture

X86

Execution Provider

Default CPU

Execution Provider Library Version

No response

Tasks

Preview Give feedback

No tasks being tracked yet.