Fake quantization when the padding of convollution layer is asymmetric or pad_w != pad_h #11525
Description
Describe the bug
Consider a very simple model which has only one convolution layer with one input and one output. If the convolution layer has asymmetric padding or padding of pad_w != pad_h, using onnxruntime.quantization.quantiza_static generates fake-quantized model like the following:
Urgency
Probably none.
System information
- OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 20.04.4 LTS
- ONNX Runtime installed from (source or binary): binary
- ONNX Runtime version: 1.11.1
- Python version: 3.9.7
- Visual Studio version (if applicable): None
- GCC/Compiler version (if compiling from source): None
- CUDA/cuDNN version: None
- GPU model and memory: None
To Reproduce
# conv with padding of pad_w != pad_h
from torch.autograd import Variable
import torch
import torch.nn as nn
import numpy as np
import onnx # version 1.11.1
import onnxruntime as rt
from onnxruntime.quantization import quantize_static, CalibrationDataReader, QuantType
input = Variable(torch.randn(1, 3, 10, 10))
conv = nn.Conv2d(3, 5, kernel_size=3, stride=2, padding=(2, 1))
torch.onnx.export(conv, input, 'conv.onnx', export_params=True, opset_version=12)
class DataReader(CalibrationDataReader):
def __init__(self, model_path, batchsize=5):
sess = rt.InferenceSession(model_path, None)
input_name = sess.get_inputs()[0].name
input_shape = sess.get_inputs()[0].shape
calib_data = np.random.uniform(-1, 1, size=[batchsize] + input_shape[1:]).astype("float32")
self.enum_data_dicts = iter([{input_name: np.expand_dims(x, axis=0)} for x in calib_data])
def get_next(self):
return next(self.enum_data_dicts, None)
dr = DataReader('conv.onnx')
quantize_static('conv.onnx', 'conv-quant.onnx', dr, per_channel=False,
activation_type=QuantType.QInt8, weight_type=QuantType.QInt8)Or you can replace 'conv.onnx' with the one of asymmetric padding to try quantizing asymmetric-padded conv layer.
Expected behavior
Quantized model (generated by neural compressor) of the structure like the following:
