[6090146] Added support inp->DQ graphs

[gcunhase]

Overview

Type: Feature enhancement

Changes:

• Added support for translating models with a standalone DequantizeLinear connected to an int8/uint8 graph input (inp->DQ).
• Skip the onnxoptimizer fuse_bn_into_conv pass when any BatchNormalization parameter is fp16, to avoid graph decomposition (see https://github.com/onnx/optimizer/blob/master/onnxoptimizer/passes/fuse_bn_into_conv.h#L40).

Original	QDQ-Translated (Before)	QDQ-Translated (After)

This also

Steps to repro

1. Install ModelOpt:

$ pip install nvidia-modelopt[onnx]

2. Download example model: resnet50-v1-12.onnx

3. Obtained explicitly-quantized (EQ) model:

$ python -m modelopt.onnx --onnx_path=resnet50-v1-12.onnx --direct_io_types

4. Obtain DLA-compatible model and cache / args files with QDQ-Translator:

$ python tools/qdq-translator/qdq_translator.py -i resnet50-v1-12.quant.onnx -o translated