rangeNet.md

PyTorch RangeNet++

Environment Setup

Setup AI Model Efficiency Toolkit

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.24.0.

Add AIMET Model Zoo to the python path

export PYTHONPATH=$PYTHONPATH:<aimet_model_zoo_path>

Package Dependencies

Install required packages

pip install -r <path to aimet-model-zoo>/aimet_zoo_torch/rangenet/requirements.txt

Dataset

Semantic Kitti can be downloaded from here:

• http://www.semantic-kitti.org/dataset.html

The folder structure and format of Semantic kitti dataset is like below:

--dataset
	--sequences
		--00
			--velodyne
				--000000.bin
				--000001.bin
			--labels
				--000000.label
				--000001.label
			--poses.txt

---

Usage

python rangenet_quanteval.py \
		--dataset-path <The path to the dataset, default is '../models/train/tasks/semantic/dataset/'>
		--use-cuda <Use cuda or cpu, default is True> \

---

Model checkpoint and configuration

• The original prepared RangeNet++ checkpoint can be downloaded from here: darknet21.tar.gz or rangeNet_plus_FP32.tar.gz
• Optimized w8a8 checkpoint can be downloaded from here: rangeNet_plus_w8a8_checkpoint.pth
• Optimized w4a8 checkpoint can be downloaded from here: rangeNet_plus_w4a8_checkpoint.pth
• The Quantization Simulation (Quantsim) Configuration file can be downloaded from here: default_config_per_channel.json (Please see this page for more information on this file).

---

Quantization Configuration (W4A8/W8A8)

• Weight quantization: 4 bits for w4a8, 8 bits for w8a8, per channel symmetric quantization
• Bias parameters are not quantized
• Activation quantization: 8 bits, asymmetric quantization
• Model inputs are quantized
• Percentile was used as quantization scheme, and the value was set to 99.99
• Bn fold and Adaround have been applied on optimized checkpoint

---

Results

Below are the mIoU results of this RangeNet++ implementation on SemanticKitti:

Model Configuration	mIoU (%)
rangeNet_plus_FP32	47.2
rangeNet_plus_FP32 + simple PTQ(w8a8)	45.0
rangeNet_plus_W8A8_checkpoint	47.1
rangeNet_plus_W4A8_checkpoint	46.8