rangeNet.md

PyTorch RangeNet++

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.

Model modifications & Experiment Setup

1. Clone the RangeNet++ repo

git clone https://github.com/PRBonn/lidar-bonnetal.git

2. Apply patches to darknet.py in the above repo using the command below:

patch /path/to/lidar-bonnetal/train/backbones/darknet.py /path/to/aimet-model-zoo/aimet_zoo_torch/rangenet/train/models/backbones/darknet.patch

path /path/to/lidar-bonnetal/train/tasks/semantic/decoders/darknet.py /path/to/aimet-model-zoo/aimet_zoo_torch/rangenet/train/tasks/semantic/decoders/darknet.patch

These changes are needed in order to meet prepare_model's requirements

2. Create a new folder to put your downloaded dataset

3. Create a new folder to put your downloaded original/optimized model

4. Add the "models/train/tasks/semantic/evaluate.py" file to your "models/train/tasks/semantic" path

5. Add AIMET Model Zoo to the python path

export PYTHONPATH=$PYTHONPATH:<aimet_model_zoo_path>

Dataset

The Semantic kitti Dataset 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

Model checkpoint and configuration

• The original prepared RangeNet++ checkpoint can be downloaded from here:
  • http://www.ipb.uni-bonn.de/html/projects/bonnetal/lidar/semantic/models/darknet21.tar.gz or
  • https://github.com/quic/aimet-model-zoo/releases/download/torch_rangenet_plus_w8a8/rangeNet_plus_FP32.tar.gz
• Optimized checkpoint can be downloaded from the:
  • https://github.com/quic/aimet-model-zoo/releases/download/torch_rangenet_plus_w8a8/rangeNet_plus_INT8_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).

Usage

To run evaluation with QuantSim in AIMET, use the following

python rangenet++_quanteval.py \
		--dataset-path <The path to the dataset, default is '../models/train/tasks/semantic/dataset/'>
		--model-orig-path <The path to the model_orig, default is '../models/train/tasks/semantic/pre_trained_model'>
		--model-optim-path <The path to the model_optim, default is '../models/train/tasks/semantic/quantized_model'>
		--use-cuda <Use cuda or cpu, default is True> \
		--batch-size <Number of images per batch, default is 1>

Quantization Configuration (W8A8)

• Weight quantization: 8 bits, 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 the PyTorch rangeNet++ model for the semantic kitti dataset:

Model Configuration	FP32 (%)	W8A8 (%)
rangeNet_plus_FP32	47.2	46.8
rangeNet_plus_W8A8_checkpoint	-	47.0