Skip to content

Latest commit

 

History

History
206 lines (190 loc) · 7.5 KB

SalsaNext.md

File metadata and controls

206 lines (190 loc) · 7.5 KB
Raw

Pytorch SalsaNext for LiDAR semantic segmentation

Environment Setup

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.25.0

Add AIMET Model Zoo to Pythonpath

  • If aimet-model-zoo source code is not ready, clone the aimet-model-zoo repo
  git clone https://github.com/quic/aimet-model-zoo.git
  export PYTHONPATH=$PYTHONPATH:<path/to/aimet-model-zoo>

Dataset

  • Semantic-kitti dataset can be downloaded from here:

  • Downloaded datasets should be arranged in one directory <dataset_path>

    • The <dataset_path> should be arranged in the following way
  <dataset_path_to_semnantic_kitti>/sequences/
  ├── 00
  │   ├── labels/
  │   ├── velodyne/
  │   ├── calib.txt
  │   ├── poses.txt
  │   ├── times.txt  
  ├── 01
  │   ├── labels/
  │   ├── velodyne/
  │   ├── calib.txt
  │   ├── poses.txt
  │   ├── times.txt

Usage

  • To run evaluation with QuantSim in AIMET, use the following
python salsanext_quanteval.py \
  --model-config <model configuration to use>
	--dataset-path <path to dataset folder> \
  • Example
python salsaNext_quanteval.py --model-config salsanext_w8a8 --dataset-path <dataset_path_to_semnantic_kitti>

Or

python salsaNext_quanteval.py --model-config salsanext_w4a8 --dataset-path <dataset_path_to_semnantic_kitti>

Available model configurations are:

  • salsanext_w8a8
  • salsanext_w4a8

Model checkpoints and configuration

  • Downloading checkpoints is handled through evaluation script. Configuration is set to default by evaluation script.
  • The model checkpoints can be downloaded from
    • FP32 checkpoint SalsaNext FP32.
    • Quantized files W8A8 SalsaNext_optimized_model.pth and SalsaNext_optimized_encoding.encodings W8A8.
    • Quantized files W4A8 SalsaNext_optimized_w4A8_model.pth and SalsaNext_optimized_w4A8_encoding.encodings W4A8.
  • The Quantization Simulation (Quantsim) Configuration file default_config.json can be downloaded from W4A8_JSON and W8A8_JSON. (Please see this page for more information on this file).

Quantization configuration for W8A8

  • Weight quantization: 8 bits
  • Activation quantization: 8 bits
  • PTQ techniques:
    • Firstly, apply batch_norm_fold API to make the folding, by batch_norm_fold.fold_all_batch_norms
    • Secondly, apply the Adaround API to optimize the weight, by AdaroundParameters(*) and Adaround.apply_adaround(*)
    • Finally, set the percentile (99.9%) as the quant scheme, by sim.set_percentile_value(99.9)
  • The checkpoint is with 3 activation output with 16 bitwidth with QAT.
    • sim.model.downCntx.conv1.input_quantizers[0].bitwidth = 16
    • sim.model.resBlock1.pool.output_quantizers[0].bitwidth = 16
    • sim.model.module_softmax.output_quantizers[0].bitwidth = 16

Quantization configuration for W4A8

  • Weight quantization: 4 bits
  • Activation quantization: 8 bits
  • PTQ techniques:
    • Set the QuantScheme with QuantScheme.training_range_learning_with_tf_init
    • Apply batch_norm_fold API to make the folding, by batch_norm_fold.fold_all_batch_norms
  • The checkpoint is with 9 activation output with 16 bitwidth with QAT.
    • sim.model.downCntx.conv3.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx.conv2.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx.conv1.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx.conv1.input_quantizers[0].bitwidth = 16
    • sim.model.downCntx.act2.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx2.act2.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx3.act2.output_quantizers[0].bitwidth = 16
    • sim.model.downCntx.act1.output_quantizers[0].bitwidth = 16
    • sim.model.module_softmax.output_quantizers[0].bitwidth = 16
  • The checkpoint is with 6 weight with 8 bitwidth with QAT.
    • sim.model.downCntx.conv1.param_quantizers['weight'].bitwidth = 8
    • sim.model.downCntx.conv3.param_quantizers['weight'].bitwidth = 8
    • sim.model.downCntx2.conv3.param_quantizers['weight'].bitwidth = 8
    • sim.model.downCntx3.conv3.param_quantizers['weight'].bitwidth = 8
    • sim.model.downCntx.conv2.param_quantizers['weight'].bitwidth = 8
    • sim.model.downCntx.bn2.param_quantizers['weight'].bitwidth = 8

Results

Model IoU avg Acc avg
FP32 W8A8 W4A8 FP32 W8A8 W4A8
SalsaNext 0.558 0.549 0.551 0.879 0.874 0.876

FP32 results

Acc avg 0.879
IoU avg 0.558
IoU class 1 [car] = 0.862
IoU class 2 [bicycle] = 0.394
IoU class 3 [motorcycle] = 0.420
IoU class 4 [truck] = 0.777
IoU class 5 [other-vehicle] = 0.420
IoU class 6 [person] = 0.621
IoU class 7 [bicyclist] = 0.683
IoU class 8 [motorcyclist] = 0.000
IoU class 9 [road] = 0.943
IoU class 10 [parking] = 0.422
IoU class 11 [sidewalk] = 0.800
IoU class 12 [other-ground] = 0.041
IoU class 13 [building] = 0.800
IoU class 14 [fence] = 0.484
IoU class 15 [vegetation] = 0.803
IoU class 16 [trunk] = 0.579
IoU class 17 [terrain] = 0.642
IoU class 18 [pole] = 0.466
IoU class 19 [traffic-sign] = 0.445

W8A8 results

Acc avg 0.874
IoU avg 0.549
IoU class 1 [car] = 0.863
IoU class 2 [bicycle] = 0.372
IoU class 3 [motorcycle] = 0.426
IoU class 4 [truck] = 0.780
IoU class 5 [other-vehicle] = 0.447
IoU class 6 [person] = 0.600
IoU class 7 [bicyclist] = 0.663
IoU class 8 [motorcyclist] = 0.000
IoU class 9 [road] = 0.936
IoU class 10 [parking] = 0.395
IoU class 11 [sidewalk] = 0.789
IoU class 12 [other-ground] = 0.035
IoU class 13 [building] = 0.791
IoU class 14 [fence] = 0.444
IoU class 15 [vegetation] = 0.796
IoU class 16 [trunk] = 0.562
IoU class 17 [terrain] = 0.647
IoU class 18 [pole] = 0.457
IoU class 19 [traffic-sign] = 0.420

W4A8 results

Acc avg 0.876
IoU avg 0.551
IoU class 1 [car] = 0.863
IoU class 2 [bicycle] = 0.381
IoU class 3 [motorcycle] = 0.437
IoU class 4 [truck] = 0.770
IoU class 5 [other-vehicle] = 0.428
IoU class 6 [person] = 0.606
IoU class 7 [bicyclist] = 0.670
IoU class 8 [motorcyclist] = 0.000
IoU class 9 [road] = 0.937
IoU class 10 [parking] = 0.378
IoU class 11 [sidewalk] = 0.788
IoU class 12 [other-ground] = 0.053
IoU class 13 [building] = 0.798
IoU class 14 [fence] = 0.467
IoU class 15 [vegetation] = 0.800
IoU class 16 [trunk] = 0.567
IoU class 17 [terrain] = 0.646
IoU class 18 [pole] = 0.456
IoU class 19 [traffic-sign] = 0.428