MobileNetV2-SSD-lite.md

PyTorch MobileNetV2-SSD-lite

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 1.22.2.

Model modifications

1. Clone the original repository

git clone https://github.com/qfgaohao/pytorch-ssd.git
cd pytorch-ssd
git checkout f61ab424d09bf3d4bb3925693579ac0a92541b0d
git apply ../aimet-model-zoo/zoo_torch/examples/ssd_mobilenetv2/patch_ssd.patch

2. Add AIMET model zoo and Pytorch-SSD to the python path

export PYTHONPATH=$PYTHONPATH:<path to parent>/pytorch-ssd
export PYTHONPATH=$PYTHONPATH:<path to parent>/aimet-model-zoo

Obtaining model checkpoint and dataset

• The original MobileNetV2-SSD-lite checkpoint can be downloaded from here:
  • https://storage.googleapis.com/models-hao/mb2-ssd-lite-mp-0_686.pth
• Optimized checkpoint can be downloaded from the Releases.
• Pascal VOC2007 dataset can be downloaded from here:
  • http://host.robots.ox.ac.uk/pascal/VOC/voc2007/index.html

Usage

• To run evaluation with QuantSim in AIMET, use the following

python ssd_mobilenetv2_quanteval.py --dataset-path <The root directory of dataset, e.g., my_path/VOCdevkit/VOC2007/>

Quantization Configuration

• Weight quantization: 8 bits, per tensor asymmetric quantization
• Bias parameters are not quantized
• Activation quantization: 8 bits, asymmetric quantization
• Model inputs are quantized
• TF_enhanced was used as quantization scheme
• Cross-layer-Equalization and Adaround have been applied on optimized checkpoint