Models-pruning

1. Prune MobileNetv1,2,3, VGG16, and EfficientNet-b0 in various way using Torch-Pruning library (https://github.com/VainF/Torch-Pruning)
2. Convert .pt to .ONNX
3. Quantize ONNX models
4. Deploy ONNX models on Rasberry Pi 3b+

• /checkpoints: .pt model files

• /deploy: Deployment code

  • deploy_onnx.py: Deploy tflite file on RPi

  python deploy_onnx_measurement.py --model [model name] --file [output_file_suffix ]

  • measure.py: Measure temperature and energy usage of RPi
  • metric_reader.py: Organize results of deployment

• /(model): Pruned .pt model files

• /models: MobileNetv1, MobileNetv2, and MobileNetv3 codes

  • (model).py: Layers are separated
  • (model)_default.py: Default model codes

• /onnx: Converted ONNX files

• draw_(graph).py:

• measure_latency: Measure the latency of each layer

• pruning.py: Prune model files using Torch-Pruning library

  • You can choose model, pruning amount, fine-tuning epochs, layers to prune, and strategy using arguments
  • python pruning.py --model [model name] --finetune_epochs [finetuning epochs] --prune [pruning amount] --layer [layers to prune] --strategy [pruning strategy]

• pruning_pt.py: Prune model files given in class

• run.sh: Automatically run codes