Resnet50-pruning

This work implements the pruning of resnet50 cat and dog classification model. Pruning tool based on https://github.com/VainF/Torch-Pruning.

environment

numpy torch torchvison matplotlib tqdm mnn

Quickstart

You may need to change your own path in the code.

dataset

wget -c https://download.microsoft.com/download/3/E/1/3E1C3F21-ECDB-4869-8368-6DEBA77B919F/kagglecatsanddogs_3367a.zip

Split the dataset into train and test at a ratio of 0.1, both of which contain cat and dog.

cd resnet50_catdog

Basic training:

python train.py --sr False

Sparse training:

python train.py --sr True --s 0.0001

Pruning and finetune:

python prune.py --percent 0.8 --sr False

Test

python test.py

mnn inference

to onnx

python pth2onnx.py

onnx2mnn

mnnconvert [-h] -f {TF,CAFFE,ONNX,TFLITE,MNN} --modelFile MODELFILE
                  [--prototxt PROTOTXT] --MNNModel MNNMODEL [--fp16 FP16]

quantize

Prepare feature quantization image

mnnquant src_mnn dst_mnn config

mnn inference

python mnn_inference.py