This repository contains reference implementations of all methods introduced in our ICML 2022 paper: SPDY: Accurate Pruning with Speedup Guarantees. This includes the DP algorithm for efficiently solving constrained layer-wise compression problems (see dpsolve() in spdy.py), the reparametrized SPDY search for injecting global information into otherwise purely layer-wise problems (see spdy.py), as well as the enhanced post trainig pruning method global AdaPrune (see adaprune.py).

The code depends on torch and torchvision. The following block shows sample commands for the various features of the repository. See --help of adaprune.py and spdy.py for additional options.

# Path to ImageNet
export DATAPATH = path/to/imagenet

# 2:4 AdaPrune   global AdaPrune
python adaprune.py rn18 imagenet nmprune --datapath ${DATAPATH}
# 4:8 AdaPrune   global AdaPrune
python adaprune.py rn18 imagenet nmprune --nmblocksize 8 --datapath ${DATAPATH}

# Generate unstr database
python adaprune.py rn18 imagenet gen --collect_to rn18_unstr --datapath ${DATAPATH}
# Generate 4block database
python adaprune.py rn18 imagenet gen --blocksize 4 --collect_to rn18_4block --datapath ${DATAPATH}

# Run SPDY search to find 2x speedup profile
python spdy.py rn18 imagenet rn18_unstr timings/rn18_unstr.txt 2 rn18_unstr_200x.txt --datapath ${DATAPATH}

# Load and evaluate profile & run global AdaPrune
python adaprune.py rn18 imagenet load --stitch_from rn18_unstr --profile rn18_unstr_200x.txt --datapath ${DATAPATH}

Currently, the repository supports several torchvision-ResNet variants. However, all the core features are implemented so that they can also easily be applied to other models by providing a few corresponding small wrapper functions, see modelutils.py and datautils.py for their ResNet implementations.

In order to work with YOLOv5 models you need to install all packages in yolov5/requirements.txt and then create a COCO calibration data folder as follows.

cp yolov5/make_calib.py COCO_ROOT
cd COCO_ROOT
python make_calib.py

After this YOLO pruning can be performed in similar fashion as described above for ResNet18.

For gradual pruning, we used SparseML v0.9; our integration is contained in the yolov5 folder. For more extensive work we would however recommend to use the official integration which supports the newest versions of YOLOv5 and SparseML. Sample SparseML pruning recipes with all our hyper-parameters (e.g. yolov5s_150x_spdy.yaml) and launch scripts (e.g. yolov5s_150x_spdy.sh) can be found in our yolov5 folder as well.

If you found this work useful, please consider citing:

@inproceedings{frantar-spdy,
  title={{SPDY}: Accurate Pruning with Speedup Guarantees}, 
  author={Elias Frantar and Dan Alistarh},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2022}
}