Skip to content

Export utility for unconstrained channel pruned models

License

Notifications You must be signed in to change notification settings

apple/ml-upscale

Repository files navigation

Unconstrained Channel Pruning · Paper

Open In Colab

UPSCALE: Unconstrained Channel Pruning @ ICML 2023
Alvin Wan, Hanxiang Hao, Kaushik Patnaik, Yueyang Xu, Omer Hadad, David Güera, Zhile Ren, Qi Shan

By removing constraints from existing pruners, we improve ImageNet accuracy for post-training pruned models by 2.1 points on average - benefiting DenseNet (+16.9), EfficientNetV2 (+7.9), and ResNet (+6.2). Furthermore, for these unconstrained pruned models, UPSCALE improves inference speeds by up to 2x over a baseline export.

Quick Start

Install our package.

pip install apple-upscale

Mask and prune channels, using the default magnitude pruner.

import torch, torchvision
from upscale import MaskingManager, PruningManager

x = torch.rand((1, 3, 224, 224), device='cuda')
model = torchvision.models.get_model('resnet18', pretrained=True).cuda()  # get any pytorch model
MaskingManager(model).importance().mask()
PruningManager(model).compute([x]).prune()

Customize Pruning

We provide a number of pruning heuristics out of the box:

You can pass the desired heuristic into the UpscaleManager.mask method call. You can also configure the pruning ratio in UpscaleManager.mask. A value of 0.25 means 25% of channels are set to zero.

from upscale.importance import LAMP
MaskingManager(model).importance(LAMP()).mask(amount=0.25)

You can also zero out channels using any method you see fit.

model.conv0.weight[:, 24] = 0

Then, run our export.

PruningManager(model).compute([x]).prune()

Advanced

You may want direct access to network segments to build a heavily-customized pruning algorithm.

for segment in MaskingManager(model).segments():
    # prune each segment in the network independently
    for layer in segment.layers:
        # layers in the segment

Development

NOTE: See src/upscale/pruning/README.md for more details on how the core export algorithm code is organized.

Clone and setup.

git clone git@github.com:apple/ml-upscale.git
cd upscale
pip install -e .

Run tests.

py.test src tests --doctest-modules

Paper

Follow the development installation instructions to have the paper code under paper/ available.

To run the baseline unconstrained export, pass baseline=True to PruningManager.prune.

PruningManager(model).compute([x]).prune(baseline=True)

To reproduce the paper results, run

python paper/main.py resnet18

Plug in any model in the torchvision.models namespace.

usage: main.py [-h] [--side {input,output} [{input,output} ...]]
               [--method {constrained,unconstrained} [{constrained,unconstrained} ...]]
               [--amount AMOUNT [AMOUNT ...]] [--epochs EPOCHS] 
               [--heuristic {l1,l2,lamp,fpgm,hrank}] [--global] [--out OUT] 
               [--force] [--latency] [--clean]
               model

positional arguments:
  model                 model to prune

options:
  -h, --help            show this help message and exit
  --side {input,output} [{input,output} ...]
                        prune which "side" -- producers, or consumers
  --method {constrained,unconstrained} [{constrained,unconstrained} ...]
                        how to handle multiple branches
  --amount AMOUNT [AMOUNT ...]
                        amounts to prune by. .6 means 60 percent pruned
  --epochs EPOCHS       number of epochs to train for
  --heuristic {l1,l2,lamp,fpgm,hrank}
                        pruning heuristic
  --global              apply heuristic globally
  --out OUT             directory to write results.csv to
  --force               force latency rerun
  --latency             measure latency locally
  --clean               clean the dataframe

Citation

If you find this useful for your research, please consider citing

@inproceedings{wan2023upscale,
  title={UPSCALE: Unconstrained Channel Pruning},
  author={Alvin Wan and Hanxiang Hao and Kaushik Patnaik and Yueyang Xu and Omer Hadad and David Guera and Zhile Ren and Qi Shan},
  booktitle={ICML},
  year={2023}
}