README.md

AutoMC

The source code of "AutoMC: Automated Model Compression Based on Domain Knowledge and Progressive Search".

Function

AutoMC is an effective and efficient automatic tool for model compression that can automatically design the effective model compression scheme according to the user demands.

1. Pretrain

To get pretrained model of VGG-13, VGG-16, VGG-19, ResNet-20, ResNet-56, ResNet-164 on CIFAR-10 and CIFAR-100, setting pretrained epoch as 200:

cd ./CAlgs
python scripts/pretrain.py {0} {1} 200 
# {0} = cifar10 or cifar100
# {1} = vgg13, vgg16, vgg19, resnet20, resnet56 or resnet164

IMPOTRTANT: After training the models, copy the model file from snapshots folder (eg. ./CAlgs/snapshots/cifar10/vgg13/train/best.finetune.*.pth.tar) to trianed_models folder (eg. ./CAlgs/trianed_models/cifar10/vgg13.pth.tar).

The trained models also are available to download at: https://drive.google.com/file/d/1t9aSagkk0dbjPmxLnXJBfD1z3T9e6nZH/view?usp=sharing

2. Sample Dataset

Sample a mini dataset from the whole dataset.

To sample mini-CIFAR-10 and mini-CIFAR-100 from CIFAR-10 and CIFAR-100 for ResNet-56 and VGG-16 respectively (sample rate is 10%):

cd ./CAlgs
python generate_mini_dataset.py {0} {1}
# {0} = cifar10 or cifar100
# {1} = vgg16 or resnet56

3. Pretrain Models on Sampled Dataset

Pretrain models on sampled dataset.

To train VGG-16 and ResNet-56 on mini-CIFAR-10 and mini-CIFAR-100, setting trained epoch as 50:

cd ./CAlgs
python scripts/pretrain.py {0} {1} 50
# {0} = mini_cifar10 or mini_cifar100
# {1} = vgg16 or resnet56

The trained models also are available to download at the link above.

4. Get Results of 6 State-of-the-art Human-invented Compression Methods

We represent LMA, LeGR, NS, SFP, HOS, LFB as 1, 2, 3, 4, 5, 7 in our code.

To get results of 6 human-invented compression methods of VGG-13, VGG-16, VGG-19, ResNet-20, ResNet-56 and ResNet-164 on CIFAR-10 and CIFAR-100, setting compression rate as 0.3 and 0.6:

cd ./CAlgs
python scripts/run.py {0} {1} {2} {3}
# {0} = cifar10 or cifar100
# {1} = vgg13, vgg16, vgg19, resnet20, resnet56 or resnet164
# {2} = 1, 2, 3, 4, 5, 7
# {3} = 0.3 or 0.6

5. Get Results of NAS Algorithms