Pytorch implementation of Deep Compression
Dataset: CIFAR10 Models: VGG16, Resnet18~
- Image classification
- Image segmentation
- Image detection
To start with..
clone the repo to your local.
git clone https://github.com/kentaroy47/Deep-Compression.Pytorch
The CIFAR10 trained models are in checkpoint.
| VGG | Res18 |
|---|---|
| 85.8% | 89.8% |
Resnet is still underfitting.
To do pruning..
The pretrained resnet18 should be in checkpoint/res18.t7. The val accuracy is about 90%.
# This will prune the default resnet18 with 50% parameters of conv2d.
# It will do 20 epochs of retraining as well.
python prune.py
Pruning results for resnet18.
| no pruning | 10% pruned | 25% pruned | 50% pruned | 75% pruned |
|---|---|---|---|---|
| 89.8% | 90.3% | 91.2% | 91.7% | 91.9% |
Even with 50% pruned weights, the accuracy is about the same (90%).
You can try arbitary pruning rates:
# This will prune 75% of conv parameters.
python prune.py --prune 0.75
To train your own model
I provided a CIFAR 10 trained resnet18 checkpoint to get started with.
If you want to train your own network, do:
# for vgg and resnet50
python train_cifar10.py --net vgg
python train_cifar10.py --net res50
Training script
You can train multiple pruned resnet by script.
sh prune.sh
TODO:
Quantization HuffmanCoding