https://github.com/arunmallya/packnet

Code for PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning
https://github.com/arunmallya/packnet

Last synced: 2 months ago
JSON representation

Code for PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning

• Host: GitHub
• URL: https://github.com/arunmallya/packnet
• Owner: arunmallya
• Created: 2017-11-16T20:39:54.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2018-10-07T03:57:55.000Z (over 5 years ago)
• Last Synced: 2024-01-26T20:34:11.577Z (5 months ago)
• Language: Python
• Homepage: https://arxiv.org/abs/1711.05769
• Size: 18.6 KB
• Stars: 223
• Watchers: 8
• Forks: 39
• Open Issues: 3
• Metadata Files:
  • Readme: README.md

Lists

• Awesome-pytorch-list - packnet
• Awesome-pytorch-list-CNVersion - packnet

README

        
## PackNet: https://arxiv.org/abs/1711.05769

Pretrained models are available here: https://uofi.box.com/s/zap2p03tnst9dfisad4u0sfupc0y1fxt  

Datasets in PyTorch format are available here: https://uofi.box.com/s/ixncr3d85guosajywhf7yridszzg5zsq  

The PyTorch-friendly Places365 dataset can be downloaded from http://places2.csail.mit.edu/download.html  

Place models in `checkpoints/` and unzipped datasets in `data/`

|               |   VGG-16 LwF |    VGG-16    |   VGG-16 BN  |   ResNet-50  | DenseNet-121 |

|:-------------:|:------------:|:------------:|:------------:|:------------:|:------------:|

| ImageNet      | 36.58 (14.75)| 29.19 (9.90) | 27.10 (8.70) | 24.33 (7.17) | 25.51 (7.85) |

| CUBS          |        34.24 |        22.56 |        20.43 |        19.59 |        20.11 |

| Stanford Cars |        22.07 |        17.09 |        14.92 |        14.03 |        16.18 |

| Flowers       |        12.15 |        11.07 |         8.59 |         8.12 |         9.07 |

Note that the numbers in the [paper](https://arxiv.org/abs/1711.05769) are averaged over multiple runs for each ordering

of datasets. The pretrained models are for a specific dataset addition ordering: (c) CUBS Birds, (s) Stanford Cars, (f) Flowers.

These numbers were obtained by evaluating the models on a Titan X (Pascal).  

Note that numbers on other GPUs might be slightly different (~0.1%) owing to cudnn algorithm selection.  

https://discuss.pytorch.org/t/slightly-different-results-on-k-40-v-s-titan-x/10064

## Requirements:

```

Python 2.7 or 3.xx

torch==0.2.0.post3

torchvision==0.1.9

torchnet (pip install git+https://github.com/pytorch/tnt.git@master)

tqdm (pip install tqdm)

```

## Training:

Check out the scripts in `src/scripts`.  

Run all code from the `src/` directory, e.g. `./scripts/run_all.sh`

## Eval:

```bash

cd src  # Run everything from src/

# Pruning-based models.

python main.py --mode eval --dataset cubs_cropped \

  --loadname ../checkpoints/csf_0.75,0.75,-1_vgg16_0.5-nobias-nobn_1.pt

# LwF models.

python lwf.py --mode eval --dataset cubs_cropped \

  --loadname ../checkpoints/csf_lwf.pt

```