Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

NetTailor [Project Page]

This repository contains the source code accompanying our CVPR 2019 paper.

NetTailor: Tuning the architecture, not just the weights.
Pedro Morgado, Nuno Vasconcelos.
In Computer Vision and Pattern Recognition, 2019.

@inproceedings{morgado_nettailor,
	title={NetTailor: Tuning the architecture, not just the weights},
	author={Pedro Morgado and Nuno Vasconcelos},
	booktitle={Computer Vision and Pattern recognition (CVPR)},
	year={2019}
}

Requirements

  1. pytorch & torchvision
  2. COCO API (from https://github.com/cocodataset/cocoapi)

Getting started

NOTE: This repo is still under development. If you find any issues running our code, missing files, etc, please do contact me.

Download and prepare demo datasets: SVHN, Flowers and Pascal VOC.

>> cd data && python prepare_data.py

If interested in Visual Decathlon results, download data and models.

>> cd data && python download_decathlon.py && cd ../checkpoints && python download_models.py

Training NetTailor

We demonstrate on three demo datasets how to use the NetTailor procedure to adapt the network architecture to a target task.

Code tour
  1. dataloaders.py: Dataloaders for demo datasets.
  2. nettailor.py and resnet.py: Model definition for the student and teacher networks, respectively, based on resnet backbone.
  3. main_student.py and main_teacher.py: Training and evaluation code for the student and teacher networks, respectively.

Usage: Refer to the three demo_xxx.py scripts. These demonstrates how to 1) train the teacher network, 2) train the student network and 3) prune and retrain the student.

To execute with default parameters, simply run:

>> python demo_svhn.py
>> python demo_flowers.py
>> python demo_voc12.py

Disclaimer: After publication, we modified the dataloader for the VOC12 and SVHN datasets to remove random cropping. This resulted in better performance than published results.

NetTailor on Visual Decathlon Challenge and trained models

We also release the final models obtained in the visual decathlon challenge. Universal blocks can be downloaded here and task-specific blocks here.

Code tour
  1. decathlon_dataloaders.py: Dataloaders for visual decathlon data.
  2. wide_nettailor.py and wide_resnet.py: Model definition for the student and teacher networks, respectively, based on wide resnet backbone.
  3. main_student_decathlon.py and main_teacher_decathlon.py: Training and evaluation code for the student and teacher networks, respectively.

Usage: Refer to the deploy_decathlon.ipynb notebook for a usage example. This notebook also showcases some predictions obtained with models trained with NetTailor on all nine datasets from the visual decathlon challenge.

If you encounter any issue when using our code or models, please let me know.

About

Source code accompanying our CVPR 2019 paper: "NetTailor: Tuning the architecture, not just the weights."

Resources

License

Releases

No releases published

Packages

No packages published