This is the official PyTorch implementation of the paper:

Gated Transfer Network for Transfer Learning

Yi Zhu and Jia Xue and Shawn Newsam

ACCV 2018

Installation

We recommend using a Conda environment. We use PyTorch 1.1, CUDA 9.0 and python 3.7.

conda create -n gtn python=3.7
conda activate gtn
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
pip install easydict

Data Preparation

Please see datasets README for more details.

Experiments

We take CUB200 as an example in the experiments folder, other experiments are similar except some hyper-parameter changes.

• Set config.py correctly (dataset path, hyper-paramters, etc.)

• python train.py

• Evaluation is done on-the-fly.

Note that, the evaluation performance on UCF101 is not the final results because it is a video dataset. If you need the final clip-level results, you need to perform aggregation (example script can be found here).

Citation

If you use this code for your research, please consider citing our paper:

@inproceedings{zhu2018GTN,
  author    = {Yi Zhu and Jia Xue and Shawn Newsam},
  title     = {Gated Transfer Network for Transfer Learning},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2018}
}