DVBE

This is an implementation for Domain-aware Visual Bias Eliminating for Generalized Zero-Shot Learning, which has been accepted by CVPR2020.

DVBE is a new state-of-the-art method for generalized zero-shot learning.

Introduction

This project is a basic implementation of DVBE by pytorch platform.

To do:

1. adding the autoML part.
2. publishing the segmentation version.

Requirements

1. Python 3.6

2. Pytorch 0.4.0

3. CUDA 8.0

Implementations

Datasets Prepare

1. Downloading correspond dataset, e.g., CUB, AWA2, aPY, and SUN. Assume your data path is ${PATH}. A provided url is: (https://pan.baidu.com/s/1RYCZzKOuhDObuO-l-Ig78A 73rw)

2. Changing the data path around the line 190 of main.py, according to your ${PATH}.

Two-stage Training

The training examples for the four datasets have been given in ./cub, ./awa2, ./apy, and ./sun.

In details, the training processing of DVBE consists of two stages, which is:

1. Run train.py to train DVBE with fixed backbone

   e.g. for training CUB

   python main.py -a dvbe -d cub -s /output --backbone resnet101 -b 128 
   				--pretrained --is_fix

2. Finetune the whole DSEN

   e.g. for training CUB

   python main.py -a dvbe -d cub -s /output --backbone resnet101
   				-b 16 --lr 0.001 \
   				--epoch 180 --resume ./checkpoints/fix.model

For reproducibility, a suggested seed can be cub: (5181,6803), awa2: (142,6059), apy: (119,4). Besies, better performance can be obtained by flipping test, when setting --args.val_flippingtest

The reimplementation results and models are soon provided!