jieeee / qfsl-zeroshot

Transductive Unbiased Embedding for Zero-Shot Learning

Transductive Unbiased Embedding for Zero-Shot Learning

Code supporting the paper:

Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, Mingli Song. Transductive Unbiased Embedding for Zero-Shot Learning. CVPR 2018

If you find this code useful in your research, please consider citing using the following BibTeX entry:

@InProceedings{Jie CVPR2018,
author = {Jie Song, Chengchao Shen, Yezhou Yang, Yang Liu, Mingli Song},
title = {Transductive Unbiased Embedding for Zero-Shot Learning},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2018}
}

Installation

This code uses Python 3.8 and PyTorch 1.9.0 cuda version 10.2.

• Installing PyTorch:

$ conda install pytorch==1.9.0 torchvision==0.10.0 torchaudio==0.9.0 cudatoolkit=10.2 -c pytorch

• Install dependencies

$ pip install -r requirements.txt

Tour

code

1.ims2file.py: Script for preparing LMDBs used in train.py and sample.py.

2.train.py: Training script.

3.sample.py: Evaluation script.

4.QFSLnet.py: Defines the QFSLmodel.

data

1.Partition into source and target classes: classes.txt,trainvalclasses.txt,testclasses.txt .

2.class attributes names: attributes.txt

3.class attributes labels: class_attribute_labels_continuous.txt

Dataset

• Download AWA2
• Download CUB
• Download SUN

Images to Image_LMDB and Information_json

Prepare for training

$ python ims2file.py --dataset_name AWA2/CUB/SUN --class_names_path path_to_dataset's_classes_names \
--dataset_img_path path_to_dataset's_Imagedir --dataset_attr_path path_to_allclasses's_attributes \
--save_path path_for_LMDB_to_save

Training

We train our model in this way:

python train.py --dataset_name AWA2 --img_encoder_name AlexNet/ResNet101/VGG19/GoogLeNet \
--train_class_path path_to_source_class --test_class_path path_to_target_class \ 
--data_path path_to_Information_json --img_path path_to_lmdb --learning_rate 0.005 \ 
--bias_weight 0.2  --num_epochs 5000 --batch_size 64 --num_workers 4   

Example

python train.py --dataset_name AWA2 --img_encoder_name AlexNet \ 
--train_class_path ../data/AWA2/standard_split/trainvalclasses.txt \ 
--test_class_path ../data/AWA2/standard_split/testclasses.txt \ 
--data_path ../data_save/AWA2/data_info.json \ 
--img_path ../data_save/AWA2/lmdb \ 
--learning_rate 0.005 --bias_weight 0.2 --num_epochs 5000 --batch_size 64 --num_workers 4 

Check training progress in src/checker/logger: Model save in src/checker/checkpoints:

Evaluation

Example

python sample.py --dataset_name AWA2 --img_encoder_name AlexNet \
 --train_class_path ../data/AWA2/standard_split/trainvalclasses.txt \
 --test_class_path ../data/AWA2/standard_split/testclasses.txt  \
  --data_path ../data_save/AWA2/data_info.json \ 
 --img_path ../data_save/AWA2/lmdb  --batch_size 64 --num_workers 4

• This script will return Mean class accuracy for target classes in conventional setting and generalized setting