pwenjay / gca-hng

GCA-HNG: Globally Correlation-Aware Hard Negative Generation

This repository serves as the official PyTorch implementation for the paper: Globally Correlation-Aware Hard Negative Generation.

It offers source code for replicating the experiments conducted on four benchmark datasets (CUB-200-2011, Cars196, SOP, and InShop) and releases the pretrained metric models.

Requirements

• Python 3.8
• PyTorch 1.8.1+cu111
• torch_scatter 2.0.9
• numpy
• tqdm
• tensorboardX
• scikit-learn
• scipy

Datasets

1. Download four benchmark datasets.
   • CUB-200-2011
   • Cars196
   • Stanford Online Products
   • InShop Clothes Retrieval
2. Extract the tgz or zip file into ./data/dataset_name/original/ folder and run the data convert scripts to transform data format. In particular, the InShop dataset does not require any data transformation.

python scripts/data_process/xxx_convert.py

3. The data folder is constructed as followed:

data:
├──  CUB200/CARS196/SOP
│  └──  class 
│    ├──  train 
│    │  ├──  Catrgory 1
│    │  ├──  Catrgory ...
│    │  └──  Catrgory M
│    └──  test
│       ├──  Catrgory 1
│       ├──  Catrgory ...
│       └──  Catrgory N
└──  IN_SHOP
   ├── img  
   │  ├──  MEN
   │  └──  WOMEN
   └──  list_eval_partition.txt

Training Process

Run the training process by designating the corresponding yml file.

python train.py --cfg scripts/cfgs/xxx.yml

Image Retrieval Evaluation

The released metric model is available at Google Drive.

Given the (full model/ metric model) with pretrained weights, run the evaluation process as follows:

python eval.py --cfg scripts/cfgs/xxx.yml --model_path xxx.pth