This repository the official PyTorch implementation of Learning Intra-Batch Connections for Deep Metric Learning. The config files contain the same parameters as used in the paper.

We use torch 1.7.1 and torchvision 0.6.0. While the training and inference should be able to be done correctly with the newer versions of the libraries, be aware that at times the network trained and tested using versions might diverge or reach lower results. We provide a evironment.yaml file to create a corresponding conda environment.

We also support mixed-precision training via Nvidia Apex and describe how to use it in usage.

As in the paper we support training on 4 datasets: CUB-200-2011, CARS 196, Stanford Online Products and In-Shop datasets.

The majority of experiments are done using ResNet50. We provide support for the entire family of ResNet and DenseNet as well as BN-Inception.

1. Clone and enter this repository:

    git clone https://github.com/dvl-tum/intra_batch.git
   
    cd intra_batch
   

2. Create an Anaconda environment for this project: To set up a conda environment containing all used packages, please fist install anaconda and then run

   1.   conda env create -f environment.yml
      

   2.  conda activate intra_batch_dml
      

   3.  pip install torch-scatter==2.0.5 -f https://pytorch-geometric.com/whl/torch-1.5.0+cu102.html
      

   4. If you want to use Apex, please follow the installation instructions on https://github.com/NVIDIA/apex

3. Download datasets: Make a data directory by typing

    mkdir data
   

   Then download the datasets using the following links and unzip them in the data directory:

   • CUB-200-2011: http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
   • Cars196: https://vision.in.tum.de/webshare/u/seidensc/intra_batch_connections/CARS.zip
   • Stanford Online Products: https://cvgl.stanford.edu/projects/lifted_struct/
   • In-Shop: http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion/InShopRetrieval.html

   We also provide a parser for Stanford Online Products and In-Shop datastes. You can find dem in the dataset/ directory. The datasets are expected to be structured as dataset/images/class/, where dataset is either CUB-200-2011, CARS, Stanford_Online_Products or In_shop and class are the classes of a given dataset. Example for CUB-200-2011:

        CUB_200_2011/images/001
        CUB_200_2011/images/002
        CUB_200_2011/images/003
        ...
        CUB_200_2011/images/200
   

4. Download our models: Please download the pretrained weights by using

    wget https://vision.in.tum.de/webshare/u/seidensc/intra_batch_connections/best_weights.zip
   

   and unzip them.

You can find config files for training and testing on each of the datasets in the config/ directory. For training and testing, you will have to input which one you want to use (see below). You will only be able to adapt some basic variables over the command line. For all others please refer to the yaml file directly.

To test to networks choose one of the config files for testing, e.g., config_cars_test.yaml to evaluate the performance on Cars196 and run:

python train.py --config_path config_cars_test.yaml --dataset_path <path to dataset> 

The default dataset path is data.

To train a network choose one of the config files for training like config_cars_train.yaml to train on Cars196 and run:

python train.py --config_path config_cars_train.yaml --dataset_path <path to dataset> --net_type <net type you want to use>

Again, if you don't specify anything, the default setting will be used. For the net type you have the following options:

resnet18, resnet32, resnet50, resnet101, resnet152, densenet121, densenet161, densenet16, densenet201, bn_inception

If you want to use apex add --is_apex 1 to the command.

If you find this code useful, please consider citing the following paper:

@inproceedings{DBLP:conf/icml/SeidenschwarzEL21,
  author    = {Jenny Seidenschwarz and
               Ismail Elezi and
               Laura Leal{-}Taix{\'{e}}},
  title     = {Learning Intra-Batch Connections for Deep Metric Learning},
  booktitle = {Proceedings of the 38th International Conference on Machine Learning,
               {ICML} 2021, 18-24 July 2021, Virtual Event},
  series    = {Proceedings of Machine Learning Research},
  volume    = {139},
  pages     = {9410--9421},
  publisher = {{PMLR}},
  year      = {2021},
}