Characterizing Generalization under Out-of-Distribution shifts in Deep Metric Learning

Links: [Paper]

---

Description

This repository contains code and implementations used in our Generalization Study on Deep Metric Learning under OOD Shifts. The underlying DML pipeline is adapted from this repository, and introduces the following novel key elements from our paper:

• It introduces ooDML, a DML benchmark comprising new and progressively harder data distribution shifts based on existing DML benchmarks (see image: black line is default split 'ood-ness', blue the possible split progression generated. Red denotes even harder splits generated with class removal). We also provide create_dataset_splits.py to apply this process to potentially other benchmarks.

• ooDML allows us to evaluate how conceptual differences hold up in more realistic and especially more challenging train-to-test distribution shifts. The change in 'Out-of-Distribution-ness' is visually highlighted when looking at each benchmark UMAP for different splits:

• Different methods and approaches to Deep Metric Learning can then simply be evaluated against a progressively more out-of-distribution train-to-test split:

• It also introduces few-shot DML as a cost-effective approach to help with retrieval under especially hard OOD shifts.

• It provides the respective scripts to run a large and diverse set, included more recent objectives and extension to DML on our proposed benchmark, ooDML.

If you wish to use the ooDML splits with your own codebase, simply copy the corresponding pickle-files from datasplits. The internal structure of the pickled dicts is explained below!

Expect further updates in the next weeks!

---

Quick Usage

• Example runs as used to produce key results in our paper are made available in OOD-Study_ZeroShot.sh for zero-shot generalization experiments and OOD-Study_FewShot.sh for few-shot generalization experiments.
• Note that for readability, runs with two different seeds are shown, but for conclusive experiments, more should be used.
• In addition, results may vary a bit, as the results published in the original paper were computed over multiple different hardware settings.
• To download the benchmark data (this repository only provides the splits as lists of train and test classes to use), simply use the download links or guides provided here. The data structure is assumed to be the same for this repository.

---

Requirements:

• PyTorch 1.2.0+ & Faiss-Gpu
• Python 3.6+
• pretrainedmodels, timm, torchvision 0.3.0+

An exemplary setup of a virtual environment containing everything needed:

(1) wget  https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
(2) bash Miniconda3-latest-Linux-x86_64.sh (say yes to append path to bashrc)
(3) source .bashrc
(4) conda create -n DL python=3.6
(5) conda activate DL
(6) conda install matplotlib scipy scikit-learn scikit-image tqdm pandas pillow
(7) conda install pytorch torchvision faiss-gpu cudatoolkit=10.0 -c pytorch
(8) pip install wandb pretrainedmodels
(9) Run the scripts!

---

The ooDML benchmark

Information about the utilised split progressions is fully available in the .pkl-files located in the datasplits directory. For each base benchmark, the respective .pkl-file contains a dictionary with the following structure:

Split_ID: 1, 2, ..., 8/9 (depending on benchmark)
└─── train: list of training classes.
|       └─── String of classname
|       ...
└─── test: List of test classes.
|       └─── String of classname
|       ...
└─── fid: FID score (using R50) for given split.
└─── test_episodes: Episode data for few-shot evaluation of DML, given per Shot-setup.
|       └─── 2
|       |    └─── Episode-ID: 1, ..., 10
|       |    |       └─── classname
|       |    |       |       └─── Support samples to use for few-shot adaptation. The complement will be used to generate the query data.
|       |    |       |       ...
|       └─── 5 (see 2-Shot setting)
|       ...
|       └─── 10 (not used in the paper)
|       ...
└─── split_train, split_val: Train/val splits for hyperparameter tuning before running on final test set. Also used in few-shot experiments to provide default train/validation splits.
|       ...

---

Detailed Usage

There are four key scripts in this repository:

• ood_main.py and ood_diva_main.py to run zero-shot ooDML-experiments on all DML criteria.
• fewshot_ood_main.py and fewshot_ood_diva_main.py to run fewshot-shot ooDML-experiments on all DML criteria.

Here are exemplary uses for each script:

python ood_main.py --checkpoint --data_hardness $split_id --kernels 6 --source $datapath --n_epochs 200 --log_online --project DML_OOD-Shift_Study --group CUB_ID-1_Margin_b12_Distance --seed 0 --gpu $gpu --bs 112 --samples_per_class 2 --loss margin --batch_mining distance --arch resnet50_frozen_normalize --embed_dim 512

python ood_diva_main.py --checkpoint --data_hardness $split_id --kernels 6 --source $datapath --n_epochs 200 --log_online --project DML_OOD-Shift_Study --group CUB_ID-1_DiVA --seed 0 --gpu $gpu --bs 108 --samples_per_class 2 --loss margin --batch_mining distance --diva_rho_decorrelation 1500 1500 1500 --diva_alpha_ssl 0.3 --diva_alpha_intra 0.3 --diva_alpha_shared 0.3 --arch resnet50_frozen_normalize --embed_dim 128

python fewshot_ood_main.py --dataset cars196 --finetune_criterion margin --finetune_shots 2 --finetune_lr_multi 10 --finetune_iter 1000 --finetune_only_last --checkpoint --data_hardness -20 --kernels 6 --source $datapath --n_epochs 200 --log_online --project DML_OOD-Shift_FewShot_Study --group CAR_Shots-2_ID-1_Multisimilarity --seed 0 --gpu $gpu --bs 112 --samples_per_class 2 --loss multisimilarity --arch resnet50_frozen_normalize --embed_dim 512

python fewshot_ood_main.py --dataset cars196 --finetune_criterion margin --finetune_shots 2 --finetune_lr_multi 10 --finetune_iter 1000 --finetune_only_last --checkpoint --data_hardness -20 --kernels 6 --source $datapath --n_epochs 200 --log_online --project DML_OOD-Shift_FewShot_Study --group CAR_Shots-2_ID-1_Multisimilarity --seed 0 --gpu $gpu --bs 112 --samples_per_class 2 --loss multisimilarity --arch resnet50_frozen_normalize --embed_dim 512

An explanation for all utilised flags is provided in the respective help-string in parameters.py.

As noted previously, the data is assumed to have the following structure:

• For CUB200-2011/CARS196:

cub200/cars196
└───images
|    └───001.Black_footed_Albatross
|           │   Black_Footed_Albatross_0001_796111
|           │   ...
|    ...

• For Stanford Online Products:

online_products
└───images
|    └───bicycle_final
|           │   111085122871_0.jpg
|    ...
|
└───Info_Files
|    │   bicycle.txt
|    │   ...

Assuming your folder is placed in e.g. <$datapath/cub200>, pass $datapath as input to --source.

---

Citation

If you use the code in this repository, insights and/or the ooDML benchmark, we kindly ask that you cite

@inproceedings{milbich2021characterizing,
      title={Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning},
      author={Timo Milbich and Karsten Roth and Samarth Sinha and Ludwig Schmidt and Marzyeh Ghassemi and Björn Ommer},
      year={2021},
      booktitle={Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS)}
}

Thanks! :)