open-metric-learning

You may think "If I need image embeddings I can simply train a vanilla classifier and take its penultimate layer". Well, it makes sense as a starting point. But there are several possible drawbacks:

• If you want to use embeddings to perform searching you need to calculate some distance among them (for example, cosine or L2). Usually, you don't directly optimize these distances during the training in the classification setup. So, you can only hope that final embeddings will have the desired properties.

• The second problem is the validation process. In the searching setup, you usually care how related your top-N outputs are to the query. The natural way to evaluate the model is to simulate searching requests to the reference set and apply one of the retrieval metrics. So, there is no guarantee that classification accuracy will correlate with these metrics.

• Finally, you may want to implement a metric learning pipeline by yourself. There is a lot of work: to use triplet loss you need to form batches in a specific way, implement different kinds of triplets mining, tracking distances, etc. For the validation, you also need to implement retrieval metrics, which include effective embeddings accumulation during the epoch, covering corner cases, etc. It's even harder if you have several gpus and use DDP. You may also want to visualize your search requests by highlighting good and bad search results. Instead of doing it by yourself, you can simply use OML for your purposes.

PML is the popular library for Metric Learning, and it includes a rich collection of losses, miners, distances, and reducers; that is why we provide straightforward examples of using them with OML. Initially, we tried to use PML, but in the end, we came up with our library, which is more pipeline / recipes oriented. That is how OML differs from PML:

• OML has Pipelines which allows training models by preparing a config and your data in the required format (it's like converting data into COCO format to train a detector from mmdetection).

• OML focuses on end-to-end pipelines and practical use cases. It has config based examples on popular benchmarks close to real life (like photos of products of thousands ids). We found some good combinations of hyperparameters on these datasets, trained and published models and their configs. Thus, it makes OML more recipes oriented than PML, and its author confirms this saying that his library is a set of tools rather the recipes, moreover, the examples in PML are mostly for CIFAR and MNIST datasets.

• OML has the Zoo of pretrained models that can be easily accessed from the code in the same way as in torchvision (when you type resnet50(pretrained=True)).

• OML is integrated with PyTorch Lightning, so, we can use the power of its Trainer. This is especially helpful when we work with DDP, so, you compare our DDP example and the PMLs one. By the way, PML also has Trainers, but it's not widely used in the examples and custom train / test functions are used instead.

We believe that having Pipelines, laconic examples, and Zoo of pretrained models sets the entry threshold to a really low value.

Metric Learning problem (also known as extreme classification problem) means a situation in which we have thousands of ids of some entities, but only a few samples for every entity. Often we assume that during the test stage (or production) we will deal with unseen entities which makes it impossible to apply the vanilla classification pipeline directly. In many cases obtained embeddings are used to perform search or matching procedures over them.

Here are a few examples of such tasks from the computer vision sphere:

• Person/Animal Re-Identification
• Face Recognition
• Landmark Recognition
• Searching engines for online shops and many others.

• embedding - model's output (also known as features vector or descriptor).
• query - a sample which is used as a request in the retrieval procedure.
• gallery set - the set of entities to search items similar to query (also known as reference or index).
• Sampler - an argument for DataLoader which is used to form batches
• Miner - the object to form pairs or triplets after the batch was formed by Sampler. It's not necessary to form the combinations of samples only inside the current batch, thus, the memory bank may be a part of Miner.
• Samples/Labels/Instances - as an example let's consider DeepFashion dataset. It includes thousands of fashion item ids (we name them labels) and several photos for each item id (we name the individual photo as instance or sample). All of the fashion item ids have their groups like "skirts", "jackets", "shorts" and so on (we name them categories). Note, we avoid using the term class to avoid misunderstanding.
• training epoch - batch samplers which we use for combination-based losses usually have a length equal to [number of labels in training dataset] / [numbers of labels in one batch]. It means that we don't observe all of the available training samples in one epoch (as opposed to vanilla classification), instead, we observe all of the available labels.

It may be comparable with the current (2022 year) SotA methods, for example, Hyp-ViT. (Few words about this approach: it's a ViT architecture trained with contrastive loss, but the embeddings were projected into some hyperbolic space. As the authors claimed, such a space is able to describe the nested structure of real-world data. So, the paper requires some heavy math to adapt the usual operations for the hyperbolical space.)

We trained the same architecture with triplet loss, fixing the rest of the parameters: training and test transformations, image size, and optimizer. See configs in Models Zoo. The trick was in heuristics in our miner and sampler:

• Category Balance Sampler forms the batches limiting the number of categories C in it. For instance, when C = 1 it puts only jackets in one batch and only jeans into another one (just an example). It automatically makes the negative pairs harder: it's more meaningful for a model to realise why two jackets are different than to understand the same about a jacket and a t-shirt.

• Hard Triplets Miner makes the task even harder keeping only the hardest triplets (with maximal positive and minimal negative distances).

Here are CMC@1 scores for 2 popular benchmarks. SOP dataset: Hyp-ViT — 85.9, ours — 86.6. DeepFashion dataset: Hyp-ViT — 92.5, ours — 92.1. Thus, utilising simple heuristics and avoiding heavy math we are able to perform on SotA level.

Recent research