Why do I need OML?
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.
What is the difference between Open Metric Learning and PyTorch Metric Learning?
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.
What is Metric Learning?
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.
Glossary (Naming convention)
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 batchesMiner
- 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.
How good may be a model trained with OML?
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.
What about Self-Supervised Learning?
Recent research in SSL definitely obtained great results. The problem is that these approaches
required an enormous amount of computing to train the model. But in our framework, we consider the most common case
when the average user has no more than a few GPUs.
At the same time, it would be unwise to ignore success in this sphere, so we still exploit it in two ways:
- As a source of checkpoints that would be great to start training with. From publications and our experience,
they are much better as initialisation than the default supervised model trained on ImageNet. Thus, we added the possibility
to initialise your models using these pretrained checkpoints only by passing an argument in the config or the constructor.
- As a source of inspiration. For example, we adapted the idea of a memory bank from MoCo for the TripletLoss.
Do I need to know other frameworks to use OML?
No, you don't. OML is a framework-agnostic. Despite we use PyTorch Lightning as a loop
runner for the experiments, we also keep the possibility to run everything on pure PyTorch.
Thus, only the tiny part of OML is Lightning-specific and we keep this logic separately from
other code (see oml.lightning
). Even when you use Lightning, you don't need to know it, since
we provide ready to use Pipelines.
The possibility of using pure PyTorch and modular structure of the code leaves a room for utilizing
OML with your favourite framework after the implementation of the necessary wrappers.
Can I use OML without any knowledge in DataScience?
Yes. To run the experiment with Pipelines
you only need to write a converter
to our format (it means preparing the
.csv
table with 5 predefined columns).
That's it!
Probably we already have a suitable pre-trained model for your domain
in our Models Zoo. In this case, you don't even need to train it.
Can OML process texts, sounds and other modalities?
You can adapt OML to make it work not only with images.
Just open one of the examples and replace Dataset
remaining the rest of the pipeline the same or almost the same.
There is several people who successfully used OML for texts in their real-world projects.
Unfortunately, we don't have ready-to-use tutorials for this kind of usage at the moment.
OML is available in PyPI:
pip install -U open-metric-learning
You can also pull the prepared image from DockerHub...
docker pull omlteam/oml:gpu
docker pull omlteam/oml:cpu
Training
import torch
from tqdm import tqdm
from oml.datasets.base import DatasetWithLabels
from oml.losses.triplet import TripletLossWithMiner
from oml.miners.inbatch_all_tri import AllTripletsMiner
from oml.models import ViTExtractor
from oml.samplers.balance import BalanceSampler
from oml.utils.download_mock_dataset import download_mock_dataset
dataset_root = "mock_dataset/"
df_train, _ = download_mock_dataset(dataset_root)
extractor = ViTExtractor("vits16_dino", arch="vits16", normalise_features=False).train()
optimizer = torch.optim.SGD(extractor.parameters(), lr=1e-6)
train_dataset = DatasetWithLabels(df_train, dataset_root=dataset_root)
criterion = TripletLossWithMiner(margin=0.1, miner=AllTripletsMiner(), need_logs=True)
sampler = BalanceSampler(train_dataset.get_labels(), n_labels=2, n_instances=2)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_sampler=sampler)
for batch in tqdm(train_loader):
embeddings = extractor(batch["input_tensors"])
loss = criterion(embeddings, batch["labels"])
loss.backward()
optimizer.step()
optimizer.zero_grad()
print(criterion.last_logs)
Validation
import torch
from tqdm import tqdm
from oml.datasets.base import DatasetQueryGallery
from oml.metrics.embeddings import EmbeddingMetrics
from oml.models import ViTExtractor
from oml.utils.download_mock_dataset import download_mock_dataset
dataset_root = "mock_dataset/"
_, df_val = download_mock_dataset(dataset_root)
extractor = ViTExtractor("vits16_dino", arch="vits16", normalise_features=False).eval()
val_dataset = DatasetQueryGallery(df_val, dataset_root=dataset_root)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=4)
calculator = EmbeddingMetrics(extra_keys=("paths",))
calculator.setup(num_samples=len(val_dataset))
with torch.no_grad():
for batch in tqdm(val_loader):
batch["embeddings"] = extractor(batch["input_tensors"])
calculator.update_data(batch)
metrics = calculator.compute_metrics()
print(calculator.metrics)
print(calculator.metrics_unreduced)
calculator.get_plot_for_queries(query_ids=[0, 2], n_instances=5)
calculator.get_plot_for_worst_queries(metric_name="OVERALL/map/5", n_queries=2, n_instances=5)
calculator.visualize()
Training + Validation [Lightning and logging]
import pytorch_lightning as pl
import torch
from oml.datasets.base import DatasetQueryGallery, DatasetWithLabels
from oml.lightning.modules.extractor import ExtractorModule
from oml.lightning.callbacks.metric import MetricValCallback
from oml.losses.triplet import TripletLossWithMiner
from oml.metrics.embeddings import EmbeddingMetrics
from oml.miners.inbatch_all_tri import AllTripletsMiner
from oml.models import ViTExtractor
from oml.samplers.balance import BalanceSampler
from oml.utils.download_mock_dataset import download_mock_dataset
from oml.lightning.pipelines.logging import (
ClearMLPipelineLogger,
MLFlowPipelineLogger,
NeptunePipelineLogger,
TensorBoardPipelineLogger,
WandBPipelineLogger,
)
dataset_root = "mock_dataset/"
df_train, df_val = download_mock_dataset(dataset_root)
extractor = ViTExtractor("vits16_dino", arch="vits16", normalise_features=False)
optimizer = torch.optim.SGD(extractor.parameters(), lr=1e-6)
train_dataset = DatasetWithLabels(df_train, dataset_root=dataset_root)
criterion = TripletLossWithMiner(margin=0.1, miner=AllTripletsMiner())
batch_sampler = BalanceSampler(train_dataset.get_labels(), n_labels=2, n_instances=3)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_sampler=batch_sampler)
val_dataset = DatasetQueryGallery(df_val, dataset_root=dataset_root)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=4)
metric_callback = MetricValCallback(metric=EmbeddingMetrics(extra_keys=[train_dataset.paths_key,]), log_images=True)
logger = TensorBoardPipelineLogger(".")
pl_model = ExtractorModule(extractor, criterion, optimizer)
trainer = pl.Trainer(max_epochs=3, callbacks=[metric_callback], num_sanity_val_steps=0, logger=logger)
trainer.fit(pl_model, train_dataloaders=train_loader, val_dataloaders=val_loader)
Using a trained model for retrieval
import torch
from oml.const import MOCK_DATASET_PATH
from oml.inference.flat import inference_on_images
from oml.models import ViTExtractor
from oml.registry.transforms import get_transforms_for_pretrained
from oml.utils.download_mock_dataset import download_mock_dataset
from oml.utils.misc_torch import pairwise_dist
_, df_val = download_mock_dataset(MOCK_DATASET_PATH)
df_val["path"] = df_val["path"].apply(lambda x: MOCK_DATASET_PATH / x)
queries = df_val[df_val["is_query"]]["path"].tolist()
galleries = df_val[df_val["is_gallery"]]["path"].tolist()
extractor = ViTExtractor.from_pretrained("vits16_dino")
transform, _ = get_transforms_for_pretrained("vits16_dino")
args = {"num_workers": 0, "batch_size": 8}
features_queries = inference_on_images(extractor, paths=queries, transform=transform, **args)
features_galleries = inference_on_images(extractor, paths=galleries, transform=transform, **args)
use_knn = False
top_k = 3
if use_knn:
from sklearn.neighbors import NearestNeighbors
knn = NearestNeighbors(algorithm="auto", p=2)
knn.fit(features_galleries)
dists, ii_closest = knn.kneighbors(features_queries, n_neighbors=top_k, return_distance=True)
else:
dist_mat = pairwise_dist(x1=features_queries, x2=features_galleries)
dists, ii_closest = torch.topk(dist_mat, dim=1, k=top_k, largest=False)
print(f"Top {top_k} items closest to queries are:\n {ii_closest}")
MORE EXAMPLES
Illustrations, explanations and tips
Pipelines provide a way to run metric learning experiments via changing only the config file.
All you need is to prepare your dataset in a required format.
See Pipelines folder for more details:
Models, trained by us.
The metrics below are for 224 x 224 images:
model | cmc1 | dataset | weights | experiment |
---|
ViTExtractor.from_pretrained("vits16_inshop") | 0.921 | DeepFashion Inshop | link | link |
ViTExtractor.from_pretrained("vits16_sop") | 0.866 | Stanford Online Products | link | link |
ViTExtractor.from_pretrained("vits16_cars") | 0.907 | CARS 196 | link | link |
ViTExtractor.from_pretrained("vits16_cub") | 0.837 | CUB 200 2011 | link | link |
Models, trained by other researchers.
Note, that some metrics on particular benchmarks are so high because they were part of the training dataset (for example unicom
).
The metrics below are for 224 x 224 images:
model | Stanford Online Products | DeepFashion InShop | CUB 200 2011 | CARS 196 |
---|
ViTUnicomExtractor.from_pretrained("vitb16_unicom") | 0.700 | 0.734 | 0.847 | 0.916 |
ViTUnicomExtractor.from_pretrained("vitb32_unicom") | 0.690 | 0.722 | 0.796 | 0.893 |
ViTUnicomExtractor.from_pretrained("vitl14_unicom") | 0.726 | 0.790 | 0.868 | 0.922 |
ViTUnicomExtractor.from_pretrained("vitl14_336px_unicom") | 0.745 | 0.810 | 0.875 | 0.924 |
ViTCLIPExtractor.from_pretrained("sber_vitb32_224") | 0.547 | 0.514 | 0.448 | 0.618 |
ViTCLIPExtractor.from_pretrained("sber_vitb16_224") | 0.565 | 0.565 | 0.524 | 0.648 |
ViTCLIPExtractor.from_pretrained("sber_vitl14_224") | 0.512 | 0.555 | 0.606 | 0.707 |
ViTCLIPExtractor.from_pretrained("openai_vitb32_224") | 0.612 | 0.491 | 0.560 | 0.693 |
ViTCLIPExtractor.from_pretrained("openai_vitb16_224") | 0.648 | 0.606 | 0.665 | 0.767 |
ViTCLIPExtractor.from_pretrained("openai_vitl14_224") | 0.670 | 0.675 | 0.745 | 0.844 |
ViTExtractor.from_pretrained("vits16_dino") | 0.648 | 0.509 | 0.627 | 0.265 |
ViTExtractor.from_pretrained("vits8_dino") | 0.651 | 0.524 | 0.661 | 0.315 |
ViTExtractor.from_pretrained("vitb16_dino") | 0.658 | 0.514 | 0.541 | 0.288 |
ViTExtractor.from_pretrained("vitb8_dino") | 0.689 | 0.599 | 0.506 | 0.313 |
ViTExtractor.from_pretrained("vits14_dinov2") | 0.566 | 0.334 | 0.797 | 0.503 |
ViTExtractor.from_pretrained("vits14_reg_dinov2") | 0.566 | 0.332 | 0.795 | 0.740 |
ViTExtractor.from_pretrained("vitb14_dinov2") | 0.565 | 0.342 | 0.842 | 0.644 |
ViTExtractor.from_pretrained("vitb14_reg_dinov2") | 0.557 | 0.324 | 0.833 | 0.828 |
ViTExtractor.from_pretrained("vitl14_dinov2") | 0.576 | 0.352 | 0.844 | 0.692 |
ViTExtractor.from_pretrained("vitl14_reg_dinov2") | 0.571 | 0.340 | 0.840 | 0.871 |
ResnetExtractor.from_pretrained("resnet50_moco_v2") | 0.493 | 0.267 | 0.264 | 0.149 |
ResnetExtractor.from_pretrained("resnet50_imagenet1k_v1") | 0.515 | 0.284 | 0.455 | 0.247 |
*The metrics may be different from the ones reported by papers,
because the version of train/val split and usage of bounding boxes may differ.
How to use models from Zoo?
from oml.const import CKPT_SAVE_ROOT as CKPT_DIR, MOCK_DATASET_PATH as DATA_DIR
from oml.models import ViTExtractor
from oml.registry.transforms import get_transforms_for_pretrained
model = ViTExtractor.from_pretrained("vits16_dino")
transforms, im_reader = get_transforms_for_pretrained("vits16_dino")
img = im_reader(DATA_DIR / "images" / "circle_1.jpg")
img_tensor = transforms(img)
features = model(img_tensor.unsqueeze(0))
print(list(ViTExtractor.pretrained_models.keys()))
model_ = ViTExtractor(weights=CKPT_DIR / "vits16_dino.ckpt", arch="vits16", normalise_features=False)
We welcome new contributors! Please, see our:
Acknowledgments
The project was started in 2020 as a module for Catalyst library.
I want to thank people who worked with me on that module:
Julia Shenshina,
Nikita Balagansky,
Sergey Kolesnikov
and others.
I would like to thank people who continue working on this pipeline when it became a separe project:
Julia Shenshina,
Misha Kindulov,
Aron Dik,
Aleksei Tarasov and
Verkhovtsev Leonid.
I also want to thank NewYorker, since the part of functionality was developed (and used) by its computer vision team led by me.