Awesome
<p align="center"> <img src="logo.png" width="250"/> </p> <div align="center"> </div>solo-learn
A library of self-supervised methods for unsupervised visual representation learning powered by PyTorch Lightning. We aim at providing SOTA self-supervised methods in a comparable environment while, at the same time, implementing training tricks. The library is self-contained, but it is possible to use the models outside of solo-learn. More details in our paper.
News
- [Jan 14 2024]: :clap: Bunch of stability improvements during 2023 :) Also added All4One.
- [Jan 07 2023]: :diving_mask: Added results, checkpoints and configs for MAE on ImageNet. Thanks to HuangChiEn.
- [Dec 31 2022]: :stars: Shiny new logo! Huge thanks to Luiz!
- [Sep 27 2022]: :pencil: Brand new config system using OmegaConf/Hydra. Adds more clarity and flexibility. New tutorials will follow soon!
- [Aug 04 2022]: :paintbrush: Added MAE and supports finetuning the backbone with
main_linear.py
, mixup, cutmix and random augment. - [Jul 13 2022]: :sparkling_heart: Added support for H5 data, improved scripts and data handling.
- [Jun 26 2022]: :fire: Added MoCo V3.
- [Jun 10 2022]: :bomb: Improved LARS.
- [Jun 09 2022]: :lollipop: Added support for WideResnet, multicrop for SwAV and equalization data augmentation.
- [May 02 2022]: :diamond_shape_with_a_dot_inside: Wrapped Dali with a DataModule, added auto resume for linear eval and Wandb run resume.
- [Apr 12 2022]: :rainbow: Improved design of models and added support to train with a fraction of data.
- [Apr 01 2022]: :mag: Added the option to use channel last conversion which considerably decreases training times.
- [Feb 04 2022]: :partying_face: Paper got accepted to JMLR.
- [Jan 31 2022]: :eye: Added ConvNeXt support with timm.
- [Dec 20 2021]: :thermometer: Added ImageNet results, scripts and checkpoints for MoCo V2+.
- [Dec 05 2021]: :notes: Separated SupCon from SimCLR and added runs.
- [Dec 01 2021]: :fountain: Added PoolFormer.
- [Nov 29 2021]: :bangbang: Breaking changes! Update your versions!!!
- [Nov 29 2021]: :book: New tutorials!
- [Nov 29 2021]: :houses: Added offline K-NN and offline UMAP.
- [Nov 29 2021]: :rotating_light: Updated PyTorch and PyTorch Lightning versions. 10% faster.
- [Nov 29 2021]: :beers: Added code of conduct, contribution instructions, issue templates and UMAP tutorial.
- [Nov 23 2021]: :space_invader: Added VIbCReg.
- [Oct 21 2021]: :triumph: Added support for object recognition via Detectron v2 and auto resume functionally that automatically tries to resume an experiment that crashed/reached a timeout.
- [Oct 10 2021]: :japanese_ogre: Restructured augmentation pipelines to allow more flexibility and multicrop. Also added multicrop for BYOL.
- [Sep 27 2021]: :pizza: Added NNSiam, NNBYOL, new tutorials for implementing new methods 1 and 2, more testing and fixed issues with custom data and linear evaluation.
- [Sep 19 2021]: :kangaroo: Added online k-NN evaluation.
- [Sep 17 2021]: :robot: Added ViT and Swin.
- [Sep 13 2021]: :book: Improved Docs and added tutorials for pretraining and offline linear eval.
- [Aug 13 2021]: :whale: DeepCluster V2 is now available.
Roadmap and help needed
- Redoing the documentation to improve clarity.
- Better and up-to-date tutorials.
- Add performance-related testing to ensure that methods perform the same across updates.
- Adding new methods (continuous effort).
Methods available
- All4One
- Barlow Twins
- BYOL
- DeepCluster V2
- DINO
- MAE
- MoCo V2+
- MoCo V3
- NNBYOL
- NNCLR
- NNSiam
- ReSSL
- SimCLR
- SimSiam
- Supervised Contrastive Learning
- SwAV
- VIbCReg
- VICReg
- W-MSE
Extra flavor
Backbones
Data
- Increased data processing speed by up to 100% using Nvidia Dali.
- Flexible augmentations.
Evaluation
- Online linear evaluation via stop-gradient for easier debugging and prototyping (optionally available for the momentum backbone as well).
- Standard offline linear evaluation.
- Online and offline K-NN evaluation.
- Automatic feature space visualization with UMAP.
Training tricks
- All the perks of PyTorch Lightning (mixed precision, gradient accumulation, clipping, and much more).
- Channel last conversion
- Multi-cropping dataloading following SwAV:
- Note: currently, only SimCLR, BYOL and SwAV support this.
- Exclude batchnorm and biases from weight decay and LARS.
- No LR scheduler for the projection head (as in SimSiam).
Logging
- Metric logging on the cloud with WandB
- Custom model checkpointing with a simple file organization.
Requirements
- torch
- torchvision
- tqdm
- einops
- wandb
- pytorch-lightning
- lightning-bolts
- torchmetrics
- scipy
- timm
Optional:
- nvidia-dali
- matplotlib
- seaborn
- pandas
- umap-learn
Installation
First clone the repo.
Then, to install solo-learn with Dali and/or UMAP support, use:
pip3 install .[dali,umap,h5] --extra-index-url https://developer.download.nvidia.com/compute/redist
If no Dali/UMAP/H5 support is needed, the repository can be installed as:
pip3 install .
For local development:
pip3 install -e .[umap,h5]
# Make sure you have pre-commit hooks installed
pre-commit install
NOTE: if you are having trouble with dali, install it following their guide.
NOTE 2: consider installing Pillow-SIMD for better loading times when not using Dali.
NOTE 3: Soon to be on pip.
Training
For pretraining the backbone, follow one of the many bash files in scripts/pretrain/
.
We are now using Hydra to handle the config files, so the common syntax is something like:
python3 main_pretrain.py \
# path to training script folder
--config-path scripts/pretrain/imagenet-100/ \
# training config name
--config-name barlow.yaml
# add new arguments (e.g. those not defined in the yaml files)
# by doing ++new_argument=VALUE
# pytorch lightning's arguments can be added here as well.
After that, for offline linear evaluation, follow the examples in scripts/linear
or scripts/finetune
for finetuning the whole backbone.
For k-NN evaluation and UMAP visualization check the scripts in scripts/{knn,umap}
.
NOTE: Files try to be up-to-date and follow as closely as possible the recommended parameters of each paper, but check them before running.
Tutorials
Please, check out our documentation and tutorials:
- Overview
- Offline linear eval
- Object detection
- Adding a new method
- Adding a new momentum method
- Visualizing features with UMAP
- Offline k-NN
If you want to contribute to solo-learn, make sure you take a look at how to contribute and follow the code of conduct
Model Zoo
All pretrained models avaiable can be downloaded directly via the tables below or programmatically by running one of the following scripts
zoo/cifar10.sh
, zoo/cifar100.sh
, zoo/imagenet100.sh
and zoo/imagenet.sh
.
Results
Note: hyperparameters may not be the best, we will be re-running the methods with lower performance eventually.
CIFAR-10
Method | Backbone | Epochs | Dali | Acc@1 | Acc@5 | Checkpoint |
---|---|---|---|---|---|---|
All4One | ResNet18 | 1000 | :x: | 93.24 | 99.88 | :link: |
Barlow Twins | ResNet18 | 1000 | :x: | 92.10 | 99.73 | :link: |
BYOL | ResNet18 | 1000 | :x: | 92.58 | 99.79 | :link: |
DeepCluster V2 | ResNet18 | 1000 | :x: | 88.85 | 99.58 | :link: |
DINO | ResNet18 | 1000 | :x: | 89.52 | 99.71 | :link: |
MoCo V2+ | ResNet18 | 1000 | :x: | 92.94 | 99.79 | :link: |
MoCo V3 | ResNet18 | 1000 | :x: | 93.10 | 99.80 | :link: |
NNCLR | ResNet18 | 1000 | :x: | 91.88 | 99.78 | :link: |
ReSSL | ResNet18 | 1000 | :x: | 90.63 | 99.62 | :link: |
SimCLR | ResNet18 | 1000 | :x: | 90.74 | 99.75 | :link: |
Simsiam | ResNet18 | 1000 | :x: | 90.51 | 99.72 | :link: |
SupCon | ResNet18 | 1000 | :x: | 93.82 | 99.65 | :link: |
SwAV | ResNet18 | 1000 | :x: | 89.17 | 99.68 | :link: |
VIbCReg | ResNet18 | 1000 | :x: | 91.18 | 99.74 | :link: |
VICReg | ResNet18 | 1000 | :x: | 92.07 | 99.74 | :link: |
W-MSE | ResNet18 | 1000 | :x: | 88.67 | 99.68 | :link: |
CIFAR-100
Method | Backbone | Epochs | Dali | Acc@1 | Acc@5 | Checkpoint |
---|---|---|---|---|---|---|
All4One | ResNet18 | 1000 | :x: | 72.17 | 93.35 | :link: |
Barlow Twins | ResNet18 | 1000 | :x: | 70.90 | 91.91 | :link: |
BYOL | ResNet18 | 1000 | :x: | 70.46 | 91.96 | :link: |
DeepCluster V2 | ResNet18 | 1000 | :x: | 63.61 | 88.09 | :link: |
DINO | ResNet18 | 1000 | :x: | 66.76 | 90.34 | :link: |
MoCo V2+ | ResNet18 | 1000 | :x: | 69.89 | 91.65 | :link: |
MoCo V3 | ResNet18 | 1000 | :x: | 68.83 | 90.57 | :link: |
NNCLR | ResNet18 | 1000 | :x: | 69.62 | 91.52 | :link: |
ReSSL | ResNet18 | 1000 | :x: | 65.92 | 89.73 | :link: |
SimCLR | ResNet18 | 1000 | :x: | 65.78 | 89.04 | :link: |
Simsiam | ResNet18 | 1000 | :x: | 66.04 | 89.62 | :link: |
SupCon | ResNet18 | 1000 | :x: | 70.38 | 89.57 | :link: |
SwAV | ResNet18 | 1000 | :x: | 64.88 | 88.78 | :link: |
VIbCReg | ResNet18 | 1000 | :x: | 67.37 | 90.07 | :link: |
VICReg | ResNet18 | 1000 | :x: | 68.54 | 90.83 | :link: |
W-MSE | ResNet18 | 1000 | :x: | 61.33 | 87.26 | :link: |
ImageNet-100
Method | Backbone | Epochs | Dali | Acc@1 (online) | Acc@1 (offline) | Acc@5 (online) | Acc@5 (offline) | Checkpoint |
---|---|---|---|---|---|---|---|---|
All4One | ResNet18 | 400 | :heavy_check_mark: | 81.93 | - | 96.23 | - | :link: |
Barlow Twins :rocket: | ResNet18 | 400 | :heavy_check_mark: | 80.38 | 80.16 | 95.28 | 95.14 | :link: |
BYOL :rocket: | ResNet18 | 400 | :heavy_check_mark: | 80.16 | 80.32 | 95.02 | 94.94 | :link: |
DeepCluster V2 | ResNet18 | 400 | :x: | 75.36 | 75.4 | 93.22 | 93.10 | :link: |
DINO | ResNet18 | 400 | :heavy_check_mark: | 74.84 | 74.92 | 92.92 | 92.78 | :link: |
DINO :sleepy: | ViT Tiny | 400 | :x: | 63.04 | TODO | 87.72 | TODO | :link: |
MoCo V2+ :rocket: | ResNet18 | 400 | :heavy_check_mark: | 78.20 | 79.28 | 95.50 | 95.18 | :link: |
MoCo V3 :rocket: | ResNet18 | 400 | :heavy_check_mark: | 80.36 | 80.36 | 95.18 | 94.96 | :link: |
MoCo V3 :rocket: | ResNet50 | 400 | :heavy_check_mark: | 85.48 | 84.58 | 96.82 | 96.70 | :link: |
NNCLR :rocket: | ResNet18 | 400 | :heavy_check_mark: | 79.80 | 80.16 | 95.28 | 95.30 | :link: |
ReSSL | ResNet18 | 400 | :heavy_check_mark: | 76.92 | 78.48 | 94.20 | 94.24 | :link: |
SimCLR :rocket: | ResNet18 | 400 | :heavy_check_mark: | 77.64 | TODO | 94.06 | TODO | :link: |
Simsiam | ResNet18 | 400 | :heavy_check_mark: | 74.54 | 78.72 | 93.16 | 94.78 | :link: |
SupCon | ResNet18 | 400 | :heavy_check_mark: | 84.40 | TODO | 95.72 | TODO | :link: |
SwAV | ResNet18 | 400 | :heavy_check_mark: | 74.04 | 74.28 | 92.70 | 92.84 | :link: |
VIbCReg | ResNet18 | 400 | :heavy_check_mark: | 79.86 | 79.38 | 94.98 | 94.60 | :link: |
VICReg :rocket: | ResNet18 | 400 | :heavy_check_mark: | 79.22 | 79.40 | 95.06 | 95.02 | :link: |
W-MSE | ResNet18 | 400 | :heavy_check_mark: | 67.60 | 69.06 | 90.94 | 91.22 | :link: |
:rocket: methods where hyperparameters were heavily tuned.
:sleepy: ViT is very compute intensive and unstable, so we are slowly running larger architectures and with a larger batch size. Atm, total batch size is 128 and we needed to use float32 precision. If you want to contribute by running it, let us know!
ImageNet
Method | Backbone | Epochs | Dali | Acc@1 (online) | Acc@1 (offline) | Acc@5 (online) | Acc@5 (offline) | Checkpoint | Finetuned Checkpoint |
---|---|---|---|---|---|---|---|---|---|
Barlow Twins | ResNet50 | 100 | :heavy_check_mark: | 67.18 | 67.23 | 87.69 | 87.98 | :link: | |
BYOL | ResNet50 | 100 | :heavy_check_mark: | 68.63 | 68.37 | 88.80 | 88.66 | :link: | |
MoCo V2+ | ResNet50 | 100 | :heavy_check_mark: | 62.61 | 66.84 | 85.40 | 87.60 | :link: | |
MAE | ViT-B/16 | 100 | :x: | ~ | 81.60 (finetuned) | ~ | 95.50 (finetuned) | :link: | :link: |
Training efficiency for DALI
We report the training efficiency of some methods using a ResNet18 with and without DALI (4 workers per GPU) in a server with an Intel i9-9820X and two RTX2080ti.
Method | Dali | Total time for 20 epochs | Time for 1 epoch | GPU memory (per GPU) |
---|---|---|---|---|
Barlow Twins | :x: | 1h 38m 27s | 4m 55s | 5097 MB |
:heavy_check_mark: | 43m 2s | 2m 10s (56% faster) | 9292 MB | |
BYOL | :x: | 1h 38m 46s | 4m 56s | 5409 MB |
:heavy_check_mark: | 50m 33s | 2m 31s (49% faster) | 9521 MB | |
NNCLR | :x: | 1h 38m 30s | 4m 55s | 5060 MB |
:heavy_check_mark: | 42m 3s | 2m 6s (64% faster) | 9244 MB |
Note: GPU memory increase doesn't scale with the model, rather it scales with the number of workers.
Citation
If you use solo-learn, please cite our paper:
@article{JMLR:v23:21-1155,
author = {Victor Guilherme Turrisi da Costa and Enrico Fini and Moin Nabi and Nicu Sebe and Elisa Ricci},
title = {solo-learn: A Library of Self-supervised Methods for Visual Representation Learning},
journal = {Journal of Machine Learning Research},
year = {2022},
volume = {23},
number = {56},
pages = {1-6},
url = {http://jmlr.org/papers/v23/21-1155.html}
}