Skip to content


Repository files navigation

GitHub license PRs Welcome

Classy Vision is no longer actively maintained.

The latest stable version is 0.7.0 and is available on pip, and has been tested to work with Pytorch 2.0.

What's New:

2020-11-20: Classy Vision v0.5 Released

New Features

  • Release Vision Transformers model implementation, with recipes(#646)
  • Implemented gradient clipping (#643)
  • Implemented gradient accumulation (#644)
  • Added support for AdamW (#636)
  • Added Precise batch norm hook (#592)
  • Added support for adaptive pooling in fully_convolutional_linear_head (#602)
  • Added support for sync batch norm group size (#534)
  • Added a CSV Hook to manually inspect model predictions
  • Added a ClassyModel tutorial (#485)
  • Migrated to Hydra 1.0 (#536)
  • Migrated off of tensorboardX (#488)

Breaking Changes

  • ClassyOptimizer API improvements
    • added OptionsView to retrieve options from the optimizer param_group
  • Removed ClassyModel.evaluation_mode (#521)
  • Removed ImageNetDataset, now a subset of ImagePathDataset (#494)
  • Renamed is_master to is_primary in distributed_util (#576)
2020-04-29: Classy Vision v0.4 Released

New Features

  • Release EfficientNet model implementation (#475)
  • Add support to convert any PyTorch model to a ClassyModel with the ability to attach heads to it (#461)
    • Added a corresponding tutorial on ClassyModel and ClassyHeads (#485)
  • Squeeze and Excitation support for ResNe(X)t and DenseNet models (#426, #427)
  • Made ClassyHooks registrable (#401) and configurable (#402)
  • Migrated to TorchElastic v0.2.0 (#464)
  • Add SyncBatchNorm support (#423)
  • Implement mixup train augmentation (#469)
  • Support LARC for SGD optimizer (#408)
  • Added convenience wrappers for Iterable datasets (#455)
  • Tensorboard improvements
    • Plot histograms of model weights to Tensorboard (#432)
    • Reduce data logged to tensorboard (#436)
  • Invalid (NaN / Inf) loss detection
  • Revamped logging (#478)
  • Add bn_weight_decay configuration option for ResNe(X)t models
  • Support specifying update_interval to Parameter Schedulers (#418)

Breaking changes

  • ClassificationTask API improvement and train_step, eval_step simplification
    • Removed local_variables from ClassificationTask (#411, #412, #413, #414, #416, #421)
    • Move use_gpu from ClassyTrainer to ClassificationTask (#468)
    • Move num_dataloader_workers out of ClassyTrainer (#477)
  • Rename lr to value in parameter schedulers (#417)
2020-03-06: Classy Vision v0.3 Released

Release notes

  • checkpoint_folder renamed to checkpoint_load_path (#379)
  • head support on DenseNet (#383)
  • Cleaner abstraction in ClassyTask/ClassyTrainer: eval_step, on_start, on_end, …
  • Speed metrics in TB (#385)
  • test_phase_period in ClassificationTask (#395)
  • support for losses with trainable parameters (#394)
  • Added presets for some typical ResNe(X)t configurations: #405)


Classy Vision is a new end-to-end, PyTorch-based framework for large-scale training of state-of-the-art image and video classification models. Previous computer vision (CV) libraries have been focused on providing components for users to build their own frameworks for their research. While this approach offers flexibility for researchers, in production settings it leads to duplicative efforts, and requires users to migrate research between frameworks and to relearn the minutiae of efficient distributed training and data loading. Our PyTorch-based CV framework offers a better solution for training at scale and for deploying to production. It offers several notable advantages:

  • Ease of use. The library features a modular, flexible design that allows anyone to train machine learning models on top of PyTorch using very simple abstractions. The system also has out-of-the-box integration with Amazon Web Services (AWS), facilitating research at scale and making it simple to move between research and production.
  • High performance. Researchers can use the framework to train Resnet50 on ImageNet in as little as 15 minutes, for example.
  • Demonstrated success in training at scale. We’ve used it to replicate the state-of-the-art results from the paper Exploring the Limits of Weakly Supervised Pretraining.
  • Integration with PyTorch Hub. AI researchers and engineers can download and fine-tune the best publically available ImageNet models with just a few lines of code.
  • Elastic training. We have also added experimental integration with PyTorch Elastic, which allows distributed training jobs to adjust as available resources in the cluster changes. It also makes distributed training robust to transient hardware failures.

Classy Vision is beta software. The project is under active development and our APIs are subject to change in future releases.


Installation Requirements

Make sure you have an up-to-date installation of PyTorch (1.6), Python (3.6) and torchvision (0.7). If you want to use GPUs, then a CUDA installation (10.1) is also required.

Installing the latest stable release

To install Classy Vision via pip:

pip install classy_vision

To install Classy Vision via conda (only works on linux):

conda install -c conda-forge classy_vision

Manual install of latest commit on main

Alternatively you can do a manual install.

git clone
cd ClassyVision
pip install .

Getting started

Classy Vision aims to support a variety of projects to be built and open sourced on top of the core library. We provide utilities for setting up a project in a standard format with some simple generated examples to get started with. To start a new project:

classy-project my-project
cd my-project

We even include a simple, synthetic, training example to show how to use Classy Vision:

 ./ --config configs/template_config.json

Voila! A few seconds later your first training run using our classification task should be done. Check out the results in the output folder:

ls output_<timestamp>/checkpoints/
checkpoint.torch model_phase-0_end.torch model_phase-1_end.torch model_phase-2_end.torch model_phase-3_end.torch

checkpoint.torch is the latest model (in this case, same as model_phase-3_end.torch), a checkpoint is saved at the end of each phase.

For more details / tutorials see the documentation section below.


Please see our tutorials to learn how to get started on Classy Vision and customize your training runs. Full documentation is available here.

Join the Classy Vision community

See the CONTRIBUTING file for how to help out.


Classy Vision is MIT licensed, as found in the LICENSE file.

Citing Classy Vision

If you use Classy Vision in your work, please use the following BibTeX entry:

  title={Classy Vision},
  author={{Adcock}, A. and {Reis}, V. and {Singh}, M. and {Yan}, Z. and {van der Maaten}, L. and {Zhang}, K. and {Motwani}, S. and {Guerin}, J. and {Goyal}, N. and {Misra}, I. and {Gustafson}, L. and {Changhan}, C. and {Goyal}, P.},
  howpublished = {\url{}},