CoreNet is a deep neural network toolkit that allows researchers and engineers to train standard and novel small and large-scale models for variety of tasks, including foundation models (e.g., CLIP and LLM), object classification, object detection, and semantic segmentation.
- What's new?
- Research efforts at Apple using CoreNet
- Installation
- Directory Structure
- Maintainers
- Contributing to CoreNet
- License
- Relationship with CVNets
- Citation
- April 2024: Version 0.1.0 of the CoreNet library includes
- OpenELM
- CatLIP
- MLX examples
Below is the list of publications from Apple that uses CoreNet. Also, training and evaluation recipes, as well as links to pre-trained models, can be found inside the projects folder. Please refer to it for further details.
- OpenELM: An Efficient Language Model Family with Open Training and Inference Framework
- CatLIP: CLIP-level Visual Recognition Accuracy with 2.7x Faster Pre-training on Web-scale Image-Text Data
- Reinforce Data, Multiply Impact: Improved Model Accuracy and Robustness with Dataset Reinforcement
- CLIP meets Model Zoo Experts: Pseudo-Supervision for Visual Enhancement
- FastVit: A Fast Hybrid Vision Transformer using Structural Reparameterization
- Bytes Are All You Need: Transformers Operating Directly on File Bytes
- MobileOne: An Improved One millisecond Mobile Backbone
- RangeAugment: Efficient Online Augmentation with Range Learning
- Separable Self-attention for Mobile Vision Transformers (MobileViTv2)
- CVNets: High performance library for Computer Vision, ACM MM'22
- MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer, ICLR'22
You will need Git LFS (instructions below) to run tests and Jupyter notebooks (instructions) in this repository, and to contribute to it so we recommend that you install and activate it first.
On Linux we recommend to use Python 3.10+ and PyTorch (version >= v2.1.0), on macOS system Python 3.9+ should be sufficient.
Note that the optional dependencies listed below are required if you'd like to make contributions and/or run tests.
For Linux (substitute apt
for your package manager):
sudo apt install git-lfs
git clone git@github.com:apple/corenet.git
cd corenet
git lfs install
git lfs pull
# The following venv command is optional, but recommended. Alternatively, you can create and activate a conda environment.
python3 -m venv venv && source venv/bin/activate
python3 -m pip install --editable .
To install optional dependencies for audio and video processing:
sudo apt install libsox-dev ffmpeg
For macOS, assuming you use Homebrew:
brew install git-lfs
git clone git@github.com:apple/corenet.git
cd corenet
cd \$(pwd -P) # See the note below.
git lfs install
git lfs pull
# The following venv command is optional, but recommended. Alternatively, you can create and activate a conda environment.
python3 -m venv venv && source venv/bin/activate
python3 -m pip install --editable .
To install optional dependencies for audio and video processing:
brew install sox ffmpeg
Note that on macOS the file system is case insensitive, and case sensitivity
can cause issues with Git. You should access the repository on disk as if the
path were case sensitive, i.e. with the same capitalization as you see when you
list the directories ls
. You can switch to such a path with the cd $(pwd -P)
command.
This section provides quick access and a brief description for important CoreNet directories.
Description | Quick Access |
---|---|
Working with the examples is an easy way to get started with CoreNet. | └── tutorials ├── train_a_new_model_on_a_new_dataset_from_scratch.ipynb ├── guide_slurm_and_multi_node_training.md ├── clip.ipynb ├── semantic_segmentation.ipynb └── object_detection.ipynb |
CoreNet provides reproducible training recipes, in addition to the pretrained model
weights and checkpoints for the publications that are listed in projects/ directory.
Publication project directories generally contain the following contents:
| └── projects ├── byteformer ├── catlip (*) ├── clip ├── fastvit ├── mobilenet_v1 ├── mobilenet_v2 ├── mobilenet_v3 ├── mobileone ├── mobilevit ├── mobilevit_v2 ├── openelm (*) ├── range_augment ├── resnet └── vit |
MLX examples demonstrate how to run CoreNet models efficiently on Apple Silicon.
Please find further information in the README.md file within the corresponding example directory.
| └──mlx_example ├── clip └── open_elm |
Models are organized by tasks (e.g. "classification"). You can find all model implementations for each
task in the corresponding task folder.
Each model class is decorated by a
| └── corenet └── modeling └── models ├── audio_classification ├── classification ├── detection ├── language_modeling ├── multi_modal_img_text └── segmentation |
Similarly to the models, datasets are also categorized by tasks. | └── corenet └── data └── datasets ├── audio_classification ├── classification ├── detection ├── language_modeling ├── multi_modal_img_text └── segmentation |
In this section, we have highlighted the rest of the key directories that implement classes corresponding to the names that are referenced in the YAML configurations. | └── corenet ├── loss_fn ├── metrics ├── optims │ └── scheduler ├── train_eval_pipelines ├── data │ ├── collate_fns │ ├── sampler │ ├── text_tokenizer │ ├── transforms │ └── video_reader └── modeling ├── layers ├── modules ├── neural_augmentor └── text_encoders |
This code is developed by Sachin, and is now maintained by Sachin, Maxwell Horton, Mohammad Sekhavat, and Yanzi Jin.
We welcome PRs from the community! You can find information about contributing to CoreNet in our contributing document.
Please remember to follow our Code of Conduct.
For license details, see LICENSE.
CoreNet evolved from CVNets, to encompass a broader range of applications beyond computer vision. Its expansion facilitated the training of foundational models, including LLMs.
If you find our work useful, please cite the following paper:
@inproceedings{mehta2022cvnets,
author = {Mehta, Sachin and Abdolhosseini, Farzad and Rastegari, Mohammad},
title = {CVNets: High Performance Library for Computer Vision},
year = {2022},
booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
series = {MM '22}
}