Navigation Menu

Skip to content

google-deepmind/detcon

Repository files navigation

Code for DetCon

This repository contains code for the ICCV 2021 paper "Efficient Visual Pretraining with Contrastive Detection" by Olivier J. Hénaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron van den Oord, Oriol Vinyals, João Carreira.

This repository includes sample code to run pretraining with DetCon. In particular, we're providing a sample script for generating the Felzenzwalb segmentations for ImageNet images (using skimage) and a pre-training experiment setup (dataloader, augmentation pipeline, optimization config, and loss definition) that describes the DetCon-B(YOL) model described in the paper. The original code uses a large grid of TPUs and internal infrastructure for training, but we've extracted the key DetCon loss+experiment in this folder so that external groups can have a reference should they want to explore a similar approaches.

This repository builds heavily from the BYOL open source release, so speed-up tricks and features in that setup may likely translate to the code here.

Running this code

Running ./setup.sh will create and activate a virtualenv and install all necessary dependencies. To enter the environment after running setup.sh, run source /tmp/detcon_venv/bin/activate.

Running bash test.sh will run a single training step on a mock image/Felzenszwalb mask as a simple validation that all dependencies are set up correctly and the DetCon pre-training can run smoothly. On our 16-core machine, running on CPU, we find this takes around 3-4 minutes.

A TFRecord dataset containing each ImageNet image, label, and its corresponding Felzenszwalb segmentation/mask can then be generated using the generate_fh_masks Python script. You will first have to download two pieces of ImageNet metadata into the same directory as the script:

wget https://raw.githubusercontent.com/tensorflow/models/master/research/slim/datasets/imagenet_metadata.txt wget https://raw.githubusercontent.com/tensorflow/models/master/research/slim/datasets/imagenet_lsvrc_2015_synsets.txt

And to run the multi-threaded mask generation script:

python generate_fh_masks_for_imagenet.py -- \
--train_directory=imagenet-train \
--output_directory=imagenet-train-fh

This single-machine, multi-threaded version of the mask generation script takes 2-3 days on a 16-core CPU machine to complete CPU-based processing of the ImageNet training and validation set. The script assumes the same ImageNet directory structure as github.com/tensorflow/models/blob/master/research/slim/datasets/build_imagenet_data.py (more details in the link).

You can then run the main training loop and execute multiple DetCon-B training steps by running from the parent directory the command:

python -m detcon.main_loop \
  --dataset_directory='/tmp/imagenet-fh-train' \
  --pretrain_epochs=100`

Note that you will need to update the dataset_directory flag, to point to the generated Felzenzwalb/image TFRecord dataset previously generated. Additionally, to use accelerators, users will need to install the correct version of jaxlib with CUDA support.

Pre-trained checkpoints

For convenience, we're providing an ImageNet-pretrained ResNet-50 and ResNet-200 pre-trained using DetCon. We also provide a number of ResNet-50 COCO-pretrained checkpoints available in the same GCS bucket. A Colab demonstrating how to load the model weights and run a forward pass on the loaded model on a sample image is linked here.

Citing this work

If you use this code in your work, please consider referencing our work:

@article{henaff2021efficient,
  title={{Efficient Visual Pretraining with Contrastive Detection}},
  author={H{\'e}naff, Olivier J and Koppula, Skanda and Alayrac, Jean-Baptiste and Oord, Aaron van den and Vinyals, Oriol and Carreira, Jo{\~a}o},
  journal={International Conference on Computer Vision},
  year={2021}
}

Disclaimer

This is not an officially supported Google product.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published