CLIP Convnext Base W 320 Laion Aesthetic S13b B82k

🚀 Model Card for CLIP-convnext_base_w-320.laion_aesthetic-s13B-b82k

This is a series of CLIP models based on ConvNeXt-Base architecture, trained on subsets of LAION-5B using OpenCLIP. It aims to explore an alternative to ViT and ResNet CLIP models, and may be more sample - efficient in a certain model scale range.

🚀 Quick Start

This README provides detailed information about the model, including its details, uses, training, evaluation, acknowledgements, and citation. You can quickly navigate to the content you're interested in through the table of contents.

✨ Features

• Alternative Architecture: Utilize the ConvNeXt - Base architecture as the image tower, offering an alternative to ViT and ResNet in CLIP models.
• Sample Efficiency: May be more sample - efficient compared to ViT - B/16 in a certain model scale range, with ImageNet Zero - Shot top - 1 of >= 70.8%.
• Augmentation and Regularization: Explore increased augmentation and regularization for the image tower, such as adding a greater scale range of RRC, random erasing, and stochastic depth.

📚 Documentation

Model Details

A series of CLIP ConvNeXt - Base (w/ wide embed dim) models were trained on subsets LAION - 5B (https://laion.ai/blog/laion - 5b/) using OpenCLIP (https://github.com/mlfoundations/open_clip).

Goals:

• Explore an alternative to ViT and ResNet (w/ AttentionPooling) CLIP models that scales well with model size and image resolution.

Firsts:

• First known ConvNeXt CLIP models trained at scale in the range of CLIP ViT - B/16 and RN50x4 models.
• First released model weights exploring increase of augmentation + regularization for image tower via adding (greater scale range of RRC, random erasing, stochastic depth).

The models use the [timm](https://github.com/rwightman/pytorch - image - models) ConvNeXt - Base model (convnext_base) as the image tower, and the same text tower as the RN50x4 (depth 12, embed dim 640) model from OpenAI CLIP. The base models are trained at 256x256 image resolution and roughly match the RN50x4 models on FLOPs and activation counts. The models with 320 in the name are trained at 320x320.

All models in this series were trained for 13B samples and have ImageNet Zero - Shot top - 1 of >= 70.8%. Comparing to ViT - B/16 at 34B SS with zero - shot of 70.2% (68.1% for 13B SS), this suggests the ConvNeXt architecture may be more sample efficient in this range of model scale. More experiments are needed to confirm.

RRC = Random Resize Crop (crop pcts), RE = Random Erasing (prob), SD = Stochastic Depth (prob) -- image tower only.

LAION - A = LAION Aesthetic, an ~900M sample subset of LAION - 2B with pHash dedupe and asthetic score filtering.

Model training was done by Ross Wightman across both the stability.ai cluster and the [JUWELS Booster](https://apps.fz - juelich.de/jsc/hps/juwels/booster - overview.html) supercomputer.

Uses

As per the original [OpenAI CLIP model card](https://github.com/openai/CLIP/blob/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1/model - card.md), this model is intended as a research output for research communities.

Direct Use: Zero - shot image classification, image and text retrieval, among others.

Downstream Use: Image classification and other image task fine - tuning, linear probe image classification, image generation guiding and conditioning, among others.

Out - of - Scope Use: As per the OpenAI models, any deployed use case of the model - whether commercial or not - is currently out of scope. Non - deployed use cases such as image search in a constrained environment are also not recommended unless there is thorough in - domain testing of the model with a specific, fixed class taxonomy. Certain use cases which would fall under the domain of surveillance and facial recognition are always out - of - scope regardless of the model's performance. Since the model has not been purposefully trained in or evaluated on any languages other than English, its use should be limited to English language use cases.

Training Details

Training Data

This model was trained with one of the following datasets:

• LAION - 2B: A 2 billion sample English subset of LAION - 5B (https://laion.ai/blog/laion - 5b/).
• LAION - Aesthetic: A 900M sample subset of LAION - 2B with pHash dedupe and asthetic score filtering.

IMPORTANT NOTE: The dataset is uncurated and may contain strongly discomforting and disturbing content. It is recommended for research purposes only. You can extract a “safe” subset by filtering out samples based on safety tags, but the possibility of harmful content still exists. It is not recommended to use it for creating ready - to - go industrial products.

Training Procedure

All models were trained with a global batch size of 81920 for 64 checkpoint intervals of 203.7M samples for a total of ~13B samples seen over training.

For 256x256 models, a slurm script w/ srun below was used on 20 8 - GPU (A100 40GB) nodes (Stability), switching to 40 4 - GPU nodes for time on JUWELS.

/opt/slurm/sbin/srun --cpu_bind=v --accel-bind=gn python -m training.main \
    --save-frequency 1 \
    --name "convnext_256" \
    --resume 'latest' \
    --train-data="pipe:aws s3 cp s3://mybucket/path/{laion{00000..xxxxx}.tar -" \
    --train-num-samples 203666042 \
    --dataset-type webdataset \
    --precision amp_bfloat16 \
    --warmup 10000 \
    --batch-size=512 \
    --epochs=64 \
    --dataset-resampled \
    --clip-grad-norm 5.0 \
    --lr 1e-3 \
    --workers=6 \
    --model "convnext_base_w" \
    --seed 0 \
    --ddp-static-graph \
    --local-loss \
    --gather-with-grad \
    --grad-checkpointing

For 320x320 models, the same settings were used but with 32 8 - GPU nodes, local batch size 320, or 64 4 - GPU nodes on JUWELs.

Evaluation

Evaluation was done with code in the [LAION CLIP Benchmark suite](https://github.com/LAION - AI/CLIP_benchmark).

Testing Data, Factors & Metrics

The testing is performed with VTAB+ (A combination of VTAB (https://arxiv.org/abs/1910.04867) w/ additional robustness datasets) for classification and COCO and Flickr for retrieval.

Results

The models achieve between 70.8 and 71.7 zero - shot top - 1 accuracy on ImageNet - 1k.

An initial round of benchmarks have been performed on a wider range of datasets, viewable at https://github.com/LAION - AI/CLIP_benchmark/blob/main/benchmark/results.ipynb

Early evaluations suggest that augreg trained models evaluate well over a wider range of resolutions.

Acknowledgements

We acknowledge stability.ai and the Gauss Centre for Supercomputing e.V. (http://gauss - centre.eu) for funding this part of work by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS Booster at Jülich Supercomputing Centre (JSC).

Citation

BibTeX:

LAION - 5B

@inproceedings{schuhmann2022laionb,
  title={{LAION}-5B: An open large-scale dataset for training next generation image-text models},
  author={Christoph Schuhmann and
          Romain Beaumont and
          Richard Vencu and
          Cade W Gordon and
          Ross Wightman and
          Mehdi Cherti and
          Theo Coombes and
          Aarush Katta and
          Clayton Mullis and
          Mitchell Wortsman and
          Patrick Schramowski and
          Srivatsa R Kundurthy and
          Katherine Crowson and
          Ludwig Schmidt and
          Robert Kaczmarczyk and
          Jenia Jitsev},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2022},
  url={https://openreview.net/forum?id=M3Y74vmsMcY}
}

OpenCLIP software

@software{ilharco_gabriel_2021_5143773,
  author       = {Ilharco, Gabriel and
                  Wortsman, Mitchell and
                  Wightman, Ross and
                  Gordon, Cade and
                  Carlini, Nicholas and
                  Taori, Rohan and
                  Dave, Achal and
                  Shankar, Vaishaal and
                  Namkoong, Hongseok and
                  Miller, John and
                  Hajishirzi, Hannaneh and
                  Farhadi, Ali and
                  Schmidt, Ludwig},
  title        = {OpenCLIP},
  month        = jul,
  year         = 2021,
  note         = {If you use this software, please cite it as below.},
  publisher    = {Zenodo},
  version      = {0.1},
  doi          = {10.5281/zenodo.5143773},
  url          = {https://doi.org/10.5281/zenodo.5143773}
}

OpenAI CLIP paper

@inproceedings{Radford2021LearningTV,
  title={Learning Transferable Visual Models From Natural Language Supervision},
  author={Alec Radford and Jong Wook Kim and Chris Hallacy and A. Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever},
  booktitle={ICML},
  year={2021}
}

@Article{liu2022convnet,
  author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
  title   = {A ConvNet for the 2020s},
  journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year    = {2022},
}

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
}

📄 License

This project is licensed under the MIT license.