CLIP Convnext Xxlarge Laion2b S34b B82k Augreg Rewind

🚀 CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-rewind

A series of CLIP ConvNeXt-XXLarge models trained on LAION-2B for zero-shot image classification.

🚀 Quick Start

This README provides detailed information about the CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-rewind model, including its details, uses, training, evaluation, and more.

✨ Features

• High Performance: Achieves over 79% ImageNet top-1 zero-shot accuracy.
• Large-scale Training: Trained on LAION-2B, a large English subset of LAION-5B.
• Advanced Architecture: Utilizes the timm ConvNeXt-XXLarge model as the image tower.

📦 Installation

No specific installation steps are provided in the original README.

💻 Usage Examples

No code examples are provided in the original README.

📚 Documentation

🔍 Model Details

Model Description

A series of CLIP ConvNeXt-XXLarge (a custom timm ConvNeXt size) models trained on LAION-2B (english), a subset of LAION-5B, using OpenCLIP.

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

The core training run was performed in pieces over a period of ~ 2 months. The global batch size for the core run was 81920. The last ~10% of training was re-done at a 95744 global batch size w/ higher LR and aug than original finish. The two were averaged together in a 'soup'. See more details in Training Details.

Goals:

• Push the size of largest convolutional CLIP image tower into the performance range of ViT-g to ViT-G w/ improved image size scaling for downstream use.

Firsts:

• Largest released ConvNeXt model pretrained (847M params w/ 198 GMAC and 125 MActs @ 256x256 for image)
• A non-ViT image tower CLIP model (with no previous image tower pretrain) achieving > 79% ImageNet top-1 zero-shot

The models utilize:

• the timm ConvNeXt-XXLarge model (convnext_xxlarge) as the image tower
• a standard projection at end of image tower
• a text tower with same size (with 1024, heads 16, depth 24) as ViT-H-14 and ViT-g-14 models

The models are trained at 256x256 image resolution. The size of the combined image + text CLIP model is 1.2B params w/ 222 GMAC and 146 MActs. At 256x256, the ConvNext-XXLarge sits just above a ViT-H-14 CLIP configuration in FLOPS and params while being lower in activation counts. It is well under both g-14 and G-14 while being between them in capabilities.

Model training done by Ross Wightman across both the stability.ai cluster and the JUWELS Booster supercomputer. See acknowledgements below.

🔎 Uses

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. This is because our safety assessment demonstrated a high need for task specific testing especially given the variability of CLIP’s performance with different class taxonomies. This makes untested and unconstrained deployment of the model in any use case currently potentially harmful.

Certain use cases which would fall under the domain of surveillance and facial recognition are always out-of-scope regardless of performance of the model. This is because the use of artificial intelligence for tasks such as these can be premature currently given the lack of testing norms and checks to ensure its fair use.

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.

Further the above notice, the LAION-5B dataset used in training of these models has additional considerations, see below.

📋 Training Details

Training Data

This model was trained with LAION-2B -- A 2 billion sample English subset of LAION-5B (https://laion.ai/blog/laion-5b/).

⚠️ Important Note

The motivation behind dataset creation is to democratize research and experimentation around large-scale multi-modal model training and handling of uncurated, large-scale datasets crawled from publically available internet. Our recommendation is therefore to use the dataset for research purposes. Be aware that this large-scale dataset is uncurated. Keep in mind that the uncurated nature of the dataset means that collected links may lead to strongly discomforting and disturbing content for a human viewer. Therefore, please use the demo links with caution and at your own risk. It is possible to extract a “safe” subset by filtering out samples based on the safety tags (using a customized trained NSFW classifier that we built). While this strongly reduces the chance for encountering potentially harmful content when viewing, we cannot entirely exclude the possibility for harmful content being still present in safe mode, so that the warning holds also there. We think that providing the dataset openly to broad research and other interested communities will allow for transparent investigation of benefits that come along with training large-scale models as well as pitfalls and dangers that may stay unreported or unnoticed when working with closed large datasets that remain restricted to a small community. Providing our dataset openly, we however do not recommend using it for creating ready-to-go industrial products, as the basic research about general properties and safety of such large-scale models, which we would like to encourage with this release, is still in progress.

Training Procedure

The main training run was done at global batch size of 81920 for 256 checkpoint intervals of 135.6M samples for a total of ~34B samples seen over training.

Many difficulties w/ both model numerical stability and cluster stability and performance were encountered while training this model. Initial attempts to train with float16 AMP and default adam beta2 resulted in loss spikes and eventually NaN blow ups. beta2 was reduced to 0.97 which helped, but the loss / zs curves were not tracking as expected. After switching to PyTorch nightlies, it was possible to use bfloat16 + AMP for training (as with rececnt H/14, g/14, and G/14 models), beta2 was returned to 0.98 and metrics improved.

JUWELS Booster has 4x A100 GPU per node w/ 4x HDR-200 IB adapters per node (200Gbit/sec per GPU). Stability setup used was 8x A100 GPU per node w/ 400Gbit/sec EFA networking per node (50 GBit/sec per GPU). Significant variation in training efficiency (throughput per GPU) as observed across the various configurations. The 1024 GPU configurations across both clusters were particularly prone to crashing (or very difficult to get running w/ a 'good' set of GPUs).

A slurm srun command line below for a 128 8-GPU (40GB A100) configuration:

srun --cpu_bind=v --accel-bind=gn python -m training.main \
    --save-frequency 1 \
    --name "xxlarge-2b-81920-bf16" \
    --resume "latest" \
    --logs "/runs" \
    --log-every-n-steps 50 \
    --train-data="pipe:aws s3 cp s3://laion5b/laion2B-data/{000000..231349}.tar -" \
    --train-num-samples 135646078 \
    --dataset-type webdataset \
    --warmup 10000 \
    --batch-size=80 \
    --epochs=256 \
    --dataset-resampled \
    --aug-cfg use_timm=True scale='(0.33, 1.0)' re_prob=0.35 \
    --precision amp_bfloat16 \
    --grad-clip-norm 5.0 \
    --lr 1e-3 \
    --workers=6 \
    --beta2 0.98 \
    --model "convnext_xxlarge" \
    --seed 0 \
    --ddp-static-graph \
    --local-loss \
    --gather-with-grad \
    --grad-checkpointing \
    --report-to "tensorboard"

For the rewind of last 10%, a higher global batch size of 95744 was used w/ a higher LR and slightly increased augmentation strength.

The slurm srun command line for 136 8-GPU (40GB A100) nodes:

srun --cpu_bind=v --accel-bind=gn python -m training.main \
    --save-frequency 1 \
    --name "xxlarge-2b-81920-r-bf16" \
    --resume "latest" \
    --logs "/runs" \
    --log-every-n-steps 50 \
    --train-data="pipe:aws s3 cp s3://laion5b/laion2B-data/{000000..231349}.tar -" \
    --train-num-samples 135646078 \
    --dataset-type webdataset \
    --warmup 10000 \
    --batch-size=88 \
    --epochs=256 \
    --dataset-resampled \
    --aug-cfg use_timm=True scale='(0.3, 1.0)' re_prob=0.4 \
    --precision amp_bfloat16 \
    --grad-clip-norm 5.0 \
    --lr 2e-3 \
    --workers=6 \
    --beta2 0.98 \
    --model "convnext_xxlarge"

📊 Evaluation

No evaluation details are provided in the original README.

🙏 Acknowledgements

Model training done by Ross Wightman across both the stability.ai cluster and the JUWELS Booster supercomputer.

📖 Citation

No citation details are provided in the original README.

🔧 Technical Details

The technical details are included in the Training Details section, such as the training data, training procedure, and the challenges encountered during training.

📄 License

The model is licensed under the MIT license.