clip-benchmark

CLIP Benchmark

The goal of this repo is to evaluate CLIP-like models on a standard set of datasets on different tasks such as zero-shot classification and zero-shot retrieval, and captioning.

Below we show the average rank (1 is the best, lower is better) of different CLIP models, evaluated on different datasets.

The current detailed results of the benchmark can be seen here or directly in the notebook.

Features

• Support for zero-shot classification and zero-shot retrieval, linear probing, and captioning.
• Support for OpenCLIP pre-trained models, Japanese CLIP, and NLLB CLIP for general multilingual abilities.
• Support various datasets from torchvision, tensorflow datasets, and VTAB.
• Support for various multilingual datasets for classification and retrieval
• Support for compositionality tasks

How to install?

pip install clip-benchmark

How to use?

To evaluate we recommend to create a models.txt like

ViT-B-32,openai

to get the list of datasets

wget https://raw.githubusercontent.com/LAION-AI/CLIP_benchmark/main/benchmark/webdatasets.txt

Then to run

clip_benchmark eval --pretrained_model models.txt \
    --dataset "webdatasets.txt" \
    --dataset_root "https://huggingface.co/datasets/clip-benchmark/wds_{dataset_cleaned}/tree/main" \
    --output "benchmark_{dataset}_{pretrained}_{model}_{language}_{task}.json"

Then to get the full table

clip_benchmark build benchmark_*.json --output benchmark.csv

Command line interface (CLI)

The easiest way to benchmark the models is using the CLI, clip_benchmark. You can specify the model to use, the dataset and the task to evaluate on. Once it is done, evaluation is performed and the results are written into a JSON file.

Using other models than openclip

It is possible to use other models than openclip ones. For example japanese-clip is supported

Here is an example of use

>>> python3 clip_benchmark/cli.py eval \
  --model_type "ja_clip" \ # flag to use japanese-clip
  --pretrained "rinna/japanese-cloob-vit-b-16" \ # now, we have `rinna/japanese-cloob-vit-b-16` or `rinna/japanese-clip-vit-b-16`. 
  --language "jp" \
  --task "zeroshot_classification"  \
  --dataset "imagenet1k"  \
  --dataset_root {ROOT_PATH} 

>>> cat result.json
{"dataset": "imagenet1k", "model": "ViT-B-32-quickgelu", "pretrained": "rinna/japanese-cloob-vit-b-16", "task": "zeroshot_classification", "metrics": {"acc1": 0.54636, "acc5": 0.72856, "mean_per_class_recall": 0.54522}, "language": "jp"}

How to add other CLIP models

Please follow these steps:

1. Add a identity file to load model in clip_benchmark/models
2. Define a loading function, that returns a tuple (model, transform, tokenizer). Please see clip_benchmark/models/open_clip.py as an example.
3. Add the function into TYPE2FUNC in clip_benchmark/models/__init__.py

Remarks:

• The new tokenizer/model must enable to do the following things as https://github.com/openai/CLIP#usage
  • tokenizer(texts).to(device) ... texts is a list of string
  • model.encode_text(tokenized_texts) ... tokenized_texts is a output from tokenizer(texts).to(device)
  • model.encode_image(images) ... images is a image tensor by the transform

CIFAR-10 example

Here is an example for CIFAR-10 zero-shot classification using OpenCLIP's pre-trained model on LAION-400m:

clip_benchmark eval --dataset=cifar10 --task=zeroshot_classification --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64

By default, the dataset is downloaded into --dataset_root, which by default is root.

Here is the content of result.json after the evaluation is done:

{
    "dataset": "cifar10", "model": "ViT-B-32-quickgelu", 
    "pretrained": "laion400m_e32", "task": "zeroshot_classification",
    "metrics": {"acc1": 0.9074, "acc5": 0.998}
}

VOC2007 example

Here is another example with VOC2007, which is a multi-label classification dataset.

clip_benchmark eval --dataset=voc2007_multilabel --task=zeroshot_classification --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64

Here is the content of result.json after the evaluation is done:

{"dataset": "voc2007_multilabel", "model": "ViT-B-32-quickgelu", "pretrained": "laion400m_e32", "task": "zeroshot_classification", "metrics": {"mean_average_precision": 0.7627869844436646}}

Here, we compute the mean average precision or mAP, more details about that metric here in the context of multi-label classification.

VTAB example

Here is an example on how to run it on VTAB classification tasks. First, you need to install VTAB's dedicated package.

pip install task_adaptation==0.1

Then, you can run it by providing the full dataset name. Example with eurosat:

clip_benchmark eval --dataset=vtab/eurosat --task=zeroshot_classification --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64

See clip_benchmark/datasets/builder.py#L634 for the full list of VTAB dataset collection.

TensorFlow dataset example

Here is an example on how to run it on Tensorflow datasets. First, you need to install tfds-nightly and timm.

pip install timm tfds-nightly

The name of the dataset follows the template tfds/<DATASET_NAME>.

Example with cifar10:

clip_benchmark eval --dataset=tfds/cifar10 --task=zeroshot_classification --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64

COCO captions retrieval example

Here is an example for COCO captions zero-shot retrieval:

clip_benchmark eval --dataset=mscoco_captions --task=zeroshot_retrieval --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64

Note that for using COCO, you also need to install pycocotools (e.g., using pip install pycocotools).

COCO captions captioning example

Here is an example for COCO captions captioning task:

clip_benchmark eval --dataset=mscoco_captions --task=captioning --model=coca_ViT-L-14 --output=result.json --pretrained mscoco_finetuned_laion2b_s13b_b90k

Note that for using COCO, you also need to install pycocotools (e.g., using pip install pycocotools) and pycocoevalcap.

Linear probing example

Full linear probing on train split, evaluate on test split:

clip_benchmark eval --dataset=cifar10 --task=linear_probe --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64 --fewshot_lr 0.1 --fewshot_epochs 20 --batch_size 512 --train_split train --test_split test

few-shot (k=5) linear probing on train split, evaluate on test split:

clip_benchmark eval --dataset=cifar10 --task=linear_probe --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64 --fewshot_k 5 --fewshot_lr 0.1 --fewshot_epochs 20 --batch_size 512 --train_split train --test_split test

Split train into train (90%), val (10%), do linear probing on train split, tune on val split, evaluate on test split:

clip_benchmark eval --dataset=cifar10 --task=linear_probe --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64 --fewshot_lr 0.1 --fewshot_epochs 20 --batch_size 512 --train_split train --val_proportion 0.1 --test_split test

For other datasets that have an official val split, one can also specify the val split:

clip_benchmark eval --dataset=fgvc_aircraft --task=linear_probe --pretrained=laion400m_e32 --model=ViT-B-32-quickgelu --output=result.json --batch_size=64 --fewshot_lr 0.1 --fewshot_epochs 20 --batch_size 512 --train_split train --val_split val --test_split test

Multilingual evaluation

We also provide datasets for evaluating multilingual models (see e.g. https://github.com/mlfoundations/open_clip#vit-b32-xlm-roberta-base, and https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_multilingual_retrieval_results.csv) by specifying --language.

For ImageNet-1k (zero-shot classification):

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=imagenet1k --output=result.json --batch_size=64 --language=<LANG>, where <LANG> can be among zh (chinese), it (italian), jp (japanese), en (english), ar (arabic).
• We also support Babel ImageNet classnames and prompts (https://github.com/gregor-ge/Babel-ImageNet), which can be used as the following: clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=babel_imagenet --output=result.json --batch_size=64 --language=<LANG>, where <LANG> is a two letter string from the ISO language code list. Supported values for language are: 'ne', 'id', 'de', 'nl', 'af', 'he', 'sq', 'uz', 'kn', 'ku', 'ta', 'lv', 'ko', 'ug', 'br', 'el', 'su', 'kk', 'sk', 'gl', 'om', 'fa', 'jv', 'cs', 'lo', 'hy', 'xh', 'hr', 'so', 'gu', 'am', 'ar', 'sa', 'ca', 'is', 'it', 'sv', 'ga', 'bg', 'vi', 'sd', 'ur', 'km', 'pl', 'hu', 'sr', 'fr', 'hi', 'fy', 'et', 'bs', 'sw', 'az', 'mk', 'es', 'mn', 'ja', 'tl', 'tr', 'gd', 'ro', 'mg', 'mr', 'sl', 'pt', 'lt', 'no', 'yi', 'uk', 'ky', 'ka', 'bn', 'or', 'my', 'en', 'ps', 'fi', 'zh', 'da', 'ml', 'be', 'eo', 'ha', 'eu', 'as', 'te', 'th', 'cy', 'si', 'ru', 'la', 'pa', 'ms'

for COCO (zero-shot retrieval):

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=multilingual_mscoco_captions --output=result.json --batch_size=64 --language=<LANG>, where <LANG> can be among es (spanish), it (italian), jp (japanese), ko (korean), pl (polish), ru (russian), tr (Turkish), zh (chinese), en (english), fr (french), de (german).

For Flickr-30k (zero-shot retrieval)

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=flickr30k --output=result.json --batch_size=64 --language=<LANG>, where <LANG> can be among en (english), zh (chinese).

For Flickr-8k (zero-shot retrieval)

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=flickr8k --output=result.json --batch_size=64 --language=<LANG>, where <LANG> can be among en (english), zh (chinese).

For Crossmodal-3600 (zero-shot retrieval)

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=crossmodal3600 --output=result.json --batch_size=64 --language=<LANG>, see supported languages here.

For Flickr30k-200 dataset, which has 1000 captions from Flickr30k dataset translated to 200 languages using NLLB-3.3B model (zero-shot retrieval)

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=flickr30k-200 --output=result.json --batch_size=64 --language=<LANG>, see supported languages here.

For XTD200 dataset, which has captions from XTD10 dataset translated to 200 languages using NLLB-3.3B model (zero-shot retrieval)

• clip_benchmark eval --model xlm-roberta-base-ViT-B-32 --pretrained laion5b_s13b_b90k --dataset=xtd200 --output=result.json --batch_size=64 --language=<LANG>, see supported languages here.

Compositionality evaluation

For Sugar Crepe:

clip_benchmark eval --model ViT-B-32 --pretrained laion400m_e32 --dataset=sugar_crepe/<TASK> --output=result.json

where <TASK> can be among add_att, add_obj, replace_att, replace_obj, replace_rel, swap_att, swap_obj. To evaluate on all the tasks together, you can do:

clip_benchmark eval --model ViT-B-32 --pretrained laion400m_e32 --dataset=sugar_crepe --output=result.json

For winoground:

clip_benchmark eval --model ViT-B-32 --pretrained laion400m_e32 --dataset=winoground --output=result.json

NB: pip install datasets is required for winoground.

Webdataset example

Here is an example on how to run it on webdatasets. First, you need to install webdataset.

pip install webdataset

Creating a webdataset

You can either convert an already supported CLIP_benchmark dataset to webdataset format, or manually create your own with the same file structure. For already supported datasets use the CLI command clip_benchmark_export_wds as in this example:

$ clip_benchmark_export_wds --dataset cifar10 --plit train --dataset_root DATA_DIR/ --output wds_cifar10/
$ clip_benchmark_export_wds --dataset cifar10 --split test --dataset_root DATA_DIR/ --output wds_cifar10/

which will convert the train and test splits for CIFAR-10 (downloaded to DATA_DIR/) and save the webdataset to wds_cifar10/ (upload to Huggingface Hub must be done manually for now). Retrieval datasets are also supported with the --retrieval flag.

For other datasets, data must be stored with the following file structure:

root_dir/
    train/
        nshards.txt
        0.tar
        1.tar
        ...
    test/
        nshards.txt
        0.tar
        ...
    classnames.txt
    zeroshot_classification_templates.txt
    dataset_type.txt

Each split should be contained in its own folder and nshards.txt should contain a single integer corresponding to the number of TAR files. The TAR files should follow webdataset format, with an image file (.webp, .png, or .jpg) and a label (.cls) for each example. Classnames and templates are required for zeroshot classification evaluation, with each classname or template on its own line. Dataset type is required for distinguishing zeroshot retrieval evaluation: the file should just contain the text retrieval.

Evaluating on a webdataset

The name of the dataset follows the template wds/<DATASET_NAME>. Note that the dataset name currently only affects the name in the results output - classnames and templates are loaded directly from the included files. The dataset root directory can be either a local path to the root_dir as specified above, or an HTTP URL pointing to a Huggingface Hub dataset file tree.

Example with vtab/cifar10 (zero-shot classification):

• local: clip_benchmark eval --dataset wds/vtab/cifar10 --dataset_root ROOT_DIR/wds_vtab-cifar10/
• remote: clip_benchmark eval --dataset wds/vtab/cifar10 --dataset_root https://huggingface.co/datasets/clip-benchmark/wds_{dataset_cleaned}/tree/main

Example with mscoco_captions (retrieval):

• local: clip_benchmark eval --dataset=wds/mscoco_captions --dataset_root ROOT_DIR/wds_vtab-mscoco_captions/ --task=zeroshot_retrieval
• remote: clip_benchmark eval --dataset=wds/mscoco_captions --dataset_root="https://huggingface.co/datasets/clip-benchmark/wds_{dataset_cleaned}/tree/main" --task=zeroshot_retrieval

Example with mscoco_captions (captioning):

• local: clip_benchmark eval --dataset=wds/mscoco_captions --dataset_root ROOT_DIR/wds_vtab-mscoco_captions/ --task=captioning
• remote: clip_benchmark eval --dataset=wds/mscoco_captions --dataset_root="https://huggingface.co/datasets/clip-benchmark/wds_{dataset_cleaned}/tree/main" --task=captioning

All other arguments remain the same as in the other examples. See https://huggingface.co/clip-benchmark for a full list of datasets that have already been uploaded to Huggingface.

Evaluate mulitple models on multiple datasets

For the purpose of benchmarking, it is possible to run the CLI with multiple pre-trained models on multiple datasets.

Pretrained models and datasets list as arguments

For models, we can provide list of pretrained model names in the form of 'model,pretrained' (so model and pretrained are comma separated). For datasets, we can provide a list of datasets. For languages, we can provide a list of languages. Example:

clip_benchmark eval --pretrained_model  ViT-B-32-quickgelu,laion400m_e32 ViT-L-14,laion400m_e32  \
--dataset cifar10 cifar100 --dataset_root "clip_benchmark_datasets/{dataset}" --language en jp \
 --output "{dataset}_{pretrained}_{model}_{language}_{task}.json"

Note that --dataset_root and --output can be now in the form of a template that depends on the dataset/model/language/task (for --output) and dataset name (for --dataset_root).

Note that If the benchmark fails at some point, it is possible to resume it by skipping already evaluated models using --skip_existing.

Pretrained models and datasets list as files

We can also provide a path to files with models (each line is in the form of 'model,pretrained' where model and pretrained are comma separated) and datasets list (one dataset per line):

clip_benchmark eval --pretrained_model  benchmark/models.txt \
--dataset benchmark/datasets.txt --dataset_root "clip_benchmark_datasets/{dataset}"  \
 --output "{dataset}_{pretrained}_{model}_{language}_{task}.json"

Examples are available in benchmark/datasets.txt and benchmark/models.txt

Multiple checkpoints from the same model

It is also common to evaluate multiple checkpoints from the same model:

clip_benchmark eval --model ViT-B-32 --pretrained *.pt \
--dataset benchmark/datasets.txt --dataset_root "clip_benchmark_datasets/{dataset}"  \
 --output "{dataset}_{pretrained}_{model}_{language}_{task}.json"

Model and dataset collections

We can also provide model collection names (openai, openclip_base, openclip_multilingual, openclip_full are supported) or dataset collection names (vtab, vtab+, retrieval, imagenet_robustness are supported):

clip_benchmark eval --pretrained_model openai openclip_base  --dataset vtab+ retrieval \
--dataset_root "clip_benchmark_datasets/{dataset}" --not quiet \
--output "{dataset}_{pretrained}_{model}_{language}_{task}.json"

See clip_benchmark/models.py#L6 and clip_benchmark/datasets/builder.py#L634 for more information about the collections.

Custom templates / prompts / classnames

It is also possible to use custom prompts by providing a custom template file and/or a custom classname file. For instance:

clip_benchmark eval --dataset "imagenet1k" --model ViT-B-32 --pretrained laion400m_e32 --custom_template_file <path> --custom_classname_file <path>

The template file can be either in the usual format https://github.com/LAION-AI/CLIP_benchmark/blob/main/clip_benchmark/datasets/en_zeroshot_classification_templates.json or in the CuPL format https://github.com/LAION-AI/CLIP_benchmark/blob/main/clip_benchmark/datasets/cupl_prompts.json to have class-specific prompts. In the case of the CuPL format, the classnames file will not be used, thus one only needs to provide the template file --custom_template_file.

For instance, the prompts from the CuPL paper https://arxiv.org/abs/2209.03320 for ImagetNet-1k can be used this way :

clip_benchmark eval --dataset "imagenet1k" --model ViT-B-32 --pretrained laion400m_e32 --custom_template_file cupl_prompts.json

Development

For development, you can also do this:

git clone https://github.com/LAION-AI/CLIP_benchmark
cd CLIP_benchmark
python setup.py install

Credits

• Thanks to OpenCLIP authors, zero-shot accuracy code is adapted from there and pre-trained models are used in the command line interface.
• Thanks to SLIP authors, some zero-shot templates and classnames are from there.
• Thanks to Wise-ft authors, Imagenet robustness datasets code is adapted from there
• Thanks to LiT authors, some zero-shot templates and classnames of VTAB datasets are from there.
• Thanks to Sugar Crepe authors for compositionality tasks evaluation on COCO
• Thanks to Babel ImageNet authors for multilingual evaluation of ImageNet-1k zero-shot classification.
• Thanks to ImageNet-W authors for ImageNet-W evaluation
• Thanks to CuPL for CuPL prompts.
• Thanks to PyCOCOevalcap and @gpucce for COCO captions image captioning evaluation.
• Thanks to @li-xirong for chinese Flickr-30k/FLickr-8k.
• Thanks to Chinese CLIP authors for chinese ImageNet-1k classnames/prompts (zero-shot classification).
• Thanks to @rinnakk and @mkshing for japanese ImageNet-1k classnames/prompts (zero-shot classification) and japanese CLIP support.
• Thanks to @KhalidAlt for arabic ImageNet-1k classnames/prompts (zero-shot classification).
• Thanks to @djghosh13 for WebDataset support.
• Thanks to @FreddeFrallan for for multilingual COCO.
• Thanks to @mitchellnw for linear probing support.
• Thanks to @teasgen for support of validation set and tuning linear probing similar to OpenAI's CLIP.
• Thanks to @visheratin for multilingual retrieval datasets support from https://arxiv.org/abs/2309.01859.
• This package was created with Cookiecutter and the audreyr/cookiecutter-pypackage project template. Thanks to the author.

History

1.6.1

• Fix missing sugar crepe example #119 thanks to @samarth4149

1.6.0

• Fix overwritten zeroshot templates issue (https://github.com/LAION-AI/CLIP_benchmark/issues/109)
• Support new multilingual retrieval datasets: Crossmodal-3600, XTD10, Flickr30k-200, and XTD200
• Support tuning linear probing on validation set

1.5.0

• Custom classnames and templates
• support wds for captioning evaluation
• support imagenet-w
• support babel imagenet
• support chinese flickr30k/8k
• support sugar crepe (compositionality)
• support (optional) sharding evaluation based on rank, for parallel runs
• fix many issues

1.4.0

• Fix silent webdataset error-handling
• Added support for wds/voc2007_multilabel
• default to float32
• add mscoco generative benchmark

1.3.0

• update flickr8k results, solve issue #48, thanks to @orchidmajumder
• Evaluate multiple models/datasets/languages using the CLI directly
• Support Japanese CLIP by rinna
• Add arabic imagenet
• updating CuPL prompts with more generated sentences + ensembled with openAI prompts
• put model in eval mode before evaluation
• Webdataset updates
• Make verbose the default

1.2.0

• Added support for loading webdatasets

1.1.0

• Added better support for multilingual eval
• Added better support for linear probing
• Added support for CuPL prompts

1.0.1

• pypi description as markdown

1.0.0

• Actual first release on PyPI.

0.1.0

• First release on PyPI.