sentence-transformers/clip-ViT-B-32-multilingual-v1

This is a multi-lingual version of the OpenAI CLIP-ViT-B32 model. You can map text (in 50+ languages) and images to a common dense vector space such that images and the matching texts are close. This model can be used for image search (users search through a large collection of images) and for multi-lingual zero-shot image classification (image labels are defined as text).

Multilingual Image Search - Demo

For a demo of multilingual image search, have a look at: Image_Search-multilingual.ipynb ( Colab version )

For more details on image search and zero-shot image classification, have a look at the documentation on SBERT.net.

Training

This model has been created using Multilingual Knowledge Distillation. As teacher model, we used the original clip-ViT-B-32 and then trained a multilingual DistilBERT model as student model. Using parallel data, the multilingual student model learns to align the teachers vector space across many languages. As a result, you get an text embedding model that works for 50+ languages.

The image encoder from CLIP is unchanged, i.e. you can use the original CLIP image encoder to encode images.

Have a look at the SBERT.net - Multilingual-Models documentation on more details and for training code.

We used the following 50+ languages to align the vector spaces: ar, bg, ca, cs, da, de, el, es, et, fa, fi, fr, fr-ca, gl, gu, he, hi, hr, hu, hy, id, it, ja, ka, ko, ku, lt, lv, mk, mn, mr, ms, my, nb, nl, pl, pt, pt, pt-br, ro, ru, sk, sl, sq, sr, sv, th, tr, uk, ur, vi, zh-cn, zh-tw.

The original multilingual DistilBERT supports 100+ lanugages. The model also work for these languages, but might not yield the best results.

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: DistilBertModel
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
  (2): Dense({'in_features': 768, 'out_features': 512, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity'})
)

Citing & Authors

This model was trained by sentence-transformers.

If you find this model helpful, feel free to cite our publication Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks:

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "http://arxiv.org/abs/1908.10084",
}