Mismatch between provided laion400 text embedding .npy files and the clip ViT-B/32 text encoder output

[TianRui-Song717]

When processing LAION400 embedding files, I found that the embedding in the provided files: https://the-eye.eu/public/AI/cah/laion400m-met-release/laion400m-embeddings/texts/text_emb_0.npy
Does not match the output of clip text encoder.

clip_encoder, _ = clip.load("ViT-B/32", device='cpu')
caption = pd.read_parquet("H:\\LAION\\embeddings\\metadata\\metadata_0.parquet").loc[0, "caption"]

tokenize_caption = clip.tokenize([caption])
with torch.no_grad():
    caption_emb = clip_encoder.encode_text(tokenize_caption)
    caption_emb = caption_emb.detach().numpy()

laion_text_emb = np.load("H:\\LAION\\embeddings\\texts\\text_emb_0.npy")[0]

Here the laion_text_emb does not match caption_emb , even when I normalize the clip output caption_emb / np.linalg.norm(caption_emb), they are still not the same, what's the problem?

[rom1504]

See #264