I cannot reproduce what your UI is doing.

[ivanlengyel]

Hi. I am trying to search over the LAION dataset using custom text.

Basically I want to reproduce what you are doing with your UI.
I downloaded the files from here: https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/

# embed text
device = 'cuda'
clip_model="ViT-B/32"
model, preprocess = clip.load(clip_model, device=device, jit=False)

text = "a dog on the snow"
text_tokens = clip.tokenize([text], truncate=True)
text_features = model.encode_text(text_tokens.to(device))
text_features /= text_features.norm(dim=-1, keepdim=True)
text_embeddings = text_features.cpu().detach().numpy().astype('float32')

# find embedding
ind = faiss.read_index("/datasets/indexs_32/text.index")
D, I = ind.search(text_embeddings, 10000)

the indexes provided by the search do not match the ones in the parquet files. And those who match the parquet files do not have nothing to do with the text query.
I was hoping that I was going to be able to reproduce and obtain similar images to the ones in your UI.
Any hint?

I think that it could be some kind of index mapping in the file ivf_old_to_new_mapping.npy but I am not able to open it.

Thank you.

[rom1504]

https://deploy.laion.ai/8f83b608504d46bb81708ec86e912220/embeddings/metadata/ did you use these parquet files sorted in alphabetical order ?

[MasterScrat]

Just to confirm: when you say "alphabetical", you do mean the first 5 files should be:

metadata_0.parquet
metadata_1.parquet
metadata_10.parquet
metadata_100.parquet
metadata_101.parquet 

as opposed to this, right?

metadata_0.parquet
metadata_1.parquet
metadata_2.parquet
metadata_3.parquet
metadata_4.parquet 

[rom1504]

Yes

[ivanlengyel]

I am using these ones.
Which dataframe indexes did you use to construct the faiss index?

I guess that not the ones that I am using...

[rom1504]

Yes that's the problem, you need to use the ones of the link I sent above

[ivanlengyel]

Ok, thank you!

[rom1504]

Yes

…

On Sun, May 1, 2022, 00:41 Florian ***@***.***> wrote: Just to confirm: when you say "alphabetical", you do mean the first 5 files should be: metadata_0.parquet metadata_1.parquet metadata_10.parquet metadata_100.parquet metadata_101.parquet as opposed to this, right? metadata_0.parquet metadata_1.parquet metadata_2.parquet metadata_3.parquet metadata_4.parquet — Reply to this email directly, view it on GitHub <#104 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAR437XAQKFXHTT72KPICLTVHWZJTANCNFSM5NJIQH5Q> . You are receiving this because you commented.Message ID: ***@***.***>