openai/CLIP
1
0
Fork 0
mirror of https://github.com/openai/CLIP.git synced 2026-11-05 01:37:26 +00:00
Code Issues 274 Projects Packages Wiki Activity

[GH-ISSUE #222] clip output is NaN #193

New issue
Closed
opened 2026-10-31 21:12:15 +00:00 by bsakaguchi · 6 comments
bsakaguchi commented 2026-10-31 21:12:15 +00:00
Owner
Copy link

Originally created by @chinoll on GitHub (Mar 7, 2022).
Original GitHub issue: https://github.com/openai/CLIP/issues/222

My CLIP will output NaN when using CUDA, but it will output normally when using CPU. How to solve this problem?

import torch
import clip
from PIL import Image
import numpy as np
device = "cuda:0" #use cuda
model, preprocess = clip.load("ViT-B/32", device=device,jit=False)

text = clip.tokenize(["a photo is comics","a photo is illustration","a photo is empty","a photo is white","a photo is black","a photo is gray","a photo is lineart"]).to(device)
with torch.no_grad():
    image = preprocess(image).unsqueeze(0).to(device)
    logits_per_image, logits_per_text = model(image, text)
    print(logits_per_image)
    probs = logits_per_image.softmax(dim=-1).cpu().numpy()[0]
print(probs)

output

tensor([[nan, nan, nan, nan, nan, nan, nan]], device='cuda:0',
       dtype=torch.float16)
[nan nan nan nan nan nan nan]

CLIP can output the probability normally when using CPU.

Originally created by @chinoll on GitHub (Mar 7, 2022). Original GitHub issue: https://github.com/openai/CLIP/issues/222 My CLIP will output NaN when using CUDA, but it will output normally when using CPU. How to solve this problem? ```python import torch import clip from PIL import Image import numpy as np device = "cuda:0" #use cuda model, preprocess = clip.load("ViT-B/32", device=device,jit=False) text = clip.tokenize(["a photo is comics","a photo is illustration","a photo is empty","a photo is white","a photo is black","a photo is gray","a photo is lineart"]).to(device) with torch.no_grad(): image = preprocess(image).unsqueeze(0).to(device) logits_per_image, logits_per_text = model(image, text) print(logits_per_image) probs = logits_per_image.softmax(dim=-1).cpu().numpy()[0] print(probs) ``` output ``` tensor([[nan, nan, nan, nan, nan, nan, nan]], device='cuda:0', dtype=torch.float16) [nan nan nan nan nan nan nan] ``` CLIP can output the probability normally when using CPU.
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@chinoll commented on GitHub (Mar 7, 2022):

torch version:1.10.2
cuda version:11.4
OS:Windows 10
python version:3.9.5

<!-- gh-comment-id:1060729572 --> @chinoll commented on GitHub (Mar 7, 2022): torch version:1.10.2 cuda version:11.4 OS:Windows 10 python version:3.9.5
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@illtellyoulater commented on GitHub (Mar 19, 2022):

same problem here, did you manage to fix it?

<!-- gh-comment-id:1072915583 --> @illtellyoulater commented on GitHub (Mar 19, 2022): same problem here, did you manage to fix it?
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@chinoll commented on GitHub (Mar 19, 2022):

same problem here, did you manage to fix it?

Are you talking about getting the model to produce the correct output? Using model.float() will make the model produce the correct output, but I want to use float16 and the model does not work properly when the input data is in the format float16.

<!-- gh-comment-id:1072950450 --> @chinoll commented on GitHub (Mar 19, 2022): > same problem here, did you manage to fix it? Are you talking about getting the model to produce the correct output? Using model.float() will make the model produce the correct output, but I want to use float16 and the model does not work properly when the input data is in the format float16.
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@illtellyoulater commented on GitHub (Mar 29, 2022):

@chinoll yeah I also wanted to make it work with the GPU (and float16)... and I think I've fixed this by using a different version of cuda toolkit.

Try uninstalling torch & torchvision and then install them again with pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html.

If it doesn't work, again uninstall them and reinstall them with pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/torch_stable.html.

<!-- gh-comment-id:1082148367 --> @illtellyoulater commented on GitHub (Mar 29, 2022): @chinoll yeah I also wanted to make it work with the GPU (and float16)... and I think I've fixed this by using a different version of cuda toolkit. Try uninstalling torch & torchvision and then install them again with `pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html`. If it doesn't work, again uninstall them and reinstall them with `pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/torch_stable.html`.
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@SunZhigang7 commented on GitHub (Mar 31, 2022):

pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html

useful!! thanks

<!-- gh-comment-id:1084802567 --> @SunZhigang7 commented on GitHub (Mar 31, 2022): > pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html useful!! thanks
bsakaguchi commented 2026-10-31 21:12:17 +00:00
Author
Owner
Copy link

@chinoll commented on GitHub (Apr 7, 2022):

@chinoll yeah I also wanted to make it work with the GPU (and float16)... and I think I've fixed this by using a different version of cuda toolkit.

Try uninstalling torch & torchvision and then install them again with pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html.

If it doesn't work, again uninstall them and reinstall them with pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/torch_stable.html.

Great!

<!-- gh-comment-id:1091028999 --> @chinoll commented on GitHub (Apr 7, 2022): > @chinoll yeah I also wanted to make it work with the GPU (and float16)... and I think I've fixed this by using a different version of cuda toolkit. > > Try uninstalling torch & torchvision and then install them again with `pip3 install torch==1.11.0+cu115 torchvision==0.12.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html`. > > If it doesn't work, again uninstall them and reinstall them with `pip install torch==1.10.1+cu102 torchvision==0.11.2+cu102 -f https://download.pytorch.org/whl/torch_stable.html`. Great!
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
codex/pin-github-workflow-refs-20260325-132838
jongwook/issue-396
jongwook-patch-1
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
openai/CLIP#193
No description provided.