【bug】do_sample= False still get random output

[penghao-wu]

Please provide more details about your case. Do you modify anything or directly run the inference.py?

[CSEEduanyu]

Please provide more details about your case. Do you modify anything or directly run the inference.py?

yes,I just changed the loading of the visual model locally, my changes should not affect the randomness.

[penghao-wu]

Can you provide thecode after modification?

[CSEEduanyu]

clip_model, processor = create_model_from_pretrained(self.vision_tower_name)

local_bin = self.vision_tower_name + "/open_clip_pytorch_model.bin" clip_model, processor = create_model_from_pretrained("convnext_xxlarge",pretrained=local_bin) just like this

[penghao-wu]

I think the problem probably comes from this local loading. But without your exact code and weight, I can't not reproduce your problem.

[CSEEduanyu]

I think the problem probably comes from this local loading. But without your exact code and weight, I can't not reproduce your problem.

how can i try webdemo of cam34b?just like https://internvl.opengvlab.com/

[penghao-wu]

We will release the public demo very soon.

[CSEEduanyu]

We will release the public demo very soon.

after some logging,i found same picture after CLIP-ConvNext get different embeding,this is why output get diff.Other vison model is ok ，What's so special about CLIP-ConvNext？

[CSEEduanyu]

@penghao-wu Or about when will the demo be available online? I wanted to verify that the offline deployment was correct

[CSEEduanyu]

@penghao-wu https://github.com/mlfoundations/open_clip/issues/865

[CSEEduanyu]

open_clip model need clip_model.eval() after create_model_from_pretrained()