Jong Wook Kim

> I think you need to install 'clip-by-openai', not 'clip': `pip3 install clip-by-openai` Please note that `clip-by-openai` is not managed/endorsed by OpenAI. I realize the OP might've installed `clip` from...

Hi, thanks for reporting this issue. I suspect it's due to [CUDA's nondeterministic behavior](https://pytorch.org/docs/stable/notes/randomness.html), as it doesn't happen in the CPU mode. Slight numerical imprecision is sometimes inevitable, especially when...

Thanks for more clues; I observed that too. I don't fully grasp what is happening under the hood, but will keep this issue open in case there's a fix possible,...

Yeah that was why I needed to require 1.7.1. Putting an assert is a good idea, thanks!

I think they're not completely redundant because the 'identity' connection in `ResidualAttentionBlock` does not go through layernorms.

We had it since when we initially iterated on transformer language model where it had ResNet features as prefix (LM RN50 in paper), and it got kinda stuck in the...

Did you try uninstalling and reinstalling the package? Alternatively, you can manually download the checkpoint and call `clip.load` with its path.

We *might* be able to provide the ImageNet classification head obtained during evaluation. Are you looking to reproduce the zero-shot performance shown in the first figure of the blog (with...

Generally speaking, I'd recommend resizing/cropping the image to the size the the model expects.

Depending on the environment/versions, the combination of conda/pip might do unexpected/undesired things .. The CLIP package should be compatible with any pytorch version 1.7.1 and after, and its [`requirements.txt`](https://github.com/openai/CLIP/blob/main/requirements.txt) was...