Feature/make installable

[boba-and-beer]

I moved everything into modules to make it PIP Installable.
I currently have created a package which can be found here: https://pypi.org/project/clip-by-openai/.
This should make CLIP a lot easier to use for individuals looking to test it out. Furthermore, I have also added the model to VectorHub, an open-source repository for vector search models I maintain, which can be found here:
https://hub.getvectorai.com/model/text_image%2Fclip

The link may be useful for people looking for low barriers to entry when trying out CLIP.

[jongwook]

Hi, thanks for your interest in CLIP and packaging it for easier use.

We cannot merge this PR at this state because it has a few issues we'd like to fix, such as the overly complex .gitignore, unintended audience and supported Python versions in setup.py, recommendation of installing pytorch and cudatoolkit via PIP in README.md, and multiple references to your company.

We are actually working on an update to make this package PIP-installable, and we are planning to make it work this way:

pip install git+https://github.com/openai/CLIP.git

so that we don't have to worry about versioning for now.

Would you like me to make edits directly on this branch before merging this PR, or would you rather close this PR entirely?

[boba-and-beer]

That's a good way to go about it. I would like you to make edits on this branch if possible. Would love to track how this goes and make whatever small contribution I can for CLIP. Found the paper and blog quite insightful and I'm surprised with the results! I have also removed all mentions of VectorAI from this PR.

[boba-and-beer]

Closing for now.

[jamt9000]

@jongwook since it seems to work fine for pytorch 1.7 could you relax the requirements to

pytorch~=1.7
torchvision==0.8.*

so that pip install works? (torchvision from source ends up having version number 0.8.0a0 hence the *)

[haltakov]

@jongwook since it seems to work fine for pytorch 1.7 could you relax the requirements to

pytorch~=1.7
torchvision==0.8.*

so that pip install works? (torchvision from source ends up having version number 0.8.0a0 hence the *)

I did some experiments with torch==1.7.0 and it fails in very subtle ways. Most of the time everything seems fine, but there is some very strange behavior at other times. If you are interested I can create a small demo showing the behavior.

[jamt9000]

@haltakov I see - yes it would be good to know how it fails

[brunodoamaral]

@haltakov I see - yes it would be good to know how it fails

See #13. I had the same strange behavior reported in this issue.

[haltakov]

@haltakov I see - yes it would be good to know how it fails

Here the experiments on Colab:

• 1.7.0: https://colab.research.google.com/drive/1KcHMeI2N-FFthMEYip3A1ZNbccdRK4XR?usp=sharing
• 1.7.1: https://colab.research.google.com/drive/1h3pyWaZ0gA4DYWoip49W1skBmkxZ_gDC?usp=sharing

@brunodoamaral - yeah it seems like the same problem