requirements.txt reduction

[largestack]

One suggestion on the repo:

• For the backend requirements.txt, I think it's a freeze of all your packages. I think focusing on the key packages without pinning to specific versions unless needed.

Initial impression from running it on 1K images from sankaku dump (anime mostly):

• Looks like a great sound technique. MobileNetV2 embedding --> PCA dimensionality reduciton --> KMeans. Is this a methodology you leveraged from elsewhere?
• SUPER fast, clustered 1K images in around 1 minutes
• Clusters seem to not be as much composition and camera angle, but more artistic style. Like carnoony, pencil drawing, vibrant bright, etc.