Img encoding

[NotNANtoN]

Added:

• an img can not be fed into Imagine (can be PIL Image or the path to a file that can be loaded with PIL.open. I added a transform pipeline that just resizes, takes the center crop and normalizes)
• a custom encoding can be used too (so users can do their embedding arithmetic and just feed in the CLIP feature vector)
• made the text in Imagine optional (this change might break some stuff in the CLI - we could also change this to be non-optional and just set it to None when not needed)
• added a setter such that the encoding of Imagine can be adjusted on the fly. See set_clip_encoding of Imagine (takes a text, img, or encoding as an input)
• adjusted the naming scheme for textpath, created a create_text_path() function that merges the name of the input image and the given text.

Please feel free to tell me what things to change or to adapt it as you like. I just thought other people would appreciate these features too.

[NotNANtoN]

I played around with the following image:

I used it as the "img" input to the Imagine model to generate this (does not look amazing, but it works):

https://user-images.githubusercontent.com/19983153/107889810-638b4500-6f15-11eb-9871-2f3f4050eb44.mp4

I used the create_img_encoding and create_text_encoding function from Imagine to get the encoding for the hot-dog image and the sentence "Yellow" and took the average of them. I fed this encoding in Imagine to generate this:

https://user-images.githubusercontent.com/19983153/107889841-93d2e380-6f15-11eb-9895-8269f3d39040.mp4

And this with "Pink":

https://user-images.githubusercontent.com/19983153/107889860-abaa6780-6f15-11eb-8445-ea18db181680.mp4

And with something more abstract "Love is the answer!":

https://user-images.githubusercontent.com/19983153/107889911-19569380-6f16-11eb-8ede-14cc603aa740.mp4

[afiaka87]

Edit: I stand corrected. This is pretty cool!.

@NotNANtoN i'm not sure if you're aware of this, but I believe this feature is already implemented. Although the saving and manipulating of CLIP embeds is cool stuff!

[afiaka87]

Oh i see you've added quite a few more knobs to turn than the original implementation. Apologies.

[lucidrains]

@NotNANtoN looks great! thank you for the contribution :)