Oh c’mon you guys...

[kcrosley-leisurelabs]

Colab notebook?

(OpenAI’s dedication to crap documentation never ceases to amaze, amirite?)

[Kreijstal]

I'd like a colab notebook too but it's probably too much to ask.

[JonathanFly]

Here you go. Just does sampling, I didn't implement image completion.

https://gist.github.com/JonathanFly/eb61f0d31680e1b890f3a53fbaf31384

[methescripter]

@JonathanFly on the codelab How can I pass images as input? thanks

[JonathanFly]

@JonathanFly on the codelab How can I pass images as input? thanks

That wasn't built in to the code released. It shouldn't be too hard to add but someone will have to do it. They do provide an encoding function for their cluster palette. I'll probably add it if I play more with this in the future.

[Kreijstal]

backlink huggingface/transformers#5088

[apeguero1]

Hey guys, I modified @JonathanFly 's notebook to support conditional input such as cropped images.

colab

[JonathanFly]

Thanks @apeguero1 you improved it a lot.

[apeguero1]

No problem @JonathanFly glad to help (:

[dryn27]

@apeguero1 I tried Runtime>Run All, but it doesn't work. GPU is not the problem, and I cannot understand what the problem is. On "Functions from run.py", on "import tensorflow as tf", it shows an error and it says:

---

NOTE: If your import is failing due to a missing package, you can
manually install dependencies using either !pip or !apt.

To view examples of installing some common dependencies, click the
"Open Examples" button below.

[apeguero1]

Thanks for pointing this out @dryn27. Somehow bumping up to tensorflow version 1.15.0 fixes the import error. Just updated the colab.

[Randy-H0]

Hey guys, I modified @JonathanFly 's notebook to support conditional input such as cropped images.

colab

Hi, Is there any way to sample with a higher resolution?

[eyalbetzalel]

@apeguero1

Hi! thanks for the colab with the conditional sampling.

is it possible to generate samples from a specific class from ImageNet?

Generating photos of cars for example.

[apeguero1]

@eyalbetzalel hmm... in theory you could just finetune the whole model on a subset of the images with a particular label. Haven't tried this yet though.

Alternatively, Maybe you could make an adapter per class starting with the HF model made here which would seem to be more efficient.

Ideally, there would be a way to train a single slightly modified architecture of the model that just adds a learned class conditional embedding to the token and positional embeddings. That would be pretty cool.

[proudbatdan]

Why’d you email me idk who you are Sent from ProtonMail Mobile

…

On Mon, Oct 12, 2020 at 3:49 AM, DJ AI ***@***.***> wrote: ***@***.***(https://github.com/eyalbetzalel) hmm... in theory you could just finetune the whole model on a subset of the images with a particular label. Haven't tried this yet though. Alternatively, Maybe you could make an [adapter](https://github.com/Adapter-Hub/adapter-transformers) per class starting with the HF model made [here](https://colab.research.google.com/github/apeguero1/image-gpt/blob/master/Transformers_Image_GPT.ipynb) which would seem to be more efficient. Ideally, there would be a way to train a single slightly modified architecture of the model that just adds a learned class conditional embedding to the token and positional embeddings. That would be pretty cool. — You are receiving this because you are subscribed to this thread. Reply to this email directly, [view it on GitHub](#2 (comment)), or [unsubscribe](https://github.com/notifications/unsubscribe-auth/ARKFAKOI2EANPI7VBF7TGJ3SKLNL3ANCNFSM4OBEEBFA).

[apeguero1]

@proudbatdan you might be signed up to watch this repo for any activity. You can always disable this.