WIP: Retrieval Augmented Diffusion Models

[isamu-isozaki]

Pulled code from patil's branch to start making the retriever class and training script for rdm. I'll base this code on

https://github.com/afiaka87/retrieval-augmented-diffusion

and

CompVis/latent-diffusion#111

[isamu-isozaki]

I found that huggingface datasets library already has faiss integration. Trying to figure out how to combine with CLIPVisionModel now

[isamu-isozaki]

@patil-suraj Hi! I moved to this branch. I started working on the retrieval class. Will keep working/testing on it tomorrow

[isamu-isozaki]

I think I almost got the retrieval class working. Will prob finish today and push then work on the training script tomorrow. Will also post example results

[isamu-isozaki]

Success! I'll update scripts and push

[isamu-isozaki]

This is an example! Currently I embed the oxford pet database with the faiss index and I can use knn to get the 10 nearest neighbors using the text clip.

The code I used for this example can be found in examples/rdm in a jupyter notebook. I'll remove the notebook later on for clean code.

[isamu-isozaki]

Next I'll start making a training script

[isamu-isozaki]

I'll try using clip-retrieval tool by laion too as an option for training. I'll double check paper on the implementation then I'll go ahead and train

[neverix]

I get

ValueError: Thecomponent<class'transformers.models.clip.image_processing_clip.CLIPImageProcessor'>of<class'diffusers.pipelines.rdm.pipeline_rdm.RDMPipeline'>cannotbeloadedasitdoesnotseemtohaveanyoftheloadingmethodsdefinedin {'ModelMixin': ['save_pretrained', 'from_pretrained'], 'SchedulerMixin': ['save_config', 'from_config'], 'DiffusionPipeline': ['save_pretrained', 'from_pretrained'], 'OnnxRuntimeModel': ['save_pretrained', 'from_pretrained'], 'PreTrainedTokenizer': ['save_pretrained', 'from_pretrained'], 'PreTrainedTokenizerFast': ['save_pretrained', 'from_pretrained'], 'PreTrainedModel': ['save_pretrained', 'from_pretrained'], 'FeatureExtractionMixin': ['save_pretrained', 'from_pretrained']}.

with fusing/rdm. What could cause this?

[isamu-isozaki]

@neverix
Is this the error you get when doing

from diffusers import RDMPipeline
pipe = RDMPipeline.from_pretrained("fusing/rdm")
pipe.to("cuda")
prompt = "a happy pineapple" images = pipe(prompt).images

or

retrieved_images = # a list of PIL images images = pipe(prompt, retrieved_images=retrieved_images).images

?

[isamu-isozaki]

I'll try to reproduce the problem and let you know. It is pretty weird since the ImageProcessor does not have a load from pretrained method so it's pretty weird trying to load it.

[neverix]

@neverix Is this the error you get when doing

from diffusers import RDMPipeline
pipe = RDMPipeline.from_pretrained("fusing/rdm")
pipe.to("cuda")
prompt = "a happy pineapple" images = pipe(prompt).images

or

retrieved_images = # a list of PIL images images = pipe(prompt, retrieved_images=retrieved_images).images

?

I get it in the model creation step, so there's not much of a difference

[isamu-isozaki]

@neverix tnx was able to reproduce. I'll try figuring out a fix tomorrow

[isamu-isozaki]

Interesting I got the same result when I did

pipeline = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")

[isamu-isozaki]

ok! I made some changes so the retriever can index with a general model if given the argument given moco, simclr, ibot etc. Next, I will wrap out the training/inference

[isamu-isozaki]

once that's done I'll fix the checks!

[isamu-isozaki]

This is for my personal research but I also will try adding ibot embeddings support too. Honestly I doubt anything will need changing

[isamu-isozaki]

I think I'll remove clip-retrieval from the script for now since

1. Training with this script will end up doing a lot of requests to their web service which is probably not their intended use case
2. The code will be way cleaner without supporting both clip-retrieval and Retriever
   Let me know if anyone has a demand for having clip-retrieval in their script and I'll add it back in!

[isamu-isozaki]

Cleaned up the inference script some more. I think a lot of the common funcs I'll abstract away into some common files like datasets just to avoid copying code wrongly.

[isamu-isozaki]

tomorrow I'll hopefully finish cleaning up the training scripts and might ask for a review again!

[isamu-isozaki]

Not finished yet but some notes

1. I can offload a lot of the processes in the dataset by having a pre-computation step where we compute the top k nearest neighbors beforehand and put them in a column. In that way, we save computation on the retrieval
2. The same can be done with a lot of the processes in diffusers, not just here. Like for example, precomputing text embeddings, vae embeddings etc.

[isamu-isozaki]

Anyways stopping a bit here for now but will resume tomorrow!

[isamu-isozaki]

@patrickvonplaten@patil-suraj Hi! I think the training scripts might take a while so I can move them to a separate pr for an easier review!

[isamu-isozaki]

For now, will be cleaning anyway!

[williamberman]

hey @isamu-isozaki if we could isolate the PR to just the pipeline and remove the collosalai pipeline and the training scripts that would be helpful for getting the PR merged

[isamu-isozaki]

@williamberman Got it! Sounds good. Will do tomorrow

[isamu-isozaki]

@williamberman Hi! Just confirming but do you think I should keep the inference scripts? Can remove them too!

[isamu-isozaki]

Let me remove it for now.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[isamu-isozaki]

Ah, let me clean up a bit more

[isamu-isozaki]

sorry got a bit preoccupied. Let me close this pr and open it once I clean up some things

[un1tz3r0]

Hi so i believe i implemented something similar using clip retrieval back in the days of disco-diffusion (we've come such a long way since then)

I took care to implement an asyncio based high performance paralell downloader, it can grab thousands of images from the URLs returned by clip-retrieval in pretty reasonable time.

The repo is here un1tz3r0/anythingdiffusion

[isamu-isozaki]

@un1tz3r0 nice! Looks awesome thanks