[Feature] Finetune text encoder in train_text_to_image_lora #3912
+230
−31
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
|
cc @sayakpaul here |
|
Thanks for adding this pipeline. Could you also share some comparative results with and without text encoder LoRA fine-tuning? Just trying to gauge the effectiveness. |
|
@sayakpaul only unettest prompt = '1girl, X X' test prompt = '1girl, >_<' test prompt = '1girl, @_@' test prompt = '1girl, =_=' with text encodertest prompt = '1girl, X X' test prompt = '1girl, >_<' test prompt = '1girl, @_@' test prompt = '1girl, =_=' test prompt = '1girl, Jitome' test prompt = '1girl, :I' test prompt = '1girl, ._.' |
|
Thanks for your hard work! Maybe let's wait for #3778 to get merged, as there are some refactoring-related changes that will make this simpler. Okay for you? |
|
@sayakpaul No problems. |
|
Gentle ping @sayakpaul - do you think we could merge this? |
|
@patrickvonplaten Should we fix some codes based on this PR? |
|
@sayakpaul I updated codes. But it caused following errors. |
|
Hi, Could you also try out the solutions provided in that thread to see if the errors persist? Also, what happens if we do this in Torch 2.0 taking advantage of SDPA disabling xformers? |
|
When disabling xformers, |
|
I tried some pattens, but all failed. |
|
@sayakpaul pytorch2.0 and |
|
Thanks for letting me know and for your efforts. Could we maybe try to dive a bit deeper to see what is the issue? For example does the issue persist when not training the text encoder? |
It also persist when training only unet. |
|
Okay. Can we verify if the version in the main branch works with the following settings?
Just trying to double down the issue. If you have other ideas to try out please let me know. |
|
PyTorch 2.0 and SDPA works well. |
|
Thanks for reporting. This is indeed weird as the |
|
The error raised when I used my branch of the script. The script of |
|
Oh okay. Let's try to investigate the differences then :-) |
|
The error solved when changing the place of applying xformers. |
|
@sayakpaul This PR is ready. |
sayakpaul
reviewed
Jul 17, 2023
Comment on lines
+474
to
+479
| xformers_version = version.parse(xformers.__version__) | ||
| if xformers_version == version.parse("0.0.16"): | ||
| logger.warn( | ||
| "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." | ||
| ) | ||
| unet.enable_xformers_memory_efficient_attention() |
sayakpaul
reviewed
Jul 17, 2023
There was a problem hiding this comment.
Excellent work here. Thanks so much for all your experiments and iterations!
Let's also update the README of the example so that readers are aware that text encoder training is supported? We also need to add a test case for for the text encoder training to test_examples.py. Then this PR is good to be shipped 🚀
Also, could we do a full run with the changes to ensure the results you're getting are as expected?
https://civitai.com/models/25613/classic-anime-expressions You can download images from this page's Params are followings, |
sayakpaul
reviewed
Jul 18, 2023
sayakpaul
reviewed
Jul 18, 2023
docs/source/en/training/lora.mdx
Outdated
Comment on lines
185
to
196
| from huggingface_hub.repocard import RepoCard | ||
| from diffusers import StableDiffusionPipeline | ||
| import torch | ||
|
|
||
| lora_model_id = "sayakpaul/sd-model-finetuned-lora-t4" | ||
| card = RepoCard.load(lora_model_id) | ||
| base_model_id = card.data.to_dict()["base_model"] | ||
|
|
||
| pipe = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16) | ||
| pipe = pipe.to("cuda") | ||
| pipe.load_lora_weights(lora_model_id) | ||
| image = pipe("A picture of a sks dog in a bucket", num_inference_steps=25).images[0] |
There was a problem hiding this comment.
Let's use a LoRA you trained with this script? We'd also need to update the example prompt :)
sayakpaul
reviewed
Jul 18, 2023
|
@williamberman could you also take a look here? |
Co-authored-by: Sayak Paul <[email protected]>
Co-authored-by: Sayak Paul <[email protected]>
Comment on lines
+470
to
+482
| if args.enable_xformers_memory_efficient_attention: | ||
| if is_xformers_available(): | ||
| import xformers | ||
|
|
||
| xformers_version = version.parse(xformers.__version__) | ||
| if xformers_version == version.parse("0.0.16"): | ||
| logger.warn( | ||
| "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details." | ||
| ) | ||
| unet.enable_xformers_memory_efficient_attention() | ||
| else: | ||
| raise ValueError("xformers is not available. Make sure it is installed correctly") | ||
|
|
There was a problem hiding this comment.
unless I'm missing something, In the future it'd be helpful to not move a code block unrelated to the PR as it makes the diff harder to read :)
There was a problem hiding this comment.
#3912 (comment)
#3912 (comment)
By moving the code block, I avoid this error.
| images.append( | ||
| pipeline(args.validation_prompt, num_inference_steps=30, generator=generator).images[0] | ||
| ) | ||
| with torch.cuda.amp.autocast(): |
There was a problem hiding this comment.
did something change that required autocast to be added?
There was a problem hiding this comment.
It also related to this error.
You can check this thread.
|
Looks basically good, a few small questions :) |
|
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. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Fixes #3418
Add train_text_encoder args in train_text_to_image_lora.py. We can finetune text encoder.
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.