* Various Fixes for Flax Dreambooth
- Correctly update the progress bar every epoch
- Allow specifying a pretrained VAE
- Allow specifying a revision to pretrained models
- Cache compiled models between invocations (speeds up TPU execution a lot!)
- Save intermediate checkpoints by specifying `save_steps`
* Don't die when save_steps is not set.
* Address CR
* Address comments
* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Suraj Patil <surajp815@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Support training SD V2 with Flax
Mostly involves supporting a v_prediction scheduler.
The implementation in #1777 doesn't take into account a recent refactor of `scheduling_utils_flax`, so this should be used instead.
* Add to other top-level files.
* misc fixes
* more comments
* Update examples/textual_inversion/textual_inversion.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* set transformers verbosity to warning
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Make xformers optional even if it is available
* Raise exception if xformers is used but not available
* Rename use_xformers to enable_xformers_memory_efficient_attention
* Add a note about xformers in README
* Reformat code style
* Section header for in-painting, inference from checkpoint.
* Inference: link to section to perform inference from checkpoint.
* Move Dreambooth in-painting instructions to the proper place.
* [Flax] Stateless schedulers, fixes and refactors
* Remove scheduling_common_flax and some renames
* Update src/diffusers/schedulers/scheduling_pndm_flax.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* expose polynomial:power and cosine_with_restarts:num_cycles using get_scheduler func, add it to train_dreambooth.py
* fix formatting
* fix style
* Update src/diffusers/optimization.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Fail if there are less images than the effective batch size.
* Remove lr-scheduler arg as it's currently ignored.
* Make guidance_scale work for batch_size > 1.
* Add state checkpointing to other training scripts
* Fix first_epoch
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Update Dreambooth checkpoint help message.
* Dreambooth docs: checkpoints, inference from a checkpoint.
* make style
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Dreambooth: save / restore training state.
* make style
* Rename vars for clarity.
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Remove unused import
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* dreambooth: fix#1566: maintain fp32 wrapper when saving a checkpoint to avoid crash when running fp16
* dreambooth: guard against passing keep_fp32_wrapper arg to older versions of accelerate. part of fix for #1566
* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update examples/dreambooth/train_dreambooth.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
easy fix for undefined name in train_dreambooth.py
import_model_class_from_model_name_or_path loads a pretrained model
and refers to args.revision in a context where args is undefined. I modified
the function to take revision as an argument and modified the invocation
of the function to pass in the revision from args. Seems like this was caused
by a cut and paste.
* add check_min_version for examples
* move __version__ to the top
* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* fix comment
* fix error_message
* adapt the install message
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
The mask and instance image were being cropped in different ways without --center_crop, causing the model to learn to ignore the mask in some cases. This PR fixes that and generate more consistent results.
* Create train_dreambooth_inpaint.py
train_dreambooth.py adapted to work with the inpaint model, generating random masks during the training
* Update train_dreambooth_inpaint.py
refactored train_dreambooth_inpaint with black
* Update train_dreambooth_inpaint.py
* Update train_dreambooth_inpaint.py
* Update train_dreambooth_inpaint.py
Fix prior preservation
* add instructions to readme, fix SD2 compatibility
* Make errors for invalid options without "--with_prior_preservation"
* Make --instance_prompt required
* Removed needless check because --instance_data_dir is marked with required
* Updated messages
* Use logger.warning instead of raise errors
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* [Better scheduler docs] Improve usage examples of schedulers
* finish
* fix warnings and add test
* finish
* more replacements
* adapt fast tests hf token
* correct more
* Apply suggestions from code review
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Integrate compatibility with euler
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Make training code usable by external scripts
Add parameter inputs to training and argument parsing function to allow this script to be used by an external call.
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
