* 7529 do not disable autocast for cuda devices
* Remove typecasting error check for non-mps platforms, as a correct autocast implementation makes it a non-issue
* add autocast fix to other training examples
* disable native_amp for dreambooth (sdxl)
* disable native_amp for pix2pix (sdxl)
* remove tests from remaining files
* disable native_amp on huggingface accelerator for every training example that uses it
* convert more usages of autocast to nullcontext, make style fixes
* make style fixes
* style.
* Empty-Commit
---------
Co-authored-by: bghira <bghira@users.github.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
you cannot specify `type="bool"` and `action="store_true"` at the same time.
remove excessive and buggy `type=bool`.
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>
* add tags for diffusers training
* add tags for diffusers training
* add tags for diffusers training
* add tags for diffusers training
* add tags for diffusers training
* add tags for diffusers training
* add dora tags for drambooth lora scripts
* style
* add is_dora arg
* style
* add dora training feature to sd 1.5 script
* added notes about DoRA training
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* add noise_offset param
* micro conditioning - wip
* image processing adjusted and moved to support micro conditioning
* change time ids to be computed inside train loop
* change time ids to be computed inside train loop
* change time ids to be computed inside train loop
* time ids shape fix
* move token replacement of validation prompt to the same section of instance prompt and class prompt
* add offset noise to sd15 advanced script
* fix token loading during validation
* fix token loading during validation in sdxl script
* a little clean
* style
* a little clean
* style
* sdxl script - a little clean + minor path fix
sd 1.5 script - change default resolution value
* ad 1.5 script - minor path fix
* fix missing comma in code example in model card
* clean up commented lines
* style
* remove time ids computed outside training loop - no longer used now that we utilize micro-conditioning, as all time ids are now computed inside the training loop
* style
* [WIP] - added draft readme, building off of examples/dreambooth/README.md
* readme
* readme
* readme
* readme
* readme
* readme
* readme
* readme
* removed --crops_coords_top_left from CLI args
* style
* fix missing shape bug due to missing RGB if statement
* add blog mention at the start of the reamde as well
* Update examples/advanced_diffusion_training/README.md
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* change note to render nicely as well
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* sd1.5 support in separate script
A quick adaptation to support people interested in using this method on 1.5 models.
* sd15 prompt text encoding and unet conversions
as per @linoytsaban 's recommendations. Testing would be appreciated,
* Readability and quality improvements
Removed some mentions of SDXL, and some arguments that don't apply to sd 1.5, and cleaned up some comments.
* make style/quality commands
* tracker rename and run-it doc
* Update examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py
* Update examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py
---------
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>
* Fixes#6418 Advanced Dreambooth LoRa Training
* change order of import to fix nit
* fix nit, use cast_training_params
* remove torch.compile fix, will move to a new PR
* remove unnecessary import
* unwrap text encoder when saving hook only for full text encoder tuning
* unwrap text encoder when saving hook only for full text encoder tuning
* save embeddings in each checkpoint as well
* save embeddings in each checkpoint as well
* save embeddings in each checkpoint as well
* Update examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* change timesteps used to calculate snr when --with_prior_preservation is enabled
* change timesteps used to calculate snr when --with_prior_preservation is enabled (canonical script)
* style
* revert canonical script to before snr gamma change
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* load pipeline for inference only if validation prompt is used
* move things outside
* load pipeline for inference only if validation prompt is used
* fix readme when validation prompt is used
---------
Co-authored-by: linoytsaban <linoy@huggingface.co>
Co-authored-by: apolinário <joaopaulo.passos@gmail.com>
* Update train_dreambooth_lora_sdxl_advanced.py
* remove global function args from dreamboothdataset class
* style
* style
---------
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* improve help tags
* style fix
* changes token_abstraction type to string.
support multiple concepts for pivotal using a comma separated string.
* style fixup
* changed logger to warning (not yet available)
* moved the token_abstraction parsing to be in the same block as where we create the mapping of identifier to token
---------
Co-authored-by: Linoy <linoy@huggingface.co>
* imports and readme bug fixes
* bug fix - ensures text_encoder params are dtype==float32 (when using pivotal tuning) even if the rest of the model is loaded in fp16
* added pivotal tuning to readme
* mapping token identifier to new inserted token in validation prompt (if used)
* correct default value of --train_text_encoder_frac
* change default value of --adam_weight_decay_text_encoder
* validation prompt generations when using pivotal tuning bug fix
* style fix
* textual inversion embeddings name change
* style fix
* bug fix - stopping text encoder optimization halfway
* readme - will include token abstraction and new inserted tokens when using pivotal tuning
- added type to --num_new_tokens_per_abstraction
* style fix
---------
Co-authored-by: Linoy Tsaban <linoy@huggingface.co>
