mirror of
https://github.com/huggingface/diffusers.git
synced 2026-01-27 17:22:53 +03:00
[WIP] Add UniDiffuser model and pipeline (#2963)
* Fix a bug of pano when not doing CFG (#3030) * Fix a bug of pano when not doing CFG * enhance code quality * apply formatting. --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * Text2video zero refinements (#3070) * fix progress bar issue in pipeline_text_to_video_zero.py. Copy scheduler after first backward * fix tensor loading in test_text_to_video_zero.py * make style && make quality * Release: v0.15.0 * [Tests] Speed up panorama tests (#3067) * fix: norm group test for UNet3D. * chore: speed up the panorama tests (fast). * set default value of _test_inference_batch_single_identical. * fix: batch_sizes default value. * [Post release] v0.16.0dev (#3072) * Adds profiling flags, computes train metrics average. (#3053) * WIP controlnet training - bugfix --streaming - bugfix running report_to!='wandb' - adds memory profile before validation * Adds final logging statement. * Sets train epochs to 11. Looking at a longer ~16ep run, we see only good validation images after ~11ep: https://wandb.ai/andsteing/controlnet_fill50k/runs/3j2hx6n8 * Removes --logging_dir (it's not used). * Adds --profile flags. * Updates --output_dir=runs/fill-circle-{timestamp}. * Compute mean of `train_metrics`. Previously `train_metrics[-1]` was logged, resulting in very bumpy train metrics. * Improves logging a bit. - adds l2_grads gradient norm logging - adds steps_per_sec - sets walltime as x coordinate of train/step - logs controlnet_params config * Adds --ccache (doesn't really help though). * minor fix in controlnet flax example (#2986) * fix the error when push_to_hub but not log validation * contronet_from_pt & controlnet_revision * add intermediate checkpointing to the guide * Bugfix --profile_steps * Sets `RACKER_PROJECT_NAME='controlnet_fill50k'`. * Logs fractional epoch. * Adds relative `walltime` metric. * Adds `StepTraceAnnotation` and uses `global_step` insetad of `step`. * Applied `black`. * Streamlines commands in README a bit. * Removes `--ccache`. This makes only a very small difference (~1 min) with this model size, so removing the option introduced in cdb3cc. * Re-ran `black`. * Update examples/controlnet/README.md Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * Converts spaces to tab. * Removes repeated args. * Skips first step (compilation) in profiling * Updates README with profiling instructions. * Unifies tabs/spaces in README. * Re-ran style & quality. --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [Pipelines] Make sure that None functions are correctly not saved (#3080) * doc string example remove from_pt (#3083) * [Tests] parallelize (#3078) * [Tests] parallelize * finish folder structuring * Parallelize tests more * Correct saving of pipelines * make sure logging level is correct * try again * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Throw deprecation warning for return_cached_folder (#3092) Throw deprecation warning * Allow SD attend and excite pipeline to work with any size output images (#2835) Allow stable diffusion attend and excite pipeline to work with any size output image. Re: #2476, #2603 * [docs] Update community pipeline docs (#2989) * update community pipeline docs * fix formatting * explain sharing workflows * Add to support Guess Mode for StableDiffusionControlnetPipleline (#2998) * add guess mode (WIP) * fix uncond/cond order * support guidance_scale=1.0 and batch != 1 * remove magic coeff * add docstring * add intergration test * add document to controlnet.mdx * made the comments a bit more explanatory * fix table * fix default value for attend-and-excite (#3099) * fix default * remvoe one line as requested by gc team (#3077) remvoe one line * ddpm custom timesteps (#3007) add custom timesteps test add custom timesteps descending order check docs timesteps -> custom_timesteps can only pass one of num_inference_steps and timesteps * Fix breaking change in `pipeline_stable_diffusion_controlnet.py` (#3118) fix breaking change * Add global pooling to controlnet (#3121) * [Bug fix] Fix img2img processor with safety checker (#3127) Fix img2img processor with safety checker * [Bug fix] Make sure correct timesteps are chosen for img2img (#3128) Make sure correct timesteps are chosen for img2img * Improve deprecation warnings (#3131) * Fix config deprecation (#3129) * Better deprecation message * Better deprecation message * Better doc string * Fixes * fix more * fix more * Improve __getattr__ * correct more * fix more * fix * Improve more * more improvements * fix more * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * make style * Fix all rest & add tests & remove old deprecation fns --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * feat: verfication of multi-gpu support for select examples. (#3126) * feat: verfication of multi-gpu support for select examples. * add: multi-gpu training sections to the relvant doc pages. * speed up attend-and-excite fast tests (#3079) * Optimize log_validation in train_controlnet_flax (#3110) extract pipeline from log_validation * make style * Correct textual inversion readme (#3145) * Update README.md * Apply suggestions from code review * Add unet act fn to other model components (#3136) Adding act fn config to the unet timestep class embedding and conv activation. The custom activation defaults to silu which is the default activation function for both the conv act and the timestep class embeddings so default behavior is not changed. The only unet which use the custom activation is the stable diffusion latent upscaler https://huggingface.co/stabilityai/sd-x2-latent-upscaler/blob/main/unet/config.json (I ran a script against the hub to confirm). The latent upscaler does not use the conv activation nor the timestep class embeddings so we don't change its behavior. * class labels timestep embeddings projection dtype cast (#3137) This mimics the dtype cast for the standard time embeddings * [ckpt loader] Allow loading the Inpaint and Img2Img pipelines, while loading a ckpt model (#2705) * [ckpt loader] Allow loading the Inpaint and Img2Img pipelines, while loading a ckpt model * Address review comment from PR * PyLint formatting * Some more pylint fixes, unrelated to our change * Another pylint fix * Styling fix * add from_ckpt method as Mixin (#2318) * add mixin class for pipeline from original sd ckpt * Improve * make style * merge main into * Improve more * fix more * up * Apply suggestions from code review * finish docs * rename * make style --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Add TensorRT SD/txt2img Community Pipeline to diffusers along with TensorRT utils (#2974) * Add SD/txt2img Community Pipeline to diffusers along with TensorRT utils Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * update installation command Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * update tensorrt installation Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * changes 1. Update setting of cache directory 2. Address comments: merge utils and pipeline code. 3. Address comments: Add section in README Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * apply make style Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> --------- Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Correct `Transformer2DModel.forward` docstring (#3074) ⚙️chore(transformer_2d) update function signature for encoder_hidden_states * Update pipeline_stable_diffusion_inpaint_legacy.py (#2903) * Update pipeline_stable_diffusion_inpaint_legacy.py * fix preprocessing of Pil images with adequate batch size * revert map * add tests * reformat * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * next try to fix the style * wth is this * Update testing_utils.py * Update testing_utils.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py * Update test_stable_diffusion_inpaint_legacy.py --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Modified altdiffusion pipline to support altdiffusion-m18 (#2993) * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 * Modified altdiffusion pipline to support altdiffusion-m18 --------- Co-authored-by: root <fulong_ye@163.com> * controlnet training resize inputs to multiple of 8 (#3135) controlnet training center crop input images to multiple of 8 The pipeline code resizes inputs to multiples of 8. Not doing this resizing in the training script is causing the encoded image to have different height/width dimensions than the encoded conditioning image (which uses a separate encoder that's part of the controlnet model). We resize and center crop the inputs to make sure they're the same size (as well as all other images in the batch). We also check that the initial resolution is a multiple of 8. * adding custom diffusion training to diffusers examples (#3031) * diffusers==0.14.0 update * custom diffusion update * custom diffusion update * custom diffusion update * custom diffusion update * custom diffusion update * custom diffusion update * custom diffusion * custom diffusion * custom diffusion * custom diffusion * custom diffusion * apply formatting and get rid of bare except. * refactor readme and other minor changes. * misc refactor. * fix: repo_id issue and loaders logging bug. * fix: save_model_card. * fix: save_model_card. * fix: save_model_card. * add: doc entry. * refactor doc,. * custom diffusion * custom diffusion * custom diffusion * apply style. * remove tralining whitespace. * fix: toctree entry. * remove unnecessary print. * custom diffusion * custom diffusion * custom diffusion test * custom diffusion xformer update * custom diffusion xformer update * custom diffusion xformer update --------- Co-authored-by: Nupur Kumari <nupurkumari@Nupurs-MacBook-Pro.local> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Nupur Kumari <nupurkumari@nupurs-mbp.wifi.local.cmu.edu> * make style * Update custom_diffusion.mdx (#3165) Add missing newlines for rendering the links correctly * Added distillation for quantization example on textual inversion. (#2760) * Added distillation for quantization example on textual inversion. Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> * refined readme and code style. Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> * Update text2images.py * refined code of model load and added compatibility check. Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> * fixed code style. Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> * fix C403 [*] Unnecessary `list` comprehension (rewrite as a `set` comprehension) Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> --------- Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> * Update Noise Autocorrelation Loss Function for Pix2PixZero Pipeline (#2942) * Update Pix2PixZero Auto-correlation Loss * Add fast inversion tests * Clarify purpose and mark as deprecated Fix inversion prompt broadcasting * Register modules set to `None` in config for `test_save_load_optional_components` * Update new tests to coordinate with #2953 * [DreamBooth] add text encoder LoRA support in the DreamBooth training script (#3130) * add: LoRA text encoder support for DreamBooth example. * fix initialization. * fix: modification call. * add: entry in the readme. * use dog dataset from hub. * fix: params to clip. * add entry to the LoRA doc. * add: tests for lora. * remove unnecessary list comprehension./ * Update Habana Gaudi documentation (#3169) * Update Habana Gaudi doc * Fix tables * Add model offload to x4 upscaler (#3187) * Add model offload to x4 upscaler * fix * [docs] Deterministic algorithms (#3172) deterministic algos * Update custom_diffusion.mdx to credit the author (#3163) * Update custom_diffusion.mdx * fix: unnecessary list comprehension. * Fix TensorRT community pipeline device set function (#3157) pass silence_dtype_warnings as kwarg Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make `from_flax` work for controlnet (#3161) fix from_flax Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [docs] Clarify training args (#3146) * clarify training arg * apply feedback * Multi Vector Textual Inversion (#3144) * Multi Vector * Improve * fix multi token * improve test * make style * Update examples/test_examples.py * Apply suggestions from code review Co-authored-by: Suraj Patil <surajp815@gmail.com> * update * Finish * Apply suggestions from code review --------- Co-authored-by: Suraj Patil <surajp815@gmail.com> * Add `Karras sigmas` to HeunDiscreteScheduler (#3160) * Add karras pattern to discrete heun scheduler * Add integration test * Fix failing CI on pytorch test on M1 (mps) --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [AudioLDM] Fix dtype of returned waveform (#3189) * Fix bug in train_dreambooth_lora (#3183) * Update train_dreambooth_lora.py fix bug * Update train_dreambooth_lora.py * [Community Pipelines] Update lpw_stable_diffusion pipeline (#3197) * Update lpw_stable_diffusion.py * fix cpu offload * Make sure VAE attention works with Torch 2_0 (#3200) * Make sure attention works with Torch 2_0 * make style * Fix more * Revert "[Community Pipelines] Update lpw_stable_diffusion pipeline" (#3201) Revert "[Community Pipelines] Update lpw_stable_diffusion pipeline (#3197)" This reverts commit9965cb50ea. * [Bug fix] Fix batch size attention head size mismatch (#3214) * fix mixed precision training on train_dreambooth_inpaint_lora (#3138) cast to weight dtype * adding enable_vae_tiling and disable_vae_tiling functions (#3225) adding enable_vae_tiling and disable_val_tiling functions * Add ControlNet v1.1 docs (#3226) Add v1.1 docs * Fix issue in maybe_convert_prompt (#3188) When the token used for textual inversion does not have any special symbols (e.g. it is not surrounded by <>), the tokenizer does not properly split the replacement tokens. Adding a space for the padding tokens fixes this. * Sync cache version check from transformers (#3179) sync cache version check from transformers * Fix docs text inversion (#3166) * Fix docs text inversion * Apply suggestions from code review * add model (#3230) * add * clean * up * clean up more * fix more tests * Improve docs further * improve * more fixes docs * Improve docs more * Update src/diffusers/models/unet_2d_condition.py * fix * up * update doc links * make fix-copies * add safety checker and watermarker to stage 3 doc page code snippets * speed optimizations docs * memory optimization docs * make style * add watermarking snippets to doc string examples * make style * use pt_to_pil helper functions in doc strings * skip mps tests * Improve safety * make style * new logic * fix * fix bad onnx design * make new stable diffusion upscale pipeline model arguments optional * define has_nsfw_concept when non-pil output type * lowercase linked to notebook name --------- Co-authored-by: William Berman <WLBberman@gmail.com> * Allow return pt x4 (#3236) * Add all files * update * Allow fp16 attn for x4 upscaler (#3239) * Add all files * update * Make sure vae is memory efficient for PT 1 * make style * fix fast test (#3241) * Adds a document on token merging (#3208) * add document on token merging. * fix headline. * fix: headline. * add some samples for comparison. * [AudioLDM] Update docs to use updated ckpt (#3240) * [AudioLDM] Update docs to use updated ckpt * make style * Release: v0.16.0 * Post release for 0.16.0 (#3244) * Post release * fix more * [docs] only mention one stage (#3246) * [docs] only mention one stage * add blurb on auto accepting --------- Co-authored-by: William Berman <WLBberman@gmail.com> * Write model card in controlnet training script (#3229) Write model card in controlnet training script. * [2064]: Add stochastic sampler (sample_dpmpp_sde) (#3020) * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * [2064]: Add stochastic sampler * Review comments * [Review comment]: Add is_torchsde_available() * [Review comment]: Test and docs * [Review comment] * [Review comment] * [Review comment] * [Review comment] * [Review comment] --------- Co-authored-by: njindal <njindal@adobe.com> * [Stochastic Sampler][Slow Test]: Cuda test fixes (#3257) [Slow Test]: Cuda test fixes Co-authored-by: njindal <njindal@adobe.com> * Remove required from tracker_project_name (#3260) Remove required from tracker_project_name. As observed by https://github.com/off99555 in https://github.com/huggingface/diffusers/issues/2695#issuecomment-1470755050, it already has a default value. * adding required parameters while calling the get_up_block and get_down_block (#3210) * removed unnecessary parameters from get_up_block and get_down_block functions * adding resnet_skip_time_act, resnet_out_scale_factor and cross_attention_norm to get_up_block and get_down_block functions --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [docs] Update interface in repaint.mdx (#3119) Update repaint.mdx accomodate to #1701 * Update IF name to XL (#3262) Co-authored-by: multimodalart <joaopaulo.passos+multimodal@gmail.com> * fix typo in score sde pipeline (#3132) * Fix typo in textual inversion JAX training script (#3123) The pipeline is built as `pipe` but then used as `pipeline`. * AudioDiffusionPipeline - fix encode method after config changes (#3114) * config fixes * deprecate get_input_dims * Revert "Revert "[Community Pipelines] Update lpw_stable_diffusion pipeline"" (#3265) Revert "Revert "[Community Pipelines] Update lpw_stable_diffusion pipeline" (#3201)" This reverts commit91a2a80eb2. * Fix community pipelines (#3266) * update notebook (#3259) Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> * [docs] add notes for stateful model changes (#3252) * [docs] add notes for stateful model changes * Update docs/source/en/optimization/fp16.mdx Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * link to accelerate docs for discarding hooks --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * [LoRA] quality of life improvements in the loading semantics and docs (#3180) * 👽 qol improvements for LoRA. * better function name? * fix: LoRA weight loading with the new format. * address Patrick's comments. * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * change wording around encouraging the use of load_lora_weights(). * fix: function name. --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Community Pipelines] EDICT pipeline implementation (#3153) * EDICT pipeline initial commit - Starting point taking from https://github.com/Joqsan/edict-diffusion * refactor __init__() method * minor refactoring * refactor scheduler code - remove scheduler and move its methods to the EDICTPipeline class * make CFG optional - refactor encode_prompt(). - include optional generator for sampling with vae. - minor variable renaming * add EDICT pipeline description to README.md * replace preprocess() with VaeImageProcessor * run make style and make quality commands --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Docs]zh translated docs update (#3245) * zh translated docs update * update _toctree * Update logging.mdx (#2863) Fix typos * Add multiple conditions to StableDiffusionControlNetInpaintPipeline (#3125) * try multi controlnet inpaint * multi controlnet inpaint * multi controlnet inpaint * Let's make sure that dreambooth always uploads to the Hub (#3272) * Update Dreambooth README * Adapt all docs as well * automatically write model card * fix * make style * Diffedit Zero-Shot Inpainting Pipeline (#2837) * Update Pix2PixZero Auto-correlation Loss * Add Stable Diffusion DiffEdit pipeline * Add draft documentation and import code * Bugfixes and refactoring * Add option to not decode latents in the inversion process * Harmonize preprocessing * Revert "Update Pix2PixZero Auto-correlation Loss" This reverts commitb218062fed. * Update annotations * rename `compute_mask` to `generate_mask` * Update documentation * Update docs * Update Docs * Fix copy * Change shape of output latents to batch first * Update docs * Add first draft for tests * Bugfix and update tests * Add `cross_attention_kwargs` support for all pipeline methods * Fix Copies * Add support for PIL image latents Add support for mask broadcasting Update docs and tests Align `mask` argument to `mask_image` Remove height and width arguments * Enable MPS Tests * Move example docstrings * Fix test * Fix test * fix pipeline inheritance * Harmonize `prepare_image_latents` with StableDiffusionPix2PixZeroPipeline * Register modules set to `None` in config for `test_save_load_optional_components` * Move fixed logic to specific test class * Clean changes to other pipelines * Update new tests to coordinate with #2953 * Update slow tests for better results * Safety to avoid potential problems with torch.inference_mode * Add reference in SD Pipeline Overview * Fix tests again * Enforce determinism in noise for generate_mask * Fix copies * Widen test tolerance for fp16 based on `test_stable_diffusion_upscale_pipeline_fp16` * Add LoraLoaderMixin and update `prepare_image_latents` * clean up repeat and reg * bugfix * Remove invalid args from docs Suppress spurious warning by repeating image before latent to mask gen * add constant learning rate with custom rule (#3133) * add constant lr with rules * add constant with rules in TYPE_TO_SCHEDULER_FUNCTION * add constant lr rate with rule * hotfix code quality * fix doc style * change name constant_with_rules to piecewise constant * Allow disabling torch 2_0 attention (#3273) * Allow disabling torch 2_0 attention * make style * Update src/diffusers/models/attention.py * [doc] add link to training script (#3271) add link to training script Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> * temp disable spectogram diffusion tests (#3278) The note-seq package throws an error on import because the default installed version of Ipython is not compatible with python 3.8 which we run in the CI. https://github.com/huggingface/diffusers/actions/runs/4830121056/jobs/8605954838#step:7:9 * Changed sample[0] to images[0] (#3304) A pipeline object stores the results in `images` not in `sample`. Current code blocks don't work. * Typo in tutorial (#3295) * Torch compile graph fix (#3286) * fix more * Fix more * fix more * Apply suggestions from code review * fix * make style * make fix-copies * fix * make sure torch compile * Clean * fix test * Postprocessing refactor img2img (#3268) * refactor img2img VaeImageProcessor.postprocess * remove copy from for init, run_safety_checker, decode_latents Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> --------- Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [Torch 2.0 compile] Fix more torch compile breaks (#3313) * Fix more torch compile breaks * add tests * Fix all * fix controlnet * fix more * Add Horace He as co-author. > > Co-authored-by: Horace He <horacehe2007@yahoo.com> * Add Horace He as co-author. Co-authored-by: Horace He <horacehe2007@yahoo.com> --------- Co-authored-by: Horace He <horacehe2007@yahoo.com> * fix: scale_lr and sync example readme and docs. (#3299) * fix: scale_lr and sync example readme and docs. * fix doc link. * Update stable_diffusion.mdx (#3310) fixed import statement * Fix missing variable assign in DeepFloyd-IF-II (#3315) Fix missing variable assign lol * Correct doc build for patch releases (#3316) Update build_documentation.yml * Add Stable Diffusion RePaint to community pipelines (#3320) * Add Stable Diffsuion RePaint to community pipelines - Adds Stable Diffsuion RePaint to community pipelines - Add Readme enty for pipeline * Fix: Remove wrong import - Remove wrong import - Minor change in comments * Fix: Code formatting of stable_diffusion_repaint * Fix: ruff errors in stable_diffusion_repaint * Fix multistep dpmsolver for cosine schedule (suitable for deepfloyd-if) (#3314) * fix multistep dpmsolver for cosine schedule (deepfloy-if) * fix a typo * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * update all dpmsolver (singlestep, multistep, dpm, dpm++) for cosine noise schedule * add test, fix style --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [docs] Improve LoRA docs (#3311) * update docs * add to toctree * apply feedback * Added input pretubation (#3292) * Added input pretubation * Fixed spelling * Update write_own_pipeline.mdx (#3323) * update controlling generation doc with latest goodies. (#3321) * [Quality] Make style (#3341) * Fix config dpm (#3343) * Add the SDE variant of DPM-Solver and DPM-Solver++ (#3344) * add SDE variant of DPM-Solver and DPM-Solver++ * add test * fix typo * fix typo * Add upsample_size to AttnUpBlock2D, AttnDownBlock2D (#3275) The argument `upsample_size` needs to be added to these modules to allow compatibility with other blocks that require this argument. * Add UniDiffuser classes to __init__ files, modify transformer block to support pre- and post-LN, add fast default tests, fix some bugs. * Update fast tests to use test checkpoints stored on the hub and to better match the reference UniDiffuser implementation. * Fix code with make style. * Revert "Fix code style with make style." This reverts commit10a174a12c. * Add self.image_encoder, self.text_decoder to list of models to offload to CPU in the enable_sequential_cpu_offload(...)/enable_model_cpu_offload(...) methods to make test_cpu_offload_forward_pass pass. * Fix code quality with make style. * Support using a data type embedding for UniDiffuser-v1. * Add fast test for checking UniDiffuser-v1 sampling. * Make changes so that the repository consistency tests pass. * Add UniDiffuser dummy objects via make fix-copies. * Fix bugs and make improvements to the UniDiffuser pipeline: - Improve batch size inference and fix bugs when num_images_per_prompt or num_prompts_per_image > 1 - Add tests for num_images_per_prompt, num_prompts_per_image > 1 - Improve check_inputs, especially regarding checking supplied latents - Add reset_mode method so that mode inference can be re-enabled after mode is set manually - Fix some warnings related to accessing class members directly instead of through their config - Small amount of refactoring in pipeline_unidiffuser.py * Fix code style with make style. * Add/edit docstrings for added classes and public pipeline methods. Also do some light refactoring. * Add documentation for UniDiffuser and fix some typos/formatting in docstrings. * Fix code with make style. * Refactor and improve the UniDiffuser convert_from_ckpt.py script. * Move the UniDiffusers convert_from_ckpy.py script to diffusers/scripts/convert_unidiffuser_to_diffusers.py * Fix code quality via make style. * Improve UniDiffuser slow tests. * make style * Fix some typos in the UniDiffuser docs. * Remove outdated logic based on transformers version in UniDiffuser pipeline __init__.py * Remove dependency on einops by refactoring einops operations to pure torch operations. * make style * Add slow test on full checkpoint for joint mode and correct expected image slices/text prefixes. * make style * Fix mixed precision issue by wrapping the offending code with the torch.autocast context manager. * Revert "Fix mixed precision issue by wrapping the offending code with the torch.autocast context manager." This reverts commit1a58958ab4. * Add fast test for CUDA/fp16 model behavior (currently failing). * Fix the mixed precision issue and add additional tests of the pipeline cuda/fp16 functionality. * make style * Use a CLIPVisionModelWithProjection instead of CLIPVisionModel for image_encoder to better match the original UniDiffuser implementation. * Make style and remove some testing code. * Fix shape errors for the 'joint' and 'img2text' modes. * Fix tests and remove some testing code. * Add option to use fixed latents for UniDiffuserPipelineSlowTests and fix issue in modeling_text_decoder.py. * Improve UniDiffuser docs, particularly the usage examples, and improve slow tests with new expected outputs. * make style * Fix examples to load model in float16. * In image-to-text mode, sample from the autoencoder moment distribution instead of always getting its mode. * make style * When encoding the image using the VAE, scale the image latents by the VAE's scaling factor. * make style * Clean up code and make slow tests pass. * make fix-copies * [docs] Fix docstring (#3334) fix docstring Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * if dreambooth lora (#3360) * update IF stage I pipelines add fixed variance schedulers and lora loading * added kv lora attn processor * allow loading into alternative lora attn processor * make vae optional * throw away predicted variance * allow loading into added kv lora layer * allow load T5 * allow pre compute text embeddings * set new variance type in schedulers * fix copies * refactor all prompt embedding code class prompts are now included in pre-encoding code max tokenizer length is now configurable embedding attention mask is now configurable * fix for when variance type is not defined on scheduler * do not pre compute validation prompt if not present * add example test for if lora dreambooth * add check for train text encoder and pre compute text embeddings * Postprocessing refactor all others (#3337) * add text2img * fix-copies * add * add all other pipelines * add * add * add * add * add * make style * style + fix copies --------- Co-authored-by: yiyixuxu <yixu310@gmail,com> * [docs] Improve safetensors docstring (#3368) * clarify safetensor docstring * fix typo * apply feedback * add: a warning message when using xformers in a PT 2.0 env. (#3365) * add: a warning message when using xformers in a PT 2.0 env. * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * StableDiffusionInpaintingPipeline - resize image w.r.t height and width (#3322) * StableDiffusionInpaintingPipeline now resizes input images and masks w.r.t to passed input height and width. Default is already set to 512. This addresses the common tensor mismatch error. Also moved type check into relevant funciton to keep main pipeline body tidy. * Fixed StableDiffusionInpaintingPrepareMaskAndMaskedImageTests Due to previous commit these tests were failing as height and width need to be passed into the prepare_mask_and_masked_image function, I have updated the code and added a height/width variable per unit test as it seemed more appropriate than the current hard coded solution * Added a resolution test to StableDiffusionInpaintPipelineSlowTests this unit test simply gets the input and resizes it into some that would fail (e.g. would throw a tensor mismatch error/not a mult of 8). Then passes it through the pipeline and verifies it produces output with correct dims w.r.t the passed height and width --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make style * [docs] Adapt a model (#3326) * first draft * apply feedback * conv_in.weight thrown away * [docs] Load safetensors (#3333) * safetensors * apply feedback * apply feedback * Apply suggestions from code review --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make style * [Docs] Fix stable_diffusion.mdx typo (#3398) Fix typo in last code block. Correct "prommpts" to "prompt" * Support ControlNet v1.1 shuffle properly (#3340) * add inferring_controlnet_cond_batch * Revert "add inferring_controlnet_cond_batch" This reverts commitabe8d6311d. * set guess_mode to True whenever global_pool_conditions is True Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * nit * add integration test --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Tests] better determinism (#3374) * enable deterministic pytorch and cuda operations. * disable manual seeding. * make style && make quality for unet_2d tests. * enable determinism for the unet2dconditional model. * add CUBLAS_WORKSPACE_CONFIG for better reproducibility. * relax tolerance (very weird issue, though). * revert to torch manual_seed() where needed. * relax more tolerance. * better placement of the cuda variable and relax more tolerance. * enable determinism for 3d condition model. * relax tolerance. * add: determinism to alt_diffusion. * relax tolerance for alt diffusion. * dance diffusion. * dance diffusion is flaky. * test_dict_tuple_outputs_equivalent edit. * fix two more tests. * fix more ddim tests. * fix: argument. * change to diff in place of difference. * fix: test_save_load call. * test_save_load_float16 call. * fix: expected_max_diff * fix: paint by example. * relax tolerance. * add determinism to 1d unet model. * torch 2.0 regressions seem to be brutal * determinism to vae. * add reason to skipping. * up tolerance. * determinism to vq. * determinism to cuda. * determinism to the generic test pipeline file. * refactor general pipelines testing a bit. * determinism to alt diffusion i2i * up tolerance for alt diff i2i and audio diff * up tolerance. * determinism to audioldm * increase tolerance for audioldm lms. * increase tolerance for paint by paint. * increase tolerance for repaint. * determinism to cycle diffusion and sd 1. * relax tol for cycle diffusion 🚲 * relax tol for sd 1.0 * relax tol for controlnet. * determinism to img var. * relax tol for img variation. * tolerance to i2i sd * make style * determinism to inpaint. * relax tolerance for inpaiting. * determinism for inpainting legacy * relax tolerance. * determinism to instruct pix2pix * determinism to model editing. * model editing tolerance. * panorama determinism * determinism to pix2pix zero. * determinism to sag. * sd 2. determinism * sd. tolerance * disallow tf32 matmul. * relax tolerance is all you need. * make style and determinism to sd 2 depth * relax tolerance for depth. * tolerance to diffedit. * tolerance to sd 2 inpaint. * up tolerance. * determinism in upscaling. * tolerance in upscaler. * more tolerance relaxation. * determinism to v pred. * up tol for v_pred * unclip determinism * determinism to unclip img2img * determinism to text to video. * determinism to last set of tests * up tol. * vq cumsum doesn't have a deterministic kernel * relax tol * relax tol * [docs] Add transformers to install (#3388) add transformers to install * [deepspeed] partial ZeRO-3 support (#3076) * [deepspeed] partial ZeRO-3 support * cleanup * improve deepspeed fixes * Improve * make style --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Add omegaconf for tests (#3400) Add omegaconfg * Fix various bugs with LoRA Dreambooth and Dreambooth script (#3353) * Improve checkpointing lora * fix more * Improve doc string * Update src/diffusers/loaders.py * make stytle * Apply suggestions from code review * Update src/diffusers/loaders.py * Apply suggestions from code review * Apply suggestions from code review * better * Fix all * Fix multi-GPU dreambooth * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix all * make style * make style --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix docker file (#3402) * up * up * fix: deepseepd_plugin retrieval from accelerate state (#3410) * [Docs] Add `sigmoid` beta_scheduler to docstrings of relevant Schedulers (#3399) * Add `sigmoid` beta scheduler to `DDPMScheduler` docstring * Add `sigmoid` beta scheduler to `RePaintScheduler` docstring --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Don't install accelerate and transformers from source (#3415) * Don't install transformers and accelerate from source (#3414) * Improve fast tests (#3416) Update pr_tests.yml * attention refactor: the trilogy (#3387) * Replace `AttentionBlock` with `Attention` * use _from_deprecated_attn_block check re: @patrickvonplaten * [Docs] update the PT 2.0 optimization doc with latest findings (#3370) * add: benchmarking stats for A100 and V100. * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * address patrick's comments. * add: rtx 4090 stats * ⚔ benchmark reports done * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * 3313 pr link. * add: plots. Co-authored-by: Pedro <pedro@huggingface.co> * fix formattimg * update number percent. --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix style rendering (#3433) * Fix style rendering. * Fix typo * unCLIP scheduler do not use note (#3417) * Replace deprecated command with environment file (#3409) Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * fix warning message pipeline loading (#3446) * add stable diffusion tensorrt img2img pipeline (#3419) * add stable diffusion tensorrt img2img pipeline Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * update docstrings Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> --------- Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * Refactor controlnet and add img2img and inpaint (#3386) * refactor controlnet and add img2img and inpaint * First draft to get pipelines to work * make style * Fix more * Fix more * More tests * Fix more * Make inpainting work * make style and more tests * Apply suggestions from code review * up * make style * Fix imports * Fix more * Fix more * Improve examples * add test * Make sure import is correctly deprecated * Make sure everything works in compile mode * make sure authorship is correctly attributed * [Scheduler] DPM-Solver (++) Inverse Scheduler (#3335) * Add DPM-Solver Multistep Inverse Scheduler * Add draft tests for DiffEdit * Add inverse sde-dpmsolver steps to tune image diversity from inverted latents * Fix tests --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Docs] Fix incomplete docstring for resnet.py (#3438) Fix incomplete docstrings for resnet.py * fix tiled vae blend extent range (#3384) fix tiled vae bleand extent range * Small update to "Next steps" section (#3443) Small update to "Next steps" section: - PyTorch 2 is recommended. - Updated improvement figures. * Allow arbitrary aspect ratio in IFSuperResolutionPipeline (#3298) * Update pipeline_if_superresolution.py Allow arbitrary aspect ratio in IFSuperResolutionPipeline by using the input image shape * IFSuperResolutionPipeline: allow the user to override the height and width through the arguments * update IFSuperResolutionPipeline width/height doc string to match StableDiffusionInpaintPipeline conventions --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Adding 'strength' parameter to StableDiffusionInpaintingPipeline (#3424) * Added explanation of 'strength' parameter * Added get_timesteps function which relies on new strength parameter * Added `strength` parameter which defaults to 1. * Swapped ordering so `noise_timestep` can be calculated before masking the image this is required when you aren't applying 100% noise to the masked region, e.g. strength < 1. * Added strength to check_inputs, throws error if out of range * Changed `prepare_latents` to initialise latents w.r.t strength inspired from the stable diffusion img2img pipeline, init latents are initialised by converting the init image into a VAE latent and adding noise (based upon the strength parameter passed in), e.g. random when strength = 1, or the init image at strength = 0. * WIP: Added a unit test for the new strength parameter in the StableDiffusionInpaintingPipeline still need to add correct regression values * Created a is_strength_max to initialise from pure random noise * Updated unit tests w.r.t new strength parameter + fixed new strength unit test * renamed parameter to avoid confusion with variable of same name * Updated regression values for new strength test - now passes * removed 'copied from' comment as this method is now different and divergent from the cpy * Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Ensure backwards compatibility for prepare_mask_and_masked_image created a return_image boolean and initialised to false * Ensure backwards compatibility for prepare_latents * Fixed copy check typo * Fixes w.r.t backward compibility changes * make style * keep function argument ordering same for backwards compatibility in callees with copied from statements * make fix-copies --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: William Berman <WLBberman@gmail.com> * [WIP] Bugfix - Pipeline.from_pretrained is broken when the pipeline is partially downloaded (#3448) Added bugfix using f strings. * Fix gradient checkpointing bugs in freezing part of models (requires_grad=False) (#3404) * gradient checkpointing bug fix * bug fix; changes for reviews * reformat * reformat --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Make dreambooth lora more robust to orig unet (#3462) * Make dreambooth lora more robust to orig unet * up * Reduce peak VRAM by releasing large attention tensors (as soon as they're unnecessary) (#3463) Release large tensors in attention (as soon as they're no longer required). Reduces peak VRAM by nearly 2 GB for 1024x1024 (even after slicing), and the savings scale up with image size. * Add min snr to text2img lora training script (#3459) add min snr to text2img lora training script * Add inpaint lora scale support (#3460) * add inpaint lora scale support * add inpaint lora scale test --------- Co-authored-by: yueyang.hyy <yueyang.hyy@alibaba-inc.com> * [From ckpt] Fix from_ckpt (#3466) * Correct from_ckpt * make style * Update full dreambooth script to work with IF (#3425) * Add IF dreambooth docs (#3470) * parameterize pass single args through tuple (#3477) * attend and excite tests disable determinism on the class level (#3478) * dreambooth docs torch.compile note (#3471) * dreambooth docs torch.compile note * Update examples/dreambooth/README.md Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * Update examples/dreambooth/README.md Co-authored-by: Pedro Cuenca <pedro@huggingface.co> --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * add: if entry in the dreambooth training docs. (#3472) * [docs] Textual inversion inference (#3473) * add textual inversion inference to docs * add to toctree --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [docs] Distributed inference (#3376) * distributed inference * move to inference section * apply feedback * update with split_between_processes * apply feedback * [{Up,Down}sample1d] explicit view kernel size as number elements in flattened indices (#3479) explicit view kernel size as number elements in flattened indices * mps & onnx tests rework (#3449) * Remove ONNX tests from PR. They are already a part of push_tests.yml. * Remove mps tests from PRs. They are already performed on push. * Fix workflow name for fast push tests. * Extract mps tests to a workflow. For better control/filtering. * Remove --extra-index-url from mps tests * Increase tolerance of mps test This test passes in my Mac (Ventura 13.3) but fails in the CI hardware (Ventura 13.2). I ran the local tests following the same steps that exist in the CI workflow. * Temporarily run mps tests on pr So we can test. * Revert "Temporarily run mps tests on pr" Tests passed, go back to running on push. * [Attention processor] Better warning message when shifting to `AttnProcessor2_0` (#3457) * add: debugging to enabling memory efficient processing * add: better warning message. * [Docs] add note on local directory path. (#3397) add note on local directory path. Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Refactor full determinism (#3485) * up * fix more * Apply suggestions from code review * fix more * fix more * Check it * Remove 16:8 * fix more * fix more * fix more * up * up * Test only stable diffusion * Test only two files * up * Try out spinning up processes that can be killed * up * Apply suggestions from code review * up * up * Fix DPM single (#3413) * Fix DPM single * add test * fix one more bug * Apply suggestions from code review Co-authored-by: StAlKeR7779 <stalkek7779@yandex.ru> --------- Co-authored-by: StAlKeR7779 <stalkek7779@yandex.ru> * Add `use_Karras_sigmas` to DPMSolverSinglestepScheduler (#3476) * add use_karras_sigmas * add karras test * add doc * Adds local_files_only bool to prevent forced online connection (#3486) * make style * [Docs] Korean translation (optimization, training) (#3488) * feat) optimization kr translation * fix) typo, italic setting * feat) dreambooth, text2image kr * feat) lora kr * fix) LoRA * fix) fp16 fix * fix) doc-builder style * fix) fp16 일부 단어 수정 * fix) fp16 style fix * fix) opt, training docs update * feat) toctree update * feat) toctree update --------- Co-authored-by: Chanran Kim <seriousran@gmail.com> * DataLoader respecting EXIF data in Training Images (#3465) * DataLoader will now bake in any transforms or image manipulations contained in the EXIF Images may have rotations stored in EXIF. Training using such images will cause those transforms to be ignored while training and thus produce unexpected results * Fixed the Dataloading EXIF issue in main DreamBooth training as well * Run make style (black & isort) * make style * feat: allow disk offload for diffuser models (#3285) * allow disk offload for diffuser models * sort import * add max_memory argument * Changed sample[0] to images[0] (#3304) A pipeline object stores the results in `images` not in `sample`. Current code blocks don't work. * Typo in tutorial (#3295) * Torch compile graph fix (#3286) * fix more * Fix more * fix more * Apply suggestions from code review * fix * make style * make fix-copies * fix * make sure torch compile * Clean * fix test * Postprocessing refactor img2img (#3268) * refactor img2img VaeImageProcessor.postprocess * remove copy from for init, run_safety_checker, decode_latents Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> --------- Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [Torch 2.0 compile] Fix more torch compile breaks (#3313) * Fix more torch compile breaks * add tests * Fix all * fix controlnet * fix more * Add Horace He as co-author. > > Co-authored-by: Horace He <horacehe2007@yahoo.com> * Add Horace He as co-author. Co-authored-by: Horace He <horacehe2007@yahoo.com> --------- Co-authored-by: Horace He <horacehe2007@yahoo.com> * fix: scale_lr and sync example readme and docs. (#3299) * fix: scale_lr and sync example readme and docs. * fix doc link. * Update stable_diffusion.mdx (#3310) fixed import statement * Fix missing variable assign in DeepFloyd-IF-II (#3315) Fix missing variable assign lol * Correct doc build for patch releases (#3316) Update build_documentation.yml * Add Stable Diffusion RePaint to community pipelines (#3320) * Add Stable Diffsuion RePaint to community pipelines - Adds Stable Diffsuion RePaint to community pipelines - Add Readme enty for pipeline * Fix: Remove wrong import - Remove wrong import - Minor change in comments * Fix: Code formatting of stable_diffusion_repaint * Fix: ruff errors in stable_diffusion_repaint * Fix multistep dpmsolver for cosine schedule (suitable for deepfloyd-if) (#3314) * fix multistep dpmsolver for cosine schedule (deepfloy-if) * fix a typo * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/diffusers/schedulers/scheduling_dpmsolver_multistep.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * update all dpmsolver (singlestep, multistep, dpm, dpm++) for cosine noise schedule * add test, fix style --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [docs] Improve LoRA docs (#3311) * update docs * add to toctree * apply feedback * Added input pretubation (#3292) * Added input pretubation * Fixed spelling * Update write_own_pipeline.mdx (#3323) * update controlling generation doc with latest goodies. (#3321) * [Quality] Make style (#3341) * Fix config dpm (#3343) * Add the SDE variant of DPM-Solver and DPM-Solver++ (#3344) * add SDE variant of DPM-Solver and DPM-Solver++ * add test * fix typo * fix typo * Add upsample_size to AttnUpBlock2D, AttnDownBlock2D (#3275) The argument `upsample_size` needs to be added to these modules to allow compatibility with other blocks that require this argument. * Rename --only_save_embeds to --save_as_full_pipeline (#3206) * Set --only_save_embeds to False by default Due to how the option is named, it makes more sense to behave like this. * Refactor only_save_embeds to save_as_full_pipeline * [AudioLDM] Generalise conversion script (#3328) Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Fix TypeError when using prompt_embeds and negative_prompt (#2982) * test: Added test case * fix: fixed type checking issue on _encode_prompt * fix: fixed copies consistency * fix: one copy was not sufficient * Fix pipeline class on README (#3345) Update README.md * Inpainting: typo in docs (#3331) Typo in docs Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Add `use_Karras_sigmas` to LMSDiscreteScheduler (#3351) * add karras sigma to lms discrete scheduler * add test for lms_scheduler karras * reformat test lms * Batched load of textual inversions (#3277) * Batched load of textual inversions - Only call resize_token_embeddings once per batch as it is the most expensive operation - Allow pretrained_model_name_or_path and token to be an optional list - Remove Dict from type annotation pretrained_model_name_or_path as it was not supported in this function - Add comment that single files (e.g. .pt/.safetensors) are supported - Add comment for token parameter - Convert token override log message from warning to info * Update src/diffusers/loaders.py Check for duplicate tokens Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update condition for None tokens --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make fix-copies * [docs] Fix docstring (#3334) fix docstring Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * if dreambooth lora (#3360) * update IF stage I pipelines add fixed variance schedulers and lora loading * added kv lora attn processor * allow loading into alternative lora attn processor * make vae optional * throw away predicted variance * allow loading into added kv lora layer * allow load T5 * allow pre compute text embeddings * set new variance type in schedulers * fix copies * refactor all prompt embedding code class prompts are now included in pre-encoding code max tokenizer length is now configurable embedding attention mask is now configurable * fix for when variance type is not defined on scheduler * do not pre compute validation prompt if not present * add example test for if lora dreambooth * add check for train text encoder and pre compute text embeddings * Postprocessing refactor all others (#3337) * add text2img * fix-copies * add * add all other pipelines * add * add * add * add * add * make style * style + fix copies --------- Co-authored-by: yiyixuxu <yixu310@gmail,com> * [docs] Improve safetensors docstring (#3368) * clarify safetensor docstring * fix typo * apply feedback * add: a warning message when using xformers in a PT 2.0 env. (#3365) * add: a warning message when using xformers in a PT 2.0 env. * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * StableDiffusionInpaintingPipeline - resize image w.r.t height and width (#3322) * StableDiffusionInpaintingPipeline now resizes input images and masks w.r.t to passed input height and width. Default is already set to 512. This addresses the common tensor mismatch error. Also moved type check into relevant funciton to keep main pipeline body tidy. * Fixed StableDiffusionInpaintingPrepareMaskAndMaskedImageTests Due to previous commit these tests were failing as height and width need to be passed into the prepare_mask_and_masked_image function, I have updated the code and added a height/width variable per unit test as it seemed more appropriate than the current hard coded solution * Added a resolution test to StableDiffusionInpaintPipelineSlowTests this unit test simply gets the input and resizes it into some that would fail (e.g. would throw a tensor mismatch error/not a mult of 8). Then passes it through the pipeline and verifies it produces output with correct dims w.r.t the passed height and width --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make style * [docs] Adapt a model (#3326) * first draft * apply feedback * conv_in.weight thrown away * [docs] Load safetensors (#3333) * safetensors * apply feedback * apply feedback * Apply suggestions from code review --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * make style * [Docs] Fix stable_diffusion.mdx typo (#3398) Fix typo in last code block. Correct "prommpts" to "prompt" * Support ControlNet v1.1 shuffle properly (#3340) * add inferring_controlnet_cond_batch * Revert "add inferring_controlnet_cond_batch" This reverts commitabe8d6311d. * set guess_mode to True whenever global_pool_conditions is True Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * nit * add integration test --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Tests] better determinism (#3374) * enable deterministic pytorch and cuda operations. * disable manual seeding. * make style && make quality for unet_2d tests. * enable determinism for the unet2dconditional model. * add CUBLAS_WORKSPACE_CONFIG for better reproducibility. * relax tolerance (very weird issue, though). * revert to torch manual_seed() where needed. * relax more tolerance. * better placement of the cuda variable and relax more tolerance. * enable determinism for 3d condition model. * relax tolerance. * add: determinism to alt_diffusion. * relax tolerance for alt diffusion. * dance diffusion. * dance diffusion is flaky. * test_dict_tuple_outputs_equivalent edit. * fix two more tests. * fix more ddim tests. * fix: argument. * change to diff in place of difference. * fix: test_save_load call. * test_save_load_float16 call. * fix: expected_max_diff * fix: paint by example. * relax tolerance. * add determinism to 1d unet model. * torch 2.0 regressions seem to be brutal * determinism to vae. * add reason to skipping. * up tolerance. * determinism to vq. * determinism to cuda. * determinism to the generic test pipeline file. * refactor general pipelines testing a bit. * determinism to alt diffusion i2i * up tolerance for alt diff i2i and audio diff * up tolerance. * determinism to audioldm * increase tolerance for audioldm lms. * increase tolerance for paint by paint. * increase tolerance for repaint. * determinism to cycle diffusion and sd 1. * relax tol for cycle diffusion 🚲 * relax tol for sd 1.0 * relax tol for controlnet. * determinism to img var. * relax tol for img variation. * tolerance to i2i sd * make style * determinism to inpaint. * relax tolerance for inpaiting. * determinism for inpainting legacy * relax tolerance. * determinism to instruct pix2pix * determinism to model editing. * model editing tolerance. * panorama determinism * determinism to pix2pix zero. * determinism to sag. * sd 2. determinism * sd. tolerance * disallow tf32 matmul. * relax tolerance is all you need. * make style and determinism to sd 2 depth * relax tolerance for depth. * tolerance to diffedit. * tolerance to sd 2 inpaint. * up tolerance. * determinism in upscaling. * tolerance in upscaler. * more tolerance relaxation. * determinism to v pred. * up tol for v_pred * unclip determinism * determinism to unclip img2img * determinism to text to video. * determinism to last set of tests * up tol. * vq cumsum doesn't have a deterministic kernel * relax tol * relax tol * [docs] Add transformers to install (#3388) add transformers to install * [deepspeed] partial ZeRO-3 support (#3076) * [deepspeed] partial ZeRO-3 support * cleanup * improve deepspeed fixes * Improve * make style --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Add omegaconf for tests (#3400) Add omegaconfg * Fix various bugs with LoRA Dreambooth and Dreambooth script (#3353) * Improve checkpointing lora * fix more * Improve doc string * Update src/diffusers/loaders.py * make stytle * Apply suggestions from code review * Update src/diffusers/loaders.py * Apply suggestions from code review * Apply suggestions from code review * better * Fix all * Fix multi-GPU dreambooth * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix all * make style * make style --------- Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix docker file (#3402) * up * up * fix: deepseepd_plugin retrieval from accelerate state (#3410) * [Docs] Add `sigmoid` beta_scheduler to docstrings of relevant Schedulers (#3399) * Add `sigmoid` beta scheduler to `DDPMScheduler` docstring * Add `sigmoid` beta scheduler to `RePaintScheduler` docstring --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Don't install accelerate and transformers from source (#3415) * Don't install transformers and accelerate from source (#3414) * Improve fast tests (#3416) Update pr_tests.yml * attention refactor: the trilogy (#3387) * Replace `AttentionBlock` with `Attention` * use _from_deprecated_attn_block check re: @patrickvonplaten * [Docs] update the PT 2.0 optimization doc with latest findings (#3370) * add: benchmarking stats for A100 and V100. * Apply suggestions from code review Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * address patrick's comments. * add: rtx 4090 stats * ⚔ benchmark reports done * Apply suggestions from code review Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * 3313 pr link. * add: plots. Co-authored-by: Pedro <pedro@huggingface.co> * fix formattimg * update number percent. --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * Fix style rendering (#3433) * Fix style rendering. * Fix typo * unCLIP scheduler do not use note (#3417) * Replace deprecated command with environment file (#3409) Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * fix warning message pipeline loading (#3446) * add stable diffusion tensorrt img2img pipeline (#3419) * add stable diffusion tensorrt img2img pipeline Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * update docstrings Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> --------- Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> * Refactor controlnet and add img2img and inpaint (#3386) * refactor controlnet and add img2img and inpaint * First draft to get pipelines to work * make style * Fix more * Fix more * More tests * Fix more * Make inpainting work * make style and more tests * Apply suggestions from code review * up * make style * Fix imports * Fix more * Fix more * Improve examples * add test * Make sure import is correctly deprecated * Make sure everything works in compile mode * make sure authorship is correctly attributed * [Scheduler] DPM-Solver (++) Inverse Scheduler (#3335) * Add DPM-Solver Multistep Inverse Scheduler * Add draft tests for DiffEdit * Add inverse sde-dpmsolver steps to tune image diversity from inverted latents * Fix tests --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * [Docs] Fix incomplete docstring for resnet.py (#3438) Fix incomplete docstrings for resnet.py * fix tiled vae blend extent range (#3384) fix tiled vae bleand extent range * Small update to "Next steps" section (#3443) Small update to "Next steps" section: - PyTorch 2 is recommended. - Updated improvement figures. * Allow arbitrary aspect ratio in IFSuperResolutionPipeline (#3298) * Update pipeline_if_superresolution.py Allow arbitrary aspect ratio in IFSuperResolutionPipeline by using the input image shape * IFSuperResolutionPipeline: allow the user to override the height and width through the arguments * update IFSuperResolutionPipeline width/height doc string to match StableDiffusionInpaintPipeline conventions --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Adding 'strength' parameter to StableDiffusionInpaintingPipeline (#3424) * Added explanation of 'strength' parameter * Added get_timesteps function which relies on new strength parameter * Added `strength` parameter which defaults to 1. * Swapped ordering so `noise_timestep` can be calculated before masking the image this is required when you aren't applying 100% noise to the masked region, e.g. strength < 1. * Added strength to check_inputs, throws error if out of range * Changed `prepare_latents` to initialise latents w.r.t strength inspired from the stable diffusion img2img pipeline, init latents are initialised by converting the init image into a VAE latent and adding noise (based upon the strength parameter passed in), e.g. random when strength = 1, or the init image at strength = 0. * WIP: Added a unit test for the new strength parameter in the StableDiffusionInpaintingPipeline still need to add correct regression values * Created a is_strength_max to initialise from pure random noise * Updated unit tests w.r.t new strength parameter + fixed new strength unit test * renamed parameter to avoid confusion with variable of same name * Updated regression values for new strength test - now passes * removed 'copied from' comment as this method is now different and divergent from the cpy * Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Ensure backwards compatibility for prepare_mask_and_masked_image created a return_image boolean and initialised to false * Ensure backwards compatibility for prepare_latents * Fixed copy check typo * Fixes w.r.t backward compibility changes * make style * keep function argument ordering same for backwards compatibility in callees with copied from statements * make fix-copies --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: William Berman <WLBberman@gmail.com> * [WIP] Bugfix - Pipeline.from_pretrained is broken when the pipeline is partially downloaded (#3448) Added bugfix using f strings. * Fix gradient checkpointing bugs in freezing part of models (requires_grad=False) (#3404) * gradient checkpointing bug fix * bug fix; changes for reviews * reformat * reformat --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Make dreambooth lora more robust to orig unet (#3462) * Make dreambooth lora more robust to orig unet * up * Reduce peak VRAM by releasing large attention tensors (as soon as they're unnecessary) (#3463) Release large tensors in attention (as soon as they're no longer required). Reduces peak VRAM by nearly 2 GB for 1024x1024 (even after slicing), and the savings scale up with image size. * Add min snr to text2img lora training script (#3459) add min snr to text2img lora training script * Add inpaint lora scale support (#3460) * add inpaint lora scale support * add inpaint lora scale test --------- Co-authored-by: yueyang.hyy <yueyang.hyy@alibaba-inc.com> * [From ckpt] Fix from_ckpt (#3466) * Correct from_ckpt * make style * Update full dreambooth script to work with IF (#3425) * Add IF dreambooth docs (#3470) * parameterize pass single args through tuple (#3477) * attend and excite tests disable determinism on the class level (#3478) * dreambooth docs torch.compile note (#3471) * dreambooth docs torch.compile note * Update examples/dreambooth/README.md Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * Update examples/dreambooth/README.md Co-authored-by: Pedro Cuenca <pedro@huggingface.co> --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> * add: if entry in the dreambooth training docs. (#3472) * [docs] Textual inversion inference (#3473) * add textual inversion inference to docs * add to toctree --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> * [docs] Distributed inference (#3376) * distributed inference * move to inference section * apply feedback * update with split_between_processes * apply feedback * [{Up,Down}sample1d] explicit view kernel size as number elements in flattened indices (#3479) explicit view kernel size as number elements in flattened indices * mps & onnx tests rework (#3449) * Remove ONNX tests from PR. They are already a part of push_tests.yml. * Remove mps tests from PRs. They are already performed on push. * Fix workflow name for fast push tests. * Extract mps tests to a workflow. For better control/filtering. * Remove --extra-index-url from mps tests * Increase tolerance of mps test This test passes in my Mac (Ventura 13.3) but fails in the CI hardware (Ventura 13.2). I ran the local tests following the same steps that exist in the CI workflow. * Temporarily run mps tests on pr So we can test. * Revert "Temporarily run mps tests on pr" Tests passed, go back to running on push. --------- Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> Co-authored-by: Ilia Larchenko <41329713+IliaLarchenko@users.noreply.github.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: YiYi Xu <yixu310@gmail.com> Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> Co-authored-by: Horace He <horacehe2007@yahoo.com> Co-authored-by: Umar <55330742+mu94-csl@users.noreply.github.com> Co-authored-by: Mylo <36931363+gitmylo@users.noreply.github.com> Co-authored-by: Markus Pobitzer <markuspobitzer@gmail.com> Co-authored-by: Cheng Lu <lucheng.lc15@gmail.com> Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> Co-authored-by: Isamu Isozaki <isamu.website@gmail.com> Co-authored-by: Cesar Aybar <csaybar@gmail.com> Co-authored-by: Will Rice <will@spokestack.io> Co-authored-by: Adrià Arrufat <1671644+arrufat@users.noreply.github.com> Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com> Co-authored-by: At-sushi <dkahw210@kyoto.zaq.ne.jp> Co-authored-by: Lucca Zenóbio <luccazen@gmail.com> Co-authored-by: Lysandre Debut <lysandre@huggingface.co> Co-authored-by: Isotr0py <41363108+Isotr0py@users.noreply.github.com> Co-authored-by: pdoane <pdoane2@gmail.com> Co-authored-by: Will Berman <wlbberman@gmail.com> Co-authored-by: yiyixuxu <yixu310@gmail,com> Co-authored-by: Rupert Menneer <71332436+rupertmenneer@users.noreply.github.com> Co-authored-by: sudowind <wfpkueecs@163.com> Co-authored-by: Takuma Mori <takuma104@gmail.com> Co-authored-by: Stas Bekman <stas00@users.noreply.github.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> Co-authored-by: Laureηt <laurentfainsin@protonmail.com> Co-authored-by: Jongwoo Han <jongwooo.han@gmail.com> Co-authored-by: asfiyab-nvidia <117682710+asfiyab-nvidia@users.noreply.github.com> Co-authored-by: clarencechen <clarencechenct@gmail.com> Co-authored-by: Laureηt <laurent@fainsin.bzh> Co-authored-by: superlabs-dev <133080491+superlabs-dev@users.noreply.github.com> Co-authored-by: Dev Aggarwal <devxpy@gmail.com> Co-authored-by: Vimarsh Chaturvedi <vimarsh.c@gmail.com> Co-authored-by: 7eu7d7 <31194890+7eu7d7@users.noreply.github.com> Co-authored-by: cmdr2 <shashank.shekhar.global@gmail.com> Co-authored-by: wfng92 <43742196+wfng92@users.noreply.github.com> Co-authored-by: Glaceon-Hyy <ffheyy0017@gmail.com> Co-authored-by: yueyang.hyy <yueyang.hyy@alibaba-inc.com> * [Community] reference only control (#3435) * add reference only control * add reference only control * add reference only control * fix lint * fix lint * reference adain * bugfix EulerAncestralDiscreteScheduler * fix style fidelity rule * fix default output size * del unused line * fix deterministic * Support for cross-attention bias / mask (#2634) * Cross-attention masks prefer qualified symbol, fix accidental Optional prefer qualified symbol in AttentionProcessor prefer qualified symbol in embeddings.py qualified symbol in transformed_2d qualify FloatTensor in unet_2d_blocks move new transformer_2d params attention_mask, encoder_attention_mask to the end of the section which is assumed (e.g. by functions such as checkpoint()) to have a stable positional param interface. regard return_dict as a special-case which is assumed to be injected separately from positional params (e.g. by create_custom_forward()). move new encoder_attention_mask param to end of CrossAttn block interfaces and Unet2DCondition interface, to maintain positional param interface. regenerate modeling_text_unet.py remove unused import unet_2d_condition encoder_attention_mask docs Co-authored-by: Pedro Cuenca <pedro@huggingface.co> versatile_diffusion/modeling_text_unet.py encoder_attention_mask docs Co-authored-by: Pedro Cuenca <pedro@huggingface.co> transformer_2d encoder_attention_mask docs Co-authored-by: Pedro Cuenca <pedro@huggingface.co> unet_2d_blocks.py: add parameter name comments Co-authored-by: Pedro Cuenca <pedro@huggingface.co> revert description. bool-to-bias treatment happens in unet_2d_condition only. comment parameter names fix copies, style * encoder_attention_mask for SimpleCrossAttnDownBlock2D, SimpleCrossAttnUpBlock2D * encoder_attention_mask for UNetMidBlock2DSimpleCrossAttn * support attention_mask, encoder_attention_mask in KCrossAttnDownBlock2D, KCrossAttnUpBlock2D, KAttentionBlock. fix binding of attention_mask, cross_attention_kwargs params in KCrossAttnDownBlock2D, KCrossAttnUpBlock2D checkpoint invocations. * fix mistake made during merge conflict resolution * regenerate versatile_diffusion * pass time embedding into checkpointed attention invocation * always assume encoder_attention_mask is a mask (i.e. not a bias). * style, fix-copies * add tests for cross-attention masks * add test for padding of attention mask * explain mask's query_tokens dim. fix explanation about broadcasting over channels; we actually broadcast over query tokens * support both masks and biases in Transformer2DModel#forward. document behaviour * fix-copies * delete attention_mask docs on the basis I never tested self-attention masking myself. not comfortable explaining it, since I don't actually understand how a self-attn mask can work in its current form: the key length will be different in every ResBlock (we don't downsample the mask when we downsample the image). * review feedback: the standard Unet blocks shouldn't pass temb to attn (only to resnet). remove from KCrossAttnDownBlock2D,KCrossAttnUpBlock2D#forward. * remove encoder_attention_mask param from SimpleCrossAttn{Up,Down}Block2D,UNetMidBlock2DSimpleCrossAttn, and mask-choice in those blocks' #forward, on the basis that they only do one type of attention, so the consumer can pass whichever type of attention_mask is appropriate. * put attention mask padding back to how it was (since the SD use-case it enabled wasn't important, and it breaks the original unclip use-case). disable the test which was added. * fix-copies * style * fix-copies * put encoder_attention_mask param back into Simple block forward interfaces, to ensure consistency of forward interface. * restore passing of emb to KAttentionBlock#forward, on the basis that removal caused test failures. restore also the passing of emb to checkpointed calls to KAttentionBlock#forward. * make simple unet2d blocks use encoder_attention_mask, but only when attention_mask is None. this should fix UnCLIP compatibility. * fix copies * do not scale the initial global step by gradient accumulation steps when loading from checkpoint (#3506) * Remove CPU latents logic for UniDiffuserPipelineFastTests. * make style * Revert "Clean up code and make slow tests pass." This reverts commitec7fb8735b. * Revert bad commit and clean up code. * add: contributor note. * Batched load of textual inversions (#3277) * Batched load of textual inversions - Only call resize_token_embeddings once per batch as it is the most expensive operation - Allow pretrained_model_name_or_path and token to be an optional list - Remove Dict from type annotation pretrained_model_name_or_path as it was not supported in this function - Add comment that single files (e.g. .pt/.safetensors) are supported - Add comment for token parameter - Convert token override log message from warning to info * Update src/diffusers/loaders.py Check for duplicate tokens Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update condition for None tokens --------- Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Revert "add: contributor note." This reverts commit302fde9409. * Re-add contributor note and refactored fast tests fixed latents code to remove CPU specific logic. * make style * Refactored the code: - Updated the checkpoint ids to the new ids where appropriate - Refactored the UniDiffuserTextDecoder methods to return only tensors (and made other changes to support this) - Cleaned up the code following suggestions by patrickvonplaten * make style * Remove padding logic from UniDiffuserTextDecoder.generate_beam since the inputs are already padded to a consistent length. * Update checkpoint id for small test v1 checkpoint to hf-internal-testing/unidiffuser-test-v1. * make style * Make improvements to the documentation. * Move ImageTextPipelineOutput documentation from /api/pipelines/unidiffuser.mdx to /api/diffusion_pipeline.mdx. * Change order of arguments for UniDiffuserTextDecoder.generate_beam. * make style * Update docs/source/en/api/pipelines/unidiffuser.mdx --------- Signed-off-by: Asfiya Baig <asfiyab@nvidia.com> Signed-off-by: Ye, Xinyu <xinyu.ye@intel.com> Co-authored-by: Ernie Chu <51432514+ernestchu@users.noreply.github.com> Co-authored-by: Sayak Paul <spsayakpaul@gmail.com> Co-authored-by: Andranik Movsisyan <48154088+19and99@users.noreply.github.com> Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Andreas Steiner <andstein@google.com> Co-authored-by: YiYi Xu <yixu310@gmail.com> Co-authored-by: Pedro Cuenca <pedro@huggingface.co> Co-authored-by: Joseph Coffland <github@joe.coffland.com> Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com> Co-authored-by: Takuma Mori <takuma104@gmail.com> Co-authored-by: Will Berman <wlbberman@gmail.com> Co-authored-by: Tommaso De Rossi <beats.by.morse@gmail.com> Co-authored-by: Cristian Garcia <cgarcia.e88@gmail.com> Co-authored-by: cmdr2 <secondary.cmdr2@gmail.com> Co-authored-by: 1lint <105617163+1lint@users.noreply.github.com> Co-authored-by: asfiyab-nvidia <117682710+asfiyab-nvidia@users.noreply.github.com> Co-authored-by: Chanchana Sornsoontorn <off9955555@gmail.com> Co-authored-by: hwuebben <wbben123@yahoo.de> Co-authored-by: superhero-7 <57797766+superhero-7@users.noreply.github.com> Co-authored-by: root <fulong_ye@163.com> Co-authored-by: nupurkmr9 <nupurkmr9@gmail.com> Co-authored-by: Nupur Kumari <nupurkumari@Nupurs-MacBook-Pro.local> Co-authored-by: Nupur Kumari <nupurkumari@nupurs-mbp.wifi.local.cmu.edu> Co-authored-by: Mishig <mishig.davaadorj@coloradocollege.edu> Co-authored-by: XinyuYe-Intel <xinyu.ye@intel.com> Co-authored-by: clarencechen <clarencechenct@gmail.com> Co-authored-by: regisss <15324346+regisss@users.noreply.github.com> Co-authored-by: Suraj Patil <surajp815@gmail.com> Co-authored-by: Youssef Adarrab <104783077+youssefadr@users.noreply.github.com> Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com> Co-authored-by: Chengrui Wang <80876977+crywang@users.noreply.github.com> Co-authored-by: SkyTNT <SKYTNT@outlook.com> Co-authored-by: Lucca Zenóbio <luccazen@gmail.com> Co-authored-by: Isaac <34376531+init-22@users.noreply.github.com> Co-authored-by: pdoane <pdoane2@gmail.com> Co-authored-by: Yuchen Fan <fyc0624@gmail.com> Co-authored-by: Nipun Jindal <jindal.nipun@gmail.com> Co-authored-by: njindal <njindal@adobe.com> Co-authored-by: apolinário <joaopaulo.passos@gmail.com> Co-authored-by: multimodalart <joaopaulo.passos+multimodal@gmail.com> Co-authored-by: Xie Zejian <xiezej@gmail.com> Co-authored-by: Jair Trejo <jairtrejo@gmail.com> Co-authored-by: Robert Dargavel Smith <teticio@gmail.com> Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan> Co-authored-by: Joqsan <6027118+Joqsan@users.noreply.github.com> Co-authored-by: NimenDavid <312648004@qq.com> Co-authored-by: M. Tolga Cangöz <46008593+standardAI@users.noreply.github.com> Co-authored-by: timegate <timegate@kaist.ac.kr> Co-authored-by: Jason Kuan <jason9075@users.noreply.github.com> Co-authored-by: Ilia Larchenko <41329713+IliaLarchenko@users.noreply.github.com> Co-authored-by: Horace He <horacehe2007@yahoo.com> Co-authored-by: Umar <55330742+mu94-csl@users.noreply.github.com> Co-authored-by: Mylo <36931363+gitmylo@users.noreply.github.com> Co-authored-by: Markus Pobitzer <markuspobitzer@gmail.com> Co-authored-by: Cheng Lu <lucheng.lc15@gmail.com> Co-authored-by: Isamu Isozaki <isamu.website@gmail.com> Co-authored-by: Cesar Aybar <csaybar@gmail.com> Co-authored-by: Will Rice <will@spokestack.io> Co-authored-by: yiyixuxu <yixu310@gmail,com> Co-authored-by: Rupert Menneer <71332436+rupertmenneer@users.noreply.github.com> Co-authored-by: sudowind <wfpkueecs@163.com> Co-authored-by: Stas Bekman <stas00@users.noreply.github.com> Co-authored-by: Laureηt <laurentfainsin@protonmail.com> Co-authored-by: Jongwoo Han <jongwooo.han@gmail.com> Co-authored-by: Laureηt <laurent@fainsin.bzh> Co-authored-by: superlabs-dev <133080491+superlabs-dev@users.noreply.github.com> Co-authored-by: Dev Aggarwal <devxpy@gmail.com> Co-authored-by: Vimarsh Chaturvedi <vimarsh.c@gmail.com> Co-authored-by: 7eu7d7 <31194890+7eu7d7@users.noreply.github.com> Co-authored-by: cmdr2 <shashank.shekhar.global@gmail.com> Co-authored-by: wfng92 <43742196+wfng92@users.noreply.github.com> Co-authored-by: Glaceon-Hyy <ffheyy0017@gmail.com> Co-authored-by: yueyang.hyy <yueyang.hyy@alibaba-inc.com> Co-authored-by: StAlKeR7779 <stalkek7779@yandex.ru> Co-authored-by: Isotr0py <41363108+Isotr0py@users.noreply.github.com> Co-authored-by: w4ffl35 <w4ffl35@ml1.net> Co-authored-by: Seongsu Park <tjdtnsu@gmail.com> Co-authored-by: Chanran Kim <seriousran@gmail.com> Co-authored-by: Ambrosiussen <paul@ambrosiussen.com> Co-authored-by: Hari Krishna <37787894+hari10599@users.noreply.github.com> Co-authored-by: Adrià Arrufat <1671644+arrufat@users.noreply.github.com> Co-authored-by: At-sushi <dkahw210@kyoto.zaq.ne.jp> Co-authored-by: Lysandre Debut <lysandre@huggingface.co> Co-authored-by: takuoko <to78314910@gmail.com> Co-authored-by: Birch-san <Birch-san@users.noreply.github.com>
This commit is contained in:
dg845
GitHub
@@ -232,6 +232,8 @@
|
||||
title: UnCLIP
|
||||
- local: api/pipelines/latent_diffusion_uncond
|
||||
title: Unconditional Latent Diffusion
|
||||
- local: api/pipelines/unidiffuser
|
||||
title: UniDiffuser
|
||||
- local: api/pipelines/versatile_diffusion
|
||||
title: Versatile Diffusion
|
||||
- local: api/pipelines/vq_diffusion
|
||||
|
||||
@@ -45,3 +45,8 @@ By default diffusion pipelines return an object of class
|
||||
By default diffusion pipelines return an object of class
|
||||
|
||||
[[autodoc]] pipelines.AudioPipelineOutput
|
||||
|
||||
## ImageTextPipelineOutput
|
||||
By default diffusion pipelines return an object of class
|
||||
|
||||
[[autodoc]] ImageTextPipelineOutput
|
||||
|
||||
204
docs/source/en/api/pipelines/unidiffuser.mdx
Normal file
204
docs/source/en/api/pipelines/unidiffuser.mdx
Normal file
@@ -0,0 +1,204 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
-->
|
||||
|
||||
# UniDiffuser
|
||||
|
||||
The UniDiffuser model was proposed in [One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale](https://arxiv.org/abs/2303.06555) by Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, Jun Zhu.
|
||||
|
||||
The abstract of the [paper](https://arxiv.org/abs/2303.06555) is the following:
|
||||
|
||||
*This paper proposes a unified diffusion framework (dubbed UniDiffuser) to fit all distributions relevant to a set of multi-modal data in one model. Our key insight is -- learning diffusion models for marginal, conditional, and joint distributions can be unified as predicting the noise in the perturbed data, where the perturbation levels (i.e. timesteps) can be different for different modalities. Inspired by the unified view, UniDiffuser learns all distributions simultaneously with a minimal modification to the original diffusion model -- perturbs data in all modalities instead of a single modality, inputs individual timesteps in different modalities, and predicts the noise of all modalities instead of a single modality. UniDiffuser is parameterized by a transformer for diffusion models to handle input types of different modalities. Implemented on large-scale paired image-text data, UniDiffuser is able to perform image, text, text-to-image, image-to-text, and image-text pair generation by setting proper timesteps without additional overhead. In particular, UniDiffuser is able to produce perceptually realistic samples in all tasks and its quantitative results (e.g., the FID and CLIP score) are not only superior to existing general-purpose models but also comparable to the bespoken models (e.g., Stable Diffusion and DALL-E 2) in representative tasks (e.g., text-to-image generation).*
|
||||
|
||||
Resources:
|
||||
|
||||
* [Paper](https://arxiv.org/abs/2303.06555).
|
||||
* [Original Code](https://github.com/thu-ml/unidiffuser).
|
||||
|
||||
Available Checkpoints are:
|
||||
- *UniDiffuser-v0 (512x512 resolution)* [thu-ml/unidiffuser-v0](https://huggingface.co/thu-ml/unidiffuser-v0)
|
||||
- *UniDiffuser-v1 (512x512 resolution)* [thu-ml/unidiffuser-v1](https://huggingface.co/thu-ml/unidiffuser-v1)
|
||||
|
||||
This pipeline was contributed by our community member [dg845](https://github.com/dg845).
|
||||
|
||||
## Available Pipelines:
|
||||
|
||||
| Pipeline | Tasks | Demo | Colab |
|
||||
|:---:|:---:|:---:|:---:|
|
||||
| [UniDiffuserPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_unidiffuser.py) | *Joint Image-Text Gen*, *Text-to-Image*, *Image-to-Text*,<br> *Image Gen*, *Text Gen*, *Image Variation*, *Text Variation* | [🤗 Spaces](https://huggingface.co/spaces/thu-ml/unidiffuser) | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/unidiffuser.ipynb) |
|
||||
|
||||
## Usage Examples
|
||||
|
||||
Because the UniDiffuser model is trained to model the joint distribution of (image, text) pairs, it is capable of performing a diverse range of generation tasks.
|
||||
|
||||
### Unconditional Image and Text Generation
|
||||
|
||||
Unconditional generation (where we start from only latents sampled from a standard Gaussian prior) from a [`UniDiffuserPipeline`] will produce a (image, text) pair:
|
||||
|
||||
```python
|
||||
import torch
|
||||
|
||||
from diffusers import UniDiffuserPipeline
|
||||
|
||||
device = "cuda"
|
||||
model_id_or_path = "thu-ml/unidiffuser-v1"
|
||||
pipe = UniDiffuserPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
|
||||
pipe.to(device)
|
||||
|
||||
# Unconditional image and text generation. The generation task is automatically inferred.
|
||||
sample = pipe(num_inference_steps=20, guidance_scale=8.0)
|
||||
image = sample.images[0]
|
||||
text = sample.text[0]
|
||||
image.save("unidiffuser_joint_sample_image.png")
|
||||
print(text)
|
||||
```
|
||||
|
||||
This is also called "joint" generation in the UniDiffusers paper, since we are sampling from the joint image-text distribution.
|
||||
|
||||
Note that the generation task is inferred from the inputs used when calling the pipeline.
|
||||
It is also possible to manually specify the unconditional generation task ("mode") manually with [`UniDiffuserPipeline.set_joint_mode`]:
|
||||
|
||||
```python
|
||||
# Equivalent to the above.
|
||||
pipe.set_joint_mode()
|
||||
sample = pipe(num_inference_steps=20, guidance_scale=8.0)
|
||||
```
|
||||
|
||||
When the mode is set manually, subsequent calls to the pipeline will use the set mode without attempting the infer the mode.
|
||||
You can reset the mode with [`UniDiffuserPipeline.reset_mode`], after which the pipeline will once again infer the mode.
|
||||
|
||||
You can also generate only an image or only text (which the UniDiffuser paper calls "marginal" generation since we sample from the marginal distribution of images and text, respectively):
|
||||
|
||||
```python
|
||||
# Unlike other generation tasks, image-only and text-only generation don't use classifier-free guidance
|
||||
# Image-only generation
|
||||
pipe.set_image_mode()
|
||||
sample_image = pipe(num_inference_steps=20).images[0]
|
||||
# Text-only generation
|
||||
pipe.set_text_mode()
|
||||
sample_text = pipe(num_inference_steps=20).text[0]
|
||||
```
|
||||
|
||||
### Text-to-Image Generation
|
||||
|
||||
UniDiffuser is also capable of sampling from conditional distributions; that is, the distribution of images conditioned on a text prompt or the distribution of texts conditioned on an image.
|
||||
Here is an example of sampling from the conditional image distribution (text-to-image generation or text-conditioned image generation):
|
||||
|
||||
```python
|
||||
import torch
|
||||
|
||||
from diffusers import UniDiffuserPipeline
|
||||
|
||||
device = "cuda"
|
||||
model_id_or_path = "thu-ml/unidiffuser-v1"
|
||||
pipe = UniDiffuserPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
|
||||
pipe.to(device)
|
||||
|
||||
# Text-to-image generation
|
||||
prompt = "an elephant under the sea"
|
||||
|
||||
sample = pipe(prompt=prompt, num_inference_steps=20, guidance_scale=8.0)
|
||||
t2i_image = sample.images[0]
|
||||
t2i_image.save("unidiffuser_text2img_sample_image.png")
|
||||
```
|
||||
|
||||
The `text2img` mode requires that either an input `prompt` or `prompt_embeds` be supplied. You can set the `text2img` mode manually with [`UniDiffuserPipeline.set_text_to_image_mode`].
|
||||
|
||||
### Image-to-Text Generation
|
||||
|
||||
Similarly, UniDiffuser can also produce text samples given an image (image-to-text or image-conditioned text generation):
|
||||
|
||||
```python
|
||||
import torch
|
||||
|
||||
from diffusers import UniDiffuserPipeline
|
||||
from diffusers.utils import load_image
|
||||
|
||||
device = "cuda"
|
||||
model_id_or_path = "thu-ml/unidiffuser-v1"
|
||||
pipe = UniDiffuserPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
|
||||
pipe.to(device)
|
||||
|
||||
# Image-to-text generation
|
||||
image_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/unidiffuser/unidiffuser_example_image.jpg"
|
||||
init_image = load_image(image_url).resize((512, 512))
|
||||
|
||||
sample = pipe(image=init_image, num_inference_steps=20, guidance_scale=8.0)
|
||||
i2t_text = sample.text[0]
|
||||
print(i2t_text)
|
||||
```
|
||||
|
||||
The `img2text` mode requires that an input `image` be supplied. You can set the `img2text` mode manually with [`UniDiffuserPipeline.set_image_to_text_mode`].
|
||||
|
||||
### Image Variation
|
||||
|
||||
The UniDiffuser authors suggest performing image variation through a "round-trip" generation method, where given an input image, we first perform an image-to-text generation, and the perform a text-to-image generation on the outputs of the first generation.
|
||||
This produces a new image which is semantically similar to the input image:
|
||||
|
||||
```python
|
||||
import torch
|
||||
|
||||
from diffusers import UniDiffuserPipeline
|
||||
from diffusers.utils import load_image
|
||||
|
||||
device = "cuda"
|
||||
model_id_or_path = "thu-ml/unidiffuser-v1"
|
||||
pipe = UniDiffuserPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
|
||||
pipe.to(device)
|
||||
|
||||
# Image variation can be performed with a image-to-text generation followed by a text-to-image generation:
|
||||
# 1. Image-to-text generation
|
||||
image_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/unidiffuser/unidiffuser_example_image.jpg"
|
||||
init_image = load_image(image_url).resize((512, 512))
|
||||
|
||||
sample = pipe(image=init_image, num_inference_steps=20, guidance_scale=8.0)
|
||||
i2t_text = sample.text[0]
|
||||
print(i2t_text)
|
||||
|
||||
# 2. Text-to-image generation
|
||||
sample = pipe(prompt=i2t_text, num_inference_steps=20, guidance_scale=8.0)
|
||||
final_image = sample.images[0]
|
||||
final_image.save("unidiffuser_image_variation_sample.png")
|
||||
```
|
||||
|
||||
### Text Variation
|
||||
|
||||
|
||||
Similarly, text variation can be performed on an input prompt with a text-to-image generation followed by a image-to-text generation:
|
||||
|
||||
```python
|
||||
import torch
|
||||
|
||||
from diffusers import UniDiffuserPipeline
|
||||
|
||||
device = "cuda"
|
||||
model_id_or_path = "thu-ml/unidiffuser-v1"
|
||||
pipe = UniDiffuserPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
|
||||
pipe.to(device)
|
||||
|
||||
# Text variation can be performed with a text-to-image generation followed by a image-to-text generation:
|
||||
# 1. Text-to-image generation
|
||||
prompt = "an elephant under the sea"
|
||||
|
||||
sample = pipe(prompt=prompt, num_inference_steps=20, guidance_scale=8.0)
|
||||
t2i_image = sample.images[0]
|
||||
t2i_image.save("unidiffuser_text2img_sample_image.png")
|
||||
|
||||
# 2. Image-to-text generation
|
||||
sample = pipe(image=t2i_image, num_inference_steps=20, guidance_scale=8.0)
|
||||
final_prompt = sample.text[0]
|
||||
print(final_prompt)
|
||||
```
|
||||
|
||||
## UniDiffuserPipeline
|
||||
[[autodoc]] UniDiffuserPipeline
|
||||
- all
|
||||
- __call__
|
||||
776
scripts/convert_unidiffuser_to_diffusers.py
Normal file
776
scripts/convert_unidiffuser_to_diffusers.py
Normal file
@@ -0,0 +1,776 @@
|
||||
# Convert the original UniDiffuser checkpoints into diffusers equivalents.
|
||||
|
||||
import argparse
|
||||
from argparse import Namespace
|
||||
|
||||
import torch
|
||||
from transformers import (
|
||||
CLIPImageProcessor,
|
||||
CLIPTextConfig,
|
||||
CLIPTextModel,
|
||||
CLIPTokenizer,
|
||||
CLIPVisionConfig,
|
||||
CLIPVisionModelWithProjection,
|
||||
GPT2Tokenizer,
|
||||
)
|
||||
|
||||
from diffusers import (
|
||||
AutoencoderKL,
|
||||
DPMSolverMultistepScheduler,
|
||||
UniDiffuserModel,
|
||||
UniDiffuserPipeline,
|
||||
UniDiffuserTextDecoder,
|
||||
)
|
||||
|
||||
|
||||
SCHEDULER_CONFIG = Namespace(
|
||||
**{
|
||||
"beta_start": 0.00085,
|
||||
"beta_end": 0.012,
|
||||
"beta_schedule": "scaled_linear",
|
||||
"solver_order": 3,
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.shave_segments
|
||||
def shave_segments(path, n_shave_prefix_segments=1):
|
||||
"""
|
||||
Removes segments. Positive values shave the first segments, negative shave the last segments.
|
||||
"""
|
||||
if n_shave_prefix_segments >= 0:
|
||||
return ".".join(path.split(".")[n_shave_prefix_segments:])
|
||||
else:
|
||||
return ".".join(path.split(".")[:n_shave_prefix_segments])
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_vae_resnet_paths
|
||||
def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
|
||||
"""
|
||||
Updates paths inside resnets to the new naming scheme (local renaming)
|
||||
"""
|
||||
mapping = []
|
||||
for old_item in old_list:
|
||||
new_item = old_item
|
||||
|
||||
new_item = new_item.replace("nin_shortcut", "conv_shortcut")
|
||||
new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
|
||||
|
||||
mapping.append({"old": old_item, "new": new_item})
|
||||
|
||||
return mapping
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.renew_vae_attention_paths
|
||||
def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
|
||||
"""
|
||||
Updates paths inside attentions to the new naming scheme (local renaming)
|
||||
"""
|
||||
mapping = []
|
||||
for old_item in old_list:
|
||||
new_item = old_item
|
||||
|
||||
new_item = new_item.replace("norm.weight", "group_norm.weight")
|
||||
new_item = new_item.replace("norm.bias", "group_norm.bias")
|
||||
|
||||
new_item = new_item.replace("q.weight", "query.weight")
|
||||
new_item = new_item.replace("q.bias", "query.bias")
|
||||
|
||||
new_item = new_item.replace("k.weight", "key.weight")
|
||||
new_item = new_item.replace("k.bias", "key.bias")
|
||||
|
||||
new_item = new_item.replace("v.weight", "value.weight")
|
||||
new_item = new_item.replace("v.bias", "value.bias")
|
||||
|
||||
new_item = new_item.replace("proj_out.weight", "proj_attn.weight")
|
||||
new_item = new_item.replace("proj_out.bias", "proj_attn.bias")
|
||||
|
||||
new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
|
||||
|
||||
mapping.append({"old": old_item, "new": new_item})
|
||||
|
||||
return mapping
|
||||
|
||||
|
||||
# Modified from diffusers.pipelines.stable_diffusion.convert_from_ckpt.assign_to_checkpoint
|
||||
# config.num_head_channels => num_head_channels
|
||||
def assign_to_checkpoint(
|
||||
paths,
|
||||
checkpoint,
|
||||
old_checkpoint,
|
||||
attention_paths_to_split=None,
|
||||
additional_replacements=None,
|
||||
num_head_channels=1,
|
||||
):
|
||||
"""
|
||||
This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
|
||||
attention layers, and takes into account additional replacements that may arise. Assigns the weights to the new
|
||||
checkpoint.
|
||||
"""
|
||||
assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
|
||||
|
||||
# Splits the attention layers into three variables.
|
||||
if attention_paths_to_split is not None:
|
||||
for path, path_map in attention_paths_to_split.items():
|
||||
old_tensor = old_checkpoint[path]
|
||||
channels = old_tensor.shape[0] // 3
|
||||
|
||||
target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
|
||||
|
||||
num_heads = old_tensor.shape[0] // num_head_channels // 3
|
||||
|
||||
old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
|
||||
query, key, value = old_tensor.split(channels // num_heads, dim=1)
|
||||
|
||||
checkpoint[path_map["query"]] = query.reshape(target_shape)
|
||||
checkpoint[path_map["key"]] = key.reshape(target_shape)
|
||||
checkpoint[path_map["value"]] = value.reshape(target_shape)
|
||||
|
||||
for path in paths:
|
||||
new_path = path["new"]
|
||||
|
||||
# These have already been assigned
|
||||
if attention_paths_to_split is not None and new_path in attention_paths_to_split:
|
||||
continue
|
||||
|
||||
# Global renaming happens here
|
||||
new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
|
||||
new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
|
||||
new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")
|
||||
|
||||
if additional_replacements is not None:
|
||||
for replacement in additional_replacements:
|
||||
new_path = new_path.replace(replacement["old"], replacement["new"])
|
||||
|
||||
# proj_attn.weight has to be converted from conv 1D to linear
|
||||
if "proj_attn.weight" in new_path:
|
||||
checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
|
||||
else:
|
||||
checkpoint[new_path] = old_checkpoint[path["old"]]
|
||||
|
||||
|
||||
# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.conv_attn_to_linear
|
||||
def conv_attn_to_linear(checkpoint):
|
||||
keys = list(checkpoint.keys())
|
||||
attn_keys = ["query.weight", "key.weight", "value.weight"]
|
||||
for key in keys:
|
||||
if ".".join(key.split(".")[-2:]) in attn_keys:
|
||||
if checkpoint[key].ndim > 2:
|
||||
checkpoint[key] = checkpoint[key][:, :, 0, 0]
|
||||
elif "proj_attn.weight" in key:
|
||||
if checkpoint[key].ndim > 2:
|
||||
checkpoint[key] = checkpoint[key][:, :, 0]
|
||||
|
||||
|
||||
def create_vae_diffusers_config(config_type):
|
||||
# Hardcoded for now
|
||||
if args.config_type == "test":
|
||||
vae_config = create_vae_diffusers_config_test()
|
||||
elif args.config_type == "big":
|
||||
vae_config = create_vae_diffusers_config_big()
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
f"Config type {config_type} is not implemented, currently only config types"
|
||||
" 'test' and 'big' are available."
|
||||
)
|
||||
return vae_config
|
||||
|
||||
|
||||
def create_unidiffuser_unet_config(config_type, version):
|
||||
# Hardcoded for now
|
||||
if args.config_type == "test":
|
||||
unet_config = create_unidiffuser_unet_config_test()
|
||||
elif args.config_type == "big":
|
||||
unet_config = create_unidiffuser_unet_config_big()
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
f"Config type {config_type} is not implemented, currently only config types"
|
||||
" 'test' and 'big' are available."
|
||||
)
|
||||
# Unidiffuser-v1 uses data type embeddings
|
||||
if version == 1:
|
||||
unet_config["use_data_type_embedding"] = True
|
||||
return unet_config
|
||||
|
||||
|
||||
def create_text_decoder_config(config_type):
|
||||
# Hardcoded for now
|
||||
if args.config_type == "test":
|
||||
text_decoder_config = create_text_decoder_config_test()
|
||||
elif args.config_type == "big":
|
||||
text_decoder_config = create_text_decoder_config_big()
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
f"Config type {config_type} is not implemented, currently only config types"
|
||||
" 'test' and 'big' are available."
|
||||
)
|
||||
return text_decoder_config
|
||||
|
||||
|
||||
# Hardcoded configs for test versions of the UniDiffuser models, corresponding to those in the fast default tests.
|
||||
def create_vae_diffusers_config_test():
|
||||
vae_config = {
|
||||
"sample_size": 32,
|
||||
"in_channels": 3,
|
||||
"out_channels": 3,
|
||||
"down_block_types": ["DownEncoderBlock2D", "DownEncoderBlock2D"],
|
||||
"up_block_types": ["UpDecoderBlock2D", "UpDecoderBlock2D"],
|
||||
"block_out_channels": [32, 64],
|
||||
"latent_channels": 4,
|
||||
"layers_per_block": 1,
|
||||
}
|
||||
return vae_config
|
||||
|
||||
|
||||
def create_unidiffuser_unet_config_test():
|
||||
unet_config = {
|
||||
"text_dim": 32,
|
||||
"clip_img_dim": 32,
|
||||
"num_text_tokens": 77,
|
||||
"num_attention_heads": 2,
|
||||
"attention_head_dim": 8,
|
||||
"in_channels": 4,
|
||||
"out_channels": 4,
|
||||
"num_layers": 2,
|
||||
"dropout": 0.0,
|
||||
"norm_num_groups": 32,
|
||||
"attention_bias": False,
|
||||
"sample_size": 16,
|
||||
"patch_size": 2,
|
||||
"activation_fn": "gelu",
|
||||
"num_embeds_ada_norm": 1000,
|
||||
"norm_type": "layer_norm",
|
||||
"block_type": "unidiffuser",
|
||||
"pre_layer_norm": False,
|
||||
"use_timestep_embedding": False,
|
||||
"norm_elementwise_affine": True,
|
||||
"use_patch_pos_embed": False,
|
||||
"ff_final_dropout": True,
|
||||
"use_data_type_embedding": False,
|
||||
}
|
||||
return unet_config
|
||||
|
||||
|
||||
def create_text_decoder_config_test():
|
||||
text_decoder_config = {
|
||||
"prefix_length": 77,
|
||||
"prefix_inner_dim": 32,
|
||||
"prefix_hidden_dim": 32,
|
||||
"vocab_size": 1025, # 1024 + 1 for new EOS token
|
||||
"n_positions": 1024,
|
||||
"n_embd": 32,
|
||||
"n_layer": 5,
|
||||
"n_head": 4,
|
||||
"n_inner": 37,
|
||||
"activation_function": "gelu",
|
||||
"resid_pdrop": 0.1,
|
||||
"embd_pdrop": 0.1,
|
||||
"attn_pdrop": 0.1,
|
||||
"layer_norm_epsilon": 1e-5,
|
||||
"initializer_range": 0.02,
|
||||
}
|
||||
return text_decoder_config
|
||||
|
||||
|
||||
# Hardcoded configs for the UniDiffuser V1 model at https://huggingface.co/thu-ml/unidiffuser-v1
|
||||
# See also https://github.com/thu-ml/unidiffuser/blob/main/configs/sample_unidiffuser_v1.py
|
||||
def create_vae_diffusers_config_big():
|
||||
vae_config = {
|
||||
"sample_size": 256,
|
||||
"in_channels": 3,
|
||||
"out_channels": 3,
|
||||
"down_block_types": ["DownEncoderBlock2D", "DownEncoderBlock2D", "DownEncoderBlock2D", "DownEncoderBlock2D"],
|
||||
"up_block_types": ["UpDecoderBlock2D", "UpDecoderBlock2D", "UpDecoderBlock2D", "UpDecoderBlock2D"],
|
||||
"block_out_channels": [128, 256, 512, 512],
|
||||
"latent_channels": 4,
|
||||
"layers_per_block": 2,
|
||||
}
|
||||
return vae_config
|
||||
|
||||
|
||||
def create_unidiffuser_unet_config_big():
|
||||
unet_config = {
|
||||
"text_dim": 64,
|
||||
"clip_img_dim": 512,
|
||||
"num_text_tokens": 77,
|
||||
"num_attention_heads": 24,
|
||||
"attention_head_dim": 64,
|
||||
"in_channels": 4,
|
||||
"out_channels": 4,
|
||||
"num_layers": 30,
|
||||
"dropout": 0.0,
|
||||
"norm_num_groups": 32,
|
||||
"attention_bias": False,
|
||||
"sample_size": 64,
|
||||
"patch_size": 2,
|
||||
"activation_fn": "gelu",
|
||||
"num_embeds_ada_norm": 1000,
|
||||
"norm_type": "layer_norm",
|
||||
"block_type": "unidiffuser",
|
||||
"pre_layer_norm": False,
|
||||
"use_timestep_embedding": False,
|
||||
"norm_elementwise_affine": True,
|
||||
"use_patch_pos_embed": False,
|
||||
"ff_final_dropout": True,
|
||||
"use_data_type_embedding": False,
|
||||
}
|
||||
return unet_config
|
||||
|
||||
|
||||
# From https://huggingface.co/gpt2/blob/main/config.json, the GPT2 checkpoint used by UniDiffuser
|
||||
def create_text_decoder_config_big():
|
||||
text_decoder_config = {
|
||||
"prefix_length": 77,
|
||||
"prefix_inner_dim": 768,
|
||||
"prefix_hidden_dim": 64,
|
||||
"vocab_size": 50258, # 50257 + 1 for new EOS token
|
||||
"n_positions": 1024,
|
||||
"n_embd": 768,
|
||||
"n_layer": 12,
|
||||
"n_head": 12,
|
||||
"n_inner": 3072,
|
||||
"activation_function": "gelu",
|
||||
"resid_pdrop": 0.1,
|
||||
"embd_pdrop": 0.1,
|
||||
"attn_pdrop": 0.1,
|
||||
"layer_norm_epsilon": 1e-5,
|
||||
"initializer_range": 0.02,
|
||||
}
|
||||
return text_decoder_config
|
||||
|
||||
|
||||
# Based on diffusers.pipelines.stable_diffusion.convert_from_ckpt.shave_segments.convert_ldm_vae_checkpoint
|
||||
def convert_vae_to_diffusers(ckpt, diffusers_model, num_head_channels=1):
|
||||
"""
|
||||
Converts a UniDiffuser autoencoder_kl.pth checkpoint to a diffusers AutoencoderKL.
|
||||
"""
|
||||
# autoencoder_kl.pth ckpt is a torch state dict
|
||||
vae_state_dict = torch.load(ckpt, map_location="cpu")
|
||||
|
||||
new_checkpoint = {}
|
||||
|
||||
new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
|
||||
new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
|
||||
new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
|
||||
new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
|
||||
new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
|
||||
new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]
|
||||
|
||||
new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
|
||||
new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
|
||||
new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
|
||||
new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
|
||||
new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
|
||||
new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]
|
||||
|
||||
new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
|
||||
new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
|
||||
new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
|
||||
new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]
|
||||
|
||||
# Retrieves the keys for the encoder down blocks only
|
||||
num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
|
||||
down_blocks = {
|
||||
layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
|
||||
}
|
||||
|
||||
# Retrieves the keys for the decoder up blocks only
|
||||
num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
|
||||
up_blocks = {
|
||||
layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
|
||||
}
|
||||
|
||||
for i in range(num_down_blocks):
|
||||
resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
|
||||
|
||||
if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
|
||||
new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
|
||||
f"encoder.down.{i}.downsample.conv.weight"
|
||||
)
|
||||
new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
|
||||
f"encoder.down.{i}.downsample.conv.bias"
|
||||
)
|
||||
|
||||
paths = renew_vae_resnet_paths(resnets)
|
||||
meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
|
||||
mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
|
||||
num_mid_res_blocks = 2
|
||||
for i in range(1, num_mid_res_blocks + 1):
|
||||
resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
|
||||
|
||||
paths = renew_vae_resnet_paths(resnets)
|
||||
meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
|
||||
mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
|
||||
paths = renew_vae_attention_paths(mid_attentions)
|
||||
meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
conv_attn_to_linear(new_checkpoint)
|
||||
|
||||
for i in range(num_up_blocks):
|
||||
block_id = num_up_blocks - 1 - i
|
||||
resnets = [
|
||||
key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
|
||||
]
|
||||
|
||||
if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
|
||||
new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
|
||||
f"decoder.up.{block_id}.upsample.conv.weight"
|
||||
]
|
||||
new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
|
||||
f"decoder.up.{block_id}.upsample.conv.bias"
|
||||
]
|
||||
|
||||
paths = renew_vae_resnet_paths(resnets)
|
||||
meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
|
||||
mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
|
||||
num_mid_res_blocks = 2
|
||||
for i in range(1, num_mid_res_blocks + 1):
|
||||
resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
|
||||
|
||||
paths = renew_vae_resnet_paths(resnets)
|
||||
meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
|
||||
mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
|
||||
paths = renew_vae_attention_paths(mid_attentions)
|
||||
meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
|
||||
assign_to_checkpoint(
|
||||
paths,
|
||||
new_checkpoint,
|
||||
vae_state_dict,
|
||||
additional_replacements=[meta_path],
|
||||
num_head_channels=num_head_channels, # not used in vae
|
||||
)
|
||||
conv_attn_to_linear(new_checkpoint)
|
||||
|
||||
missing_keys, unexpected_keys = diffusers_model.load_state_dict(new_checkpoint)
|
||||
for missing_key in missing_keys:
|
||||
print(f"Missing key: {missing_key}")
|
||||
for unexpected_key in unexpected_keys:
|
||||
print(f"Unexpected key: {unexpected_key}")
|
||||
|
||||
return diffusers_model
|
||||
|
||||
|
||||
def convert_uvit_block_to_diffusers_block(
|
||||
uvit_state_dict,
|
||||
new_state_dict,
|
||||
block_prefix,
|
||||
new_prefix="transformer.transformer_",
|
||||
skip_connection=False,
|
||||
):
|
||||
"""
|
||||
Maps the keys in a UniDiffuser transformer block (`Block`) to the keys in a diffusers transformer block
|
||||
(`UTransformerBlock`/`UniDiffuserBlock`).
|
||||
"""
|
||||
prefix = new_prefix + block_prefix
|
||||
if skip_connection:
|
||||
new_state_dict[prefix + ".skip.skip_linear.weight"] = uvit_state_dict[block_prefix + ".skip_linear.weight"]
|
||||
new_state_dict[prefix + ".skip.skip_linear.bias"] = uvit_state_dict[block_prefix + ".skip_linear.bias"]
|
||||
new_state_dict[prefix + ".skip.norm.weight"] = uvit_state_dict[block_prefix + ".norm1.weight"]
|
||||
new_state_dict[prefix + ".skip.norm.bias"] = uvit_state_dict[block_prefix + ".norm1.bias"]
|
||||
|
||||
# Create the prefix string for out_blocks.
|
||||
prefix += ".block"
|
||||
|
||||
# Split up attention qkv.weight into to_q.weight, to_k.weight, to_v.weight
|
||||
qkv = uvit_state_dict[block_prefix + ".attn.qkv.weight"]
|
||||
new_attn_keys = [".attn1.to_q.weight", ".attn1.to_k.weight", ".attn1.to_v.weight"]
|
||||
new_attn_keys = [prefix + key for key in new_attn_keys]
|
||||
shape = qkv.shape[0] // len(new_attn_keys)
|
||||
for i, attn_key in enumerate(new_attn_keys):
|
||||
new_state_dict[attn_key] = qkv[i * shape : (i + 1) * shape]
|
||||
|
||||
new_state_dict[prefix + ".attn1.to_out.0.weight"] = uvit_state_dict[block_prefix + ".attn.proj.weight"]
|
||||
new_state_dict[prefix + ".attn1.to_out.0.bias"] = uvit_state_dict[block_prefix + ".attn.proj.bias"]
|
||||
new_state_dict[prefix + ".norm1.weight"] = uvit_state_dict[block_prefix + ".norm2.weight"]
|
||||
new_state_dict[prefix + ".norm1.bias"] = uvit_state_dict[block_prefix + ".norm2.bias"]
|
||||
new_state_dict[prefix + ".ff.net.0.proj.weight"] = uvit_state_dict[block_prefix + ".mlp.fc1.weight"]
|
||||
new_state_dict[prefix + ".ff.net.0.proj.bias"] = uvit_state_dict[block_prefix + ".mlp.fc1.bias"]
|
||||
new_state_dict[prefix + ".ff.net.2.weight"] = uvit_state_dict[block_prefix + ".mlp.fc2.weight"]
|
||||
new_state_dict[prefix + ".ff.net.2.bias"] = uvit_state_dict[block_prefix + ".mlp.fc2.bias"]
|
||||
new_state_dict[prefix + ".norm3.weight"] = uvit_state_dict[block_prefix + ".norm3.weight"]
|
||||
new_state_dict[prefix + ".norm3.bias"] = uvit_state_dict[block_prefix + ".norm3.bias"]
|
||||
|
||||
return uvit_state_dict, new_state_dict
|
||||
|
||||
|
||||
def convert_uvit_to_diffusers(ckpt, diffusers_model):
|
||||
"""
|
||||
Converts a UniDiffuser uvit_v*.pth checkpoint to a diffusers UniDiffusersModel.
|
||||
"""
|
||||
# uvit_v*.pth ckpt is a torch state dict
|
||||
uvit_state_dict = torch.load(ckpt, map_location="cpu")
|
||||
|
||||
new_state_dict = {}
|
||||
|
||||
# Input layers
|
||||
new_state_dict["vae_img_in.proj.weight"] = uvit_state_dict["patch_embed.proj.weight"]
|
||||
new_state_dict["vae_img_in.proj.bias"] = uvit_state_dict["patch_embed.proj.bias"]
|
||||
new_state_dict["clip_img_in.weight"] = uvit_state_dict["clip_img_embed.weight"]
|
||||
new_state_dict["clip_img_in.bias"] = uvit_state_dict["clip_img_embed.bias"]
|
||||
new_state_dict["text_in.weight"] = uvit_state_dict["text_embed.weight"]
|
||||
new_state_dict["text_in.bias"] = uvit_state_dict["text_embed.bias"]
|
||||
|
||||
new_state_dict["pos_embed"] = uvit_state_dict["pos_embed"]
|
||||
|
||||
# Handle data type token embeddings for UniDiffuser-v1
|
||||
if "token_embedding.weight" in uvit_state_dict and diffusers_model.use_data_type_embedding:
|
||||
new_state_dict["data_type_pos_embed_token"] = uvit_state_dict["pos_embed_token"]
|
||||
new_state_dict["data_type_token_embedding.weight"] = uvit_state_dict["token_embedding.weight"]
|
||||
|
||||
# Also initialize the PatchEmbedding in UTransformer2DModel with the PatchEmbedding from the checkpoint.
|
||||
# This isn't used in the current implementation, so might want to remove.
|
||||
new_state_dict["transformer.pos_embed.proj.weight"] = uvit_state_dict["patch_embed.proj.weight"]
|
||||
new_state_dict["transformer.pos_embed.proj.bias"] = uvit_state_dict["patch_embed.proj.bias"]
|
||||
|
||||
# Output layers
|
||||
new_state_dict["transformer.norm_out.weight"] = uvit_state_dict["norm.weight"]
|
||||
new_state_dict["transformer.norm_out.bias"] = uvit_state_dict["norm.bias"]
|
||||
|
||||
new_state_dict["vae_img_out.weight"] = uvit_state_dict["decoder_pred.weight"]
|
||||
new_state_dict["vae_img_out.bias"] = uvit_state_dict["decoder_pred.bias"]
|
||||
new_state_dict["clip_img_out.weight"] = uvit_state_dict["clip_img_out.weight"]
|
||||
new_state_dict["clip_img_out.bias"] = uvit_state_dict["clip_img_out.bias"]
|
||||
new_state_dict["text_out.weight"] = uvit_state_dict["text_out.weight"]
|
||||
new_state_dict["text_out.bias"] = uvit_state_dict["text_out.bias"]
|
||||
|
||||
# in_blocks
|
||||
in_blocks_prefixes = {".".join(layer.split(".")[:2]) for layer in uvit_state_dict if "in_blocks" in layer}
|
||||
for in_block_prefix in list(in_blocks_prefixes):
|
||||
convert_uvit_block_to_diffusers_block(uvit_state_dict, new_state_dict, in_block_prefix)
|
||||
|
||||
# mid_block
|
||||
# Assume there's only one mid block
|
||||
convert_uvit_block_to_diffusers_block(uvit_state_dict, new_state_dict, "mid_block")
|
||||
|
||||
# out_blocks
|
||||
out_blocks_prefixes = {".".join(layer.split(".")[:2]) for layer in uvit_state_dict if "out_blocks" in layer}
|
||||
for out_block_prefix in list(out_blocks_prefixes):
|
||||
convert_uvit_block_to_diffusers_block(uvit_state_dict, new_state_dict, out_block_prefix, skip_connection=True)
|
||||
|
||||
missing_keys, unexpected_keys = diffusers_model.load_state_dict(new_state_dict)
|
||||
for missing_key in missing_keys:
|
||||
print(f"Missing key: {missing_key}")
|
||||
for unexpected_key in unexpected_keys:
|
||||
print(f"Unexpected key: {unexpected_key}")
|
||||
|
||||
return diffusers_model
|
||||
|
||||
|
||||
def convert_caption_decoder_to_diffusers(ckpt, diffusers_model):
|
||||
"""
|
||||
Converts a UniDiffuser caption_decoder.pth checkpoint to a diffusers UniDiffuserTextDecoder.
|
||||
"""
|
||||
# caption_decoder.pth ckpt is a torch state dict
|
||||
checkpoint_state_dict = torch.load(ckpt, map_location="cpu")
|
||||
decoder_state_dict = {}
|
||||
# Remove the "module." prefix, if necessary
|
||||
caption_decoder_key = "module."
|
||||
for key in checkpoint_state_dict:
|
||||
if key.startswith(caption_decoder_key):
|
||||
decoder_state_dict[key.replace(caption_decoder_key, "")] = checkpoint_state_dict.get(key)
|
||||
else:
|
||||
decoder_state_dict[key] = checkpoint_state_dict.get(key)
|
||||
|
||||
new_state_dict = {}
|
||||
|
||||
# Encoder and Decoder
|
||||
new_state_dict["encode_prefix.weight"] = decoder_state_dict["encode_prefix.weight"]
|
||||
new_state_dict["encode_prefix.bias"] = decoder_state_dict["encode_prefix.bias"]
|
||||
new_state_dict["decode_prefix.weight"] = decoder_state_dict["decode_prefix.weight"]
|
||||
new_state_dict["decode_prefix.bias"] = decoder_state_dict["decode_prefix.bias"]
|
||||
|
||||
# Internal GPT2LMHeadModel transformer model
|
||||
for key, val in decoder_state_dict.items():
|
||||
if key.startswith("gpt"):
|
||||
suffix = key[len("gpt") :]
|
||||
new_state_dict["transformer" + suffix] = val
|
||||
|
||||
missing_keys, unexpected_keys = diffusers_model.load_state_dict(new_state_dict)
|
||||
for missing_key in missing_keys:
|
||||
print(f"Missing key: {missing_key}")
|
||||
for unexpected_key in unexpected_keys:
|
||||
print(f"Unexpected key: {unexpected_key}")
|
||||
|
||||
return diffusers_model
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
|
||||
parser.add_argument(
|
||||
"--caption_decoder_checkpoint_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=False,
|
||||
help="Path to caption decoder checkpoint to convert.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--uvit_checkpoint_path", default=None, type=str, required=False, help="Path to U-ViT checkpoint to convert."
|
||||
)
|
||||
parser.add_argument(
|
||||
"--vae_checkpoint_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=False,
|
||||
help="Path to VAE checkpoint to convert.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--pipeline_output_path",
|
||||
default=None,
|
||||
type=str,
|
||||
required=True,
|
||||
help="Path to save the output pipeline to.",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--config_type",
|
||||
default="test",
|
||||
type=str,
|
||||
help=(
|
||||
"Config type to use. Should be 'test' to create small models for testing or 'big' to convert a full"
|
||||
" checkpoint."
|
||||
),
|
||||
)
|
||||
parser.add_argument(
|
||||
"--version",
|
||||
default=0,
|
||||
type=int,
|
||||
help="The UniDiffuser model type to convert to. Should be 0 for UniDiffuser-v0 and 1 for UniDiffuser-v1.",
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
# Convert the VAE model.
|
||||
if args.vae_checkpoint_path is not None:
|
||||
vae_config = create_vae_diffusers_config(args.config_type)
|
||||
vae = AutoencoderKL(**vae_config)
|
||||
vae = convert_vae_to_diffusers(args.vae_checkpoint_path, vae)
|
||||
|
||||
# Convert the U-ViT ("unet") model.
|
||||
if args.uvit_checkpoint_path is not None:
|
||||
unet_config = create_unidiffuser_unet_config(args.config_type, args.version)
|
||||
unet = UniDiffuserModel(**unet_config)
|
||||
unet = convert_uvit_to_diffusers(args.uvit_checkpoint_path, unet)
|
||||
|
||||
# Convert the caption decoder ("text_decoder") model.
|
||||
if args.caption_decoder_checkpoint_path is not None:
|
||||
text_decoder_config = create_text_decoder_config(args.config_type)
|
||||
text_decoder = UniDiffuserTextDecoder(**text_decoder_config)
|
||||
text_decoder = convert_caption_decoder_to_diffusers(args.caption_decoder_checkpoint_path, text_decoder)
|
||||
|
||||
# Scheduler is the same for both the test and big models.
|
||||
scheduler_config = SCHEDULER_CONFIG
|
||||
scheduler = DPMSolverMultistepScheduler(
|
||||
beta_start=scheduler_config.beta_start,
|
||||
beta_end=scheduler_config.beta_end,
|
||||
beta_schedule=scheduler_config.beta_schedule,
|
||||
solver_order=scheduler_config.solver_order,
|
||||
)
|
||||
|
||||
if args.config_type == "test":
|
||||
# Make a small random CLIPTextModel
|
||||
torch.manual_seed(0)
|
||||
clip_text_encoder_config = CLIPTextConfig(
|
||||
bos_token_id=0,
|
||||
eos_token_id=2,
|
||||
hidden_size=32,
|
||||
intermediate_size=37,
|
||||
layer_norm_eps=1e-05,
|
||||
num_attention_heads=4,
|
||||
num_hidden_layers=5,
|
||||
pad_token_id=1,
|
||||
vocab_size=1000,
|
||||
)
|
||||
text_encoder = CLIPTextModel(clip_text_encoder_config)
|
||||
clip_tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
|
||||
|
||||
# Make a small random CLIPVisionModel and accompanying CLIPImageProcessor
|
||||
torch.manual_seed(0)
|
||||
clip_image_encoder_config = CLIPVisionConfig(
|
||||
image_size=32,
|
||||
patch_size=2,
|
||||
num_channels=3,
|
||||
hidden_size=32,
|
||||
projection_dim=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
dropout=0.1,
|
||||
attention_dropout=0.1,
|
||||
initializer_range=0.02,
|
||||
)
|
||||
image_encoder = CLIPVisionModelWithProjection(clip_image_encoder_config)
|
||||
image_processor = CLIPImageProcessor(crop_size=32, size=32)
|
||||
|
||||
# Note that the text_decoder should already have its token embeddings resized.
|
||||
text_tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-GPT2Model")
|
||||
eos = "<|EOS|>"
|
||||
special_tokens_dict = {"eos_token": eos}
|
||||
text_tokenizer.add_special_tokens(special_tokens_dict)
|
||||
elif args.config_type == "big":
|
||||
text_encoder = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14")
|
||||
clip_tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
|
||||
|
||||
image_encoder = CLIPVisionModelWithProjection.from_pretrained("openai/clip-vit-base-patch32")
|
||||
image_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
||||
|
||||
# Note that the text_decoder should already have its token embeddings resized.
|
||||
text_tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
||||
eos = "<|EOS|>"
|
||||
special_tokens_dict = {"eos_token": eos}
|
||||
text_tokenizer.add_special_tokens(special_tokens_dict)
|
||||
else:
|
||||
raise NotImplementedError(
|
||||
f"Config type {args.config_type} is not implemented, currently only config types"
|
||||
" 'test' and 'big' are available."
|
||||
)
|
||||
|
||||
pipeline = UniDiffuserPipeline(
|
||||
vae=vae,
|
||||
text_encoder=text_encoder,
|
||||
image_encoder=image_encoder,
|
||||
image_processor=image_processor,
|
||||
clip_tokenizer=clip_tokenizer,
|
||||
text_decoder=text_decoder,
|
||||
text_tokenizer=text_tokenizer,
|
||||
unet=unet,
|
||||
scheduler=scheduler,
|
||||
)
|
||||
pipeline.save_pretrained(args.pipeline_output_path)
|
||||
@@ -129,6 +129,7 @@ else:
|
||||
IFInpaintingSuperResolutionPipeline,
|
||||
IFPipeline,
|
||||
IFSuperResolutionPipeline,
|
||||
ImageTextPipelineOutput,
|
||||
KandinskyImg2ImgPipeline,
|
||||
KandinskyInpaintPipeline,
|
||||
KandinskyPipeline,
|
||||
@@ -161,6 +162,9 @@ else:
|
||||
TextToVideoZeroPipeline,
|
||||
UnCLIPImageVariationPipeline,
|
||||
UnCLIPPipeline,
|
||||
UniDiffuserModel,
|
||||
UniDiffuserPipeline,
|
||||
UniDiffuserTextDecoder,
|
||||
VersatileDiffusionDualGuidedPipeline,
|
||||
VersatileDiffusionImageVariationPipeline,
|
||||
VersatileDiffusionPipeline,
|
||||
|
||||
@@ -89,6 +89,7 @@ else:
|
||||
from .stable_diffusion_safe import StableDiffusionPipelineSafe
|
||||
from .text_to_video_synthesis import TextToVideoSDPipeline, TextToVideoZeroPipeline
|
||||
from .unclip import UnCLIPImageVariationPipeline, UnCLIPPipeline
|
||||
from .unidiffuser import ImageTextPipelineOutput, UniDiffuserModel, UniDiffuserPipeline, UniDiffuserTextDecoder
|
||||
from .versatile_diffusion import (
|
||||
VersatileDiffusionDualGuidedPipeline,
|
||||
VersatileDiffusionImageVariationPipeline,
|
||||
|
||||
20
src/diffusers/pipelines/unidiffuser/__init__.py
Normal file
20
src/diffusers/pipelines/unidiffuser/__init__.py
Normal file
@@ -0,0 +1,20 @@
|
||||
from ...utils import (
|
||||
OptionalDependencyNotAvailable,
|
||||
is_torch_available,
|
||||
is_transformers_available,
|
||||
is_transformers_version,
|
||||
)
|
||||
|
||||
|
||||
try:
|
||||
if not (is_transformers_available() and is_torch_available()):
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
from ...utils.dummy_torch_and_transformers_objects import (
|
||||
ImageTextPipelineOutput,
|
||||
UniDiffuserPipeline,
|
||||
)
|
||||
else:
|
||||
from .modeling_text_decoder import UniDiffuserTextDecoder
|
||||
from .modeling_uvit import UniDiffuserModel, UTransformer2DModel
|
||||
from .pipeline_unidiffuser import ImageTextPipelineOutput, UniDiffuserPipeline
|
||||
294
src/diffusers/pipelines/unidiffuser/modeling_text_decoder.py
Normal file
294
src/diffusers/pipelines/unidiffuser/modeling_text_decoder.py
Normal file
@@ -0,0 +1,294 @@
|
||||
from typing import Optional
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from torch import nn
|
||||
from transformers import GPT2Config, GPT2LMHeadModel
|
||||
from transformers.modeling_utils import ModuleUtilsMixin
|
||||
|
||||
from ...configuration_utils import ConfigMixin, register_to_config
|
||||
from ...models import ModelMixin
|
||||
|
||||
|
||||
# Modified from ClipCaptionModel in https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py
|
||||
class UniDiffuserTextDecoder(ModelMixin, ConfigMixin, ModuleUtilsMixin):
|
||||
"""
|
||||
Text decoder model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is used to
|
||||
generate text from the UniDiffuser image-text embedding.
|
||||
|
||||
Parameters:
|
||||
prefix_length (`int`):
|
||||
Max number of prefix tokens that will be supplied to the model.
|
||||
prefix_inner_dim (`int`):
|
||||
The hidden size of the the incoming prefix embeddings. For UniDiffuser, this would be the hidden dim of the
|
||||
CLIP text encoder.
|
||||
prefix_hidden_dim (`int`, *optional*):
|
||||
Hidden dim of the MLP if we encode the prefix.
|
||||
vocab_size (`int`, *optional*, defaults to 50257):
|
||||
Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`].
|
||||
n_positions (`int`, *optional*, defaults to 1024):
|
||||
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
||||
just in case (e.g., 512 or 1024 or 2048).
|
||||
n_embd (`int`, *optional*, defaults to 768):
|
||||
Dimensionality of the embeddings and hidden states.
|
||||
n_layer (`int`, *optional*, defaults to 12):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
n_head (`int`, *optional*, defaults to 12):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
n_inner (`int`, *optional*, defaults to None):
|
||||
Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
|
||||
activation_function (`str`, *optional*, defaults to `"gelu"`):
|
||||
Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
|
||||
resid_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
||||
embd_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout ratio for the embeddings.
|
||||
attn_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout ratio for the attention.
|
||||
layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
|
||||
The epsilon to use in the layer normalization layers.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
scale_attn_weights (`bool`, *optional*, defaults to `True`):
|
||||
Scale attention weights by dividing by sqrt(hidden_size)..
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models).
|
||||
scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
|
||||
Whether to additionally scale attention weights by `1 / layer_idx + 1`.
|
||||
reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
|
||||
Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
|
||||
dot-product/softmax to float() when training with mixed precision.
|
||||
"""
|
||||
|
||||
@register_to_config
|
||||
def __init__(
|
||||
self,
|
||||
prefix_length: int,
|
||||
prefix_inner_dim: int,
|
||||
prefix_hidden_dim: Optional[int] = None,
|
||||
vocab_size: int = 50257, # Start of GPT2 config args
|
||||
n_positions: int = 1024,
|
||||
n_embd: int = 768,
|
||||
n_layer: int = 12,
|
||||
n_head: int = 12,
|
||||
n_inner: Optional[int] = None,
|
||||
activation_function: str = "gelu_new",
|
||||
resid_pdrop: float = 0.1,
|
||||
embd_pdrop: float = 0.1,
|
||||
attn_pdrop: float = 0.1,
|
||||
layer_norm_epsilon: float = 1e-5,
|
||||
initializer_range: float = 0.02,
|
||||
scale_attn_weights: bool = True,
|
||||
use_cache: bool = True,
|
||||
scale_attn_by_inverse_layer_idx: bool = False,
|
||||
reorder_and_upcast_attn: bool = False,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.prefix_length = prefix_length
|
||||
|
||||
if prefix_inner_dim != n_embd and prefix_hidden_dim is None:
|
||||
raise ValueError(
|
||||
f"`prefix_hidden_dim` cannot be `None` when `prefix_inner_dim`: {prefix_hidden_dim} and"
|
||||
f" `n_embd`: {n_embd} are not equal."
|
||||
)
|
||||
|
||||
self.prefix_inner_dim = prefix_inner_dim
|
||||
self.prefix_hidden_dim = prefix_hidden_dim
|
||||
|
||||
self.encode_prefix = (
|
||||
nn.Linear(self.prefix_inner_dim, self.prefix_hidden_dim)
|
||||
if self.prefix_hidden_dim is not None
|
||||
else nn.Identity()
|
||||
)
|
||||
self.decode_prefix = (
|
||||
nn.Linear(self.prefix_hidden_dim, n_embd) if self.prefix_hidden_dim is not None else nn.Identity()
|
||||
)
|
||||
|
||||
gpt_config = GPT2Config(
|
||||
vocab_size=vocab_size,
|
||||
n_positions=n_positions,
|
||||
n_embd=n_embd,
|
||||
n_layer=n_layer,
|
||||
n_head=n_head,
|
||||
n_inner=n_inner,
|
||||
activation_function=activation_function,
|
||||
resid_pdrop=resid_pdrop,
|
||||
embd_pdrop=embd_pdrop,
|
||||
attn_pdrop=attn_pdrop,
|
||||
layer_norm_epsilon=layer_norm_epsilon,
|
||||
initializer_range=initializer_range,
|
||||
scale_attn_weights=scale_attn_weights,
|
||||
use_cache=use_cache,
|
||||
scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
|
||||
reorder_and_upcast_attn=reorder_and_upcast_attn,
|
||||
)
|
||||
self.transformer = GPT2LMHeadModel(gpt_config)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
input_ids: torch.Tensor,
|
||||
prefix_embeds: torch.Tensor,
|
||||
attention_mask: Optional[torch.Tensor] = None,
|
||||
labels: Optional[torch.Tensor] = None,
|
||||
):
|
||||
"""
|
||||
Args:
|
||||
input_ids (`torch.Tensor` of shape `(N, max_seq_len)`):
|
||||
Text tokens to use for inference.
|
||||
prefix_embeds (`torch.Tensor` of shape `(N, prefix_length, 768)`):
|
||||
Prefix embedding to preprend to the embedded tokens.
|
||||
attention_mask (`torch.Tensor` of shape `(N, prefix_length + max_seq_len, 768)`, *optional*):
|
||||
Attention mask for the prefix embedding.
|
||||
labels (`torch.Tensor`, *optional*):
|
||||
Labels to use for language modeling.
|
||||
"""
|
||||
embedding_text = self.transformer.transformer.wte(input_ids)
|
||||
hidden = self.encode_prefix(prefix_embeds)
|
||||
prefix_embeds = self.decode_prefix(hidden)
|
||||
embedding_cat = torch.cat((prefix_embeds, embedding_text), dim=1)
|
||||
|
||||
if labels is not None:
|
||||
dummy_token = self.get_dummy_token(input_ids.shape[0], input_ids.device)
|
||||
labels = torch.cat((dummy_token, input_ids), dim=1)
|
||||
out = self.transformer(inputs_embeds=embedding_cat, labels=labels, attention_mask=attention_mask)
|
||||
if self.prefix_hidden_dim is not None:
|
||||
return out, hidden
|
||||
else:
|
||||
return out
|
||||
|
||||
def get_dummy_token(self, batch_size: int, device: torch.device) -> torch.Tensor:
|
||||
return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device)
|
||||
|
||||
def encode(self, prefix):
|
||||
return self.encode_prefix(prefix)
|
||||
|
||||
@torch.no_grad()
|
||||
def generate_captions(self, features, eos_token_id, device):
|
||||
"""
|
||||
Generate captions given text embedding features. Returns list[L].
|
||||
|
||||
Args:
|
||||
features (`torch.Tensor` of shape `(B, L, D)`):
|
||||
Text embedding features to generate captions from.
|
||||
eos_token_id (`int`):
|
||||
The token ID of the EOS token for the text decoder model.
|
||||
device:
|
||||
Device to perform text generation on.
|
||||
|
||||
Returns:
|
||||
`List[str]`: A list of strings generated from the decoder model.
|
||||
"""
|
||||
|
||||
features = torch.split(features, 1, dim=0)
|
||||
generated_tokens = []
|
||||
generated_seq_lengths = []
|
||||
for feature in features:
|
||||
feature = self.decode_prefix(feature.to(device)) # back to the clip feature
|
||||
# Only support beam search for now
|
||||
output_tokens, seq_lengths = self.generate_beam(
|
||||
input_embeds=feature, device=device, eos_token_id=eos_token_id
|
||||
)
|
||||
generated_tokens.append(output_tokens[0])
|
||||
generated_seq_lengths.append(seq_lengths[0])
|
||||
generated_tokens = torch.stack(generated_tokens)
|
||||
generated_seq_lengths = torch.stack(generated_seq_lengths)
|
||||
return generated_tokens, generated_seq_lengths
|
||||
|
||||
@torch.no_grad()
|
||||
def generate_beam(
|
||||
self,
|
||||
input_ids=None,
|
||||
input_embeds=None,
|
||||
device=None,
|
||||
beam_size: int = 5,
|
||||
entry_length: int = 67,
|
||||
temperature: float = 1.0,
|
||||
eos_token_id: Optional[int] = None,
|
||||
):
|
||||
"""
|
||||
Generates text using the given tokenizer and text prompt or token embedding via beam search. This
|
||||
implementation is based on the beam search implementation from the [original UniDiffuser
|
||||
code](https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py#L89).
|
||||
|
||||
Args:
|
||||
eos_token_id (`int`, *optional*):
|
||||
The token ID of the EOS token for the text decoder model.
|
||||
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
|
||||
Tokenizer indices of input sequence tokens in the vocabulary. One of `input_ids` and `input_embeds`
|
||||
must be supplied.
|
||||
input_embeds (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
|
||||
An embedded representation to directly pass to the transformer as a prefix for beam search. One of
|
||||
`input_ids` and `input_embeds` must be supplied.
|
||||
device:
|
||||
The device to perform beam search on.
|
||||
beam_size (`int`, *optional*, defaults to `5`):
|
||||
The number of best states to store during beam search.
|
||||
entry_length (`int`, *optional*, defaults to `67`):
|
||||
The number of iterations to run beam search.
|
||||
temperature (`float`, *optional*, defaults to 1.0):
|
||||
The temperature to use when performing the softmax over logits from the decoding model.
|
||||
|
||||
Returns:
|
||||
`Tuple(torch.Tensor, torch.Tensor)`: A tuple of tensors where the first element is a tensor of generated
|
||||
token sequences sorted by score in descending order, and the second element is the sequence lengths
|
||||
corresponding to those sequences.
|
||||
"""
|
||||
# Generates text until stop_token is reached using beam search with the desired beam size.
|
||||
stop_token_index = eos_token_id
|
||||
tokens = None
|
||||
scores = None
|
||||
seq_lengths = torch.ones(beam_size, device=device, dtype=torch.int)
|
||||
is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool)
|
||||
|
||||
if input_embeds is not None:
|
||||
generated = input_embeds
|
||||
else:
|
||||
generated = self.transformer.transformer.wte(input_ids)
|
||||
|
||||
for i in range(entry_length):
|
||||
outputs = self.transformer(inputs_embeds=generated)
|
||||
logits = outputs.logits
|
||||
logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0)
|
||||
logits = logits.softmax(-1).log()
|
||||
|
||||
if scores is None:
|
||||
scores, next_tokens = logits.topk(beam_size, -1)
|
||||
generated = generated.expand(beam_size, *generated.shape[1:])
|
||||
next_tokens, scores = next_tokens.permute(1, 0), scores.squeeze(0)
|
||||
if tokens is None:
|
||||
tokens = next_tokens
|
||||
else:
|
||||
tokens = tokens.expand(beam_size, *tokens.shape[1:])
|
||||
tokens = torch.cat((tokens, next_tokens), dim=1)
|
||||
else:
|
||||
logits[is_stopped] = -float(np.inf)
|
||||
logits[is_stopped, 0] = 0
|
||||
scores_sum = scores[:, None] + logits
|
||||
seq_lengths[~is_stopped] += 1
|
||||
scores_sum_average = scores_sum / seq_lengths[:, None]
|
||||
scores_sum_average, next_tokens = scores_sum_average.view(-1).topk(beam_size, -1)
|
||||
next_tokens_source = next_tokens // scores_sum.shape[1]
|
||||
seq_lengths = seq_lengths[next_tokens_source]
|
||||
next_tokens = next_tokens % scores_sum.shape[1]
|
||||
next_tokens = next_tokens.unsqueeze(1)
|
||||
tokens = tokens[next_tokens_source]
|
||||
tokens = torch.cat((tokens, next_tokens), dim=1)
|
||||
generated = generated[next_tokens_source]
|
||||
scores = scores_sum_average * seq_lengths
|
||||
is_stopped = is_stopped[next_tokens_source]
|
||||
|
||||
next_token_embed = self.transformer.transformer.wte(next_tokens.squeeze()).view(generated.shape[0], 1, -1)
|
||||
generated = torch.cat((generated, next_token_embed), dim=1)
|
||||
is_stopped = is_stopped + next_tokens.eq(stop_token_index).squeeze()
|
||||
if is_stopped.all():
|
||||
break
|
||||
|
||||
scores = scores / seq_lengths
|
||||
order = scores.argsort(descending=True)
|
||||
# tokens tensors are already padded to max_seq_length
|
||||
output_texts = [tokens[i] for i in order]
|
||||
output_texts = torch.stack(output_texts, dim=0)
|
||||
seq_lengths = torch.tensor([seq_lengths[i] for i in order], dtype=seq_lengths.dtype)
|
||||
return output_texts, seq_lengths
|
||||
1196
src/diffusers/pipelines/unidiffuser/modeling_uvit.py
Normal file
1196
src/diffusers/pipelines/unidiffuser/modeling_uvit.py
Normal file
File diff suppressed because it is too large
Load Diff
1422
src/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py
Normal file
1422
src/diffusers/pipelines/unidiffuser/pipeline_unidiffuser.py
Normal file
File diff suppressed because it is too large
Load Diff
@@ -152,6 +152,21 @@ class IFSuperResolutionPipeline(metaclass=DummyObject):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class ImageTextPipelineOutput(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class KandinskyImg2ImgPipeline(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
@@ -632,6 +647,51 @@ class UnCLIPPipeline(metaclass=DummyObject):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class UniDiffuserModel(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class UniDiffuserPipeline(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class UniDiffuserTextDecoder(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, **kwargs):
|
||||
requires_backends(cls, ["torch", "transformers"])
|
||||
|
||||
|
||||
class VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject):
|
||||
_backends = ["torch", "transformers"]
|
||||
|
||||
|
||||
0
tests/pipelines/unidiffuser/__init__.py
Normal file
0
tests/pipelines/unidiffuser/__init__.py
Normal file
670
tests/pipelines/unidiffuser/test_unidiffuser.py
Normal file
670
tests/pipelines/unidiffuser/test_unidiffuser.py
Normal file
@@ -0,0 +1,670 @@
|
||||
import gc
|
||||
import random
|
||||
import unittest
|
||||
|
||||
import numpy as np
|
||||
import torch
|
||||
from PIL import Image
|
||||
from transformers import (
|
||||
CLIPImageProcessor,
|
||||
CLIPTextModel,
|
||||
CLIPTokenizer,
|
||||
CLIPVisionModelWithProjection,
|
||||
GPT2Tokenizer,
|
||||
)
|
||||
|
||||
from diffusers import (
|
||||
AutoencoderKL,
|
||||
DPMSolverMultistepScheduler,
|
||||
UniDiffuserModel,
|
||||
UniDiffuserPipeline,
|
||||
UniDiffuserTextDecoder,
|
||||
)
|
||||
from diffusers.utils import floats_tensor, load_image, randn_tensor, slow, torch_device
|
||||
from diffusers.utils.testing_utils import require_torch_gpu
|
||||
|
||||
from ..pipeline_params import TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS, TEXT_GUIDED_IMAGE_VARIATION_PARAMS
|
||||
from ..test_pipelines_common import PipelineTesterMixin
|
||||
|
||||
|
||||
class UniDiffuserPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
|
||||
pipeline_class = UniDiffuserPipeline
|
||||
params = TEXT_GUIDED_IMAGE_VARIATION_PARAMS
|
||||
batch_params = TEXT_GUIDED_IMAGE_VARIATION_BATCH_PARAMS
|
||||
|
||||
def get_dummy_components(self):
|
||||
unet = UniDiffuserModel.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="unet",
|
||||
)
|
||||
|
||||
scheduler = DPMSolverMultistepScheduler(
|
||||
beta_start=0.00085,
|
||||
beta_end=0.012,
|
||||
beta_schedule="scaled_linear",
|
||||
solver_order=3,
|
||||
)
|
||||
|
||||
vae = AutoencoderKL.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="vae",
|
||||
)
|
||||
|
||||
text_encoder = CLIPTextModel.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="text_encoder",
|
||||
)
|
||||
clip_tokenizer = CLIPTokenizer.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="clip_tokenizer",
|
||||
)
|
||||
|
||||
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="image_encoder",
|
||||
)
|
||||
# From the Stable Diffusion Image Variation pipeline tests
|
||||
image_processor = CLIPImageProcessor(crop_size=32, size=32)
|
||||
# image_processor = CLIPImageProcessor.from_pretrained("hf-internal-testing/tiny-random-clip")
|
||||
|
||||
text_tokenizer = GPT2Tokenizer.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="text_tokenizer",
|
||||
)
|
||||
text_decoder = UniDiffuserTextDecoder.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-diffusers-test",
|
||||
subfolder="text_decoder",
|
||||
)
|
||||
|
||||
components = {
|
||||
"vae": vae,
|
||||
"text_encoder": text_encoder,
|
||||
"image_encoder": image_encoder,
|
||||
"image_processor": image_processor,
|
||||
"clip_tokenizer": clip_tokenizer,
|
||||
"text_decoder": text_decoder,
|
||||
"text_tokenizer": text_tokenizer,
|
||||
"unet": unet,
|
||||
"scheduler": scheduler,
|
||||
}
|
||||
|
||||
return components
|
||||
|
||||
def get_dummy_inputs(self, device, seed=0):
|
||||
image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
|
||||
image = image.cpu().permute(0, 2, 3, 1)[0]
|
||||
image = Image.fromarray(np.uint8(image)).convert("RGB")
|
||||
if str(device).startswith("mps"):
|
||||
generator = torch.manual_seed(seed)
|
||||
else:
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
inputs = {
|
||||
"prompt": "an elephant under the sea",
|
||||
"image": image,
|
||||
"generator": generator,
|
||||
"num_inference_steps": 2,
|
||||
"guidance_scale": 6.0,
|
||||
"output_type": "numpy",
|
||||
}
|
||||
return inputs
|
||||
|
||||
def get_fixed_latents(self, device, seed=0):
|
||||
if type(device) == str:
|
||||
device = torch.device(device)
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
# Hardcode the shapes for now.
|
||||
prompt_latents = randn_tensor((1, 77, 32), generator=generator, device=device, dtype=torch.float32)
|
||||
vae_latents = randn_tensor((1, 4, 16, 16), generator=generator, device=device, dtype=torch.float32)
|
||||
clip_latents = randn_tensor((1, 1, 32), generator=generator, device=device, dtype=torch.float32)
|
||||
|
||||
latents = {
|
||||
"prompt_latents": prompt_latents,
|
||||
"vae_latents": vae_latents,
|
||||
"clip_latents": clip_latents,
|
||||
}
|
||||
return latents
|
||||
|
||||
def get_dummy_inputs_with_latents(self, device, seed=0):
|
||||
# image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
|
||||
# image = image.cpu().permute(0, 2, 3, 1)[0]
|
||||
# image = Image.fromarray(np.uint8(image)).convert("RGB")
|
||||
image = load_image(
|
||||
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/unidiffuser/unidiffuser_example_image.jpg",
|
||||
)
|
||||
image = image.resize((32, 32))
|
||||
latents = self.get_fixed_latents(device, seed=seed)
|
||||
|
||||
if str(device).startswith("mps"):
|
||||
generator = torch.manual_seed(seed)
|
||||
else:
|
||||
generator = torch.Generator(device=device).manual_seed(seed)
|
||||
|
||||
inputs = {
|
||||
"prompt": "an elephant under the sea",
|
||||
"image": image,
|
||||
"generator": generator,
|
||||
"num_inference_steps": 2,
|
||||
"guidance_scale": 6.0,
|
||||
"output_type": "numpy",
|
||||
"prompt_latents": latents.get("prompt_latents"),
|
||||
"vae_latents": latents.get("vae_latents"),
|
||||
"clip_latents": latents.get("clip_latents"),
|
||||
}
|
||||
return inputs
|
||||
|
||||
def test_unidiffuser_default_joint_v0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'joint'
|
||||
unidiffuser_pipe.set_joint_mode()
|
||||
assert unidiffuser_pipe.mode == "joint"
|
||||
|
||||
# inputs = self.get_dummy_inputs(device)
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
sample = unidiffuser_pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.5760, 0.6270, 0.6571, 0.4965, 0.4638, 0.5663, 0.5254, 0.5068, 0.5716])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-3
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_joint_no_cfg_v0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'joint'
|
||||
unidiffuser_pipe.set_joint_mode()
|
||||
assert unidiffuser_pipe.mode == "joint"
|
||||
|
||||
# inputs = self.get_dummy_inputs(device)
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
# Set guidance scale to 1.0 to turn off CFG
|
||||
inputs["guidance_scale"] = 1.0
|
||||
sample = unidiffuser_pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.5760, 0.6270, 0.6571, 0.4965, 0.4638, 0.5663, 0.5254, 0.5068, 0.5716])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-3
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_text2img_v0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'text2img'
|
||||
unidiffuser_pipe.set_text_to_image_mode()
|
||||
assert unidiffuser_pipe.mode == "text2img"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete image for text-conditioned image generation
|
||||
del inputs["image"]
|
||||
image = unidiffuser_pipe(**inputs).images
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_slice = np.array([0.5758, 0.6269, 0.6570, 0.4967, 0.4639, 0.5664, 0.5257, 0.5067, 0.5715])
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
|
||||
|
||||
def test_unidiffuser_default_image_0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img'
|
||||
unidiffuser_pipe.set_image_mode()
|
||||
assert unidiffuser_pipe.mode == "img"
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
# Delete prompt and image for unconditional ("marginal") text generation.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
image = unidiffuser_pipe(**inputs).images
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_slice = np.array([0.5760, 0.6270, 0.6571, 0.4966, 0.4638, 0.5663, 0.5254, 0.5068, 0.5715])
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
|
||||
|
||||
def test_unidiffuser_default_text_v0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img'
|
||||
unidiffuser_pipe.set_text_mode()
|
||||
assert unidiffuser_pipe.mode == "text"
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
# Delete prompt and image for unconditional ("marginal") text generation.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_img2text_v0(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img2text'
|
||||
unidiffuser_pipe.set_image_to_text_mode()
|
||||
assert unidiffuser_pipe.mode == "img2text"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete text for image-conditioned text generation
|
||||
del inputs["prompt"]
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_joint_v1(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained("hf-internal-testing/unidiffuser-test-v1")
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'joint'
|
||||
unidiffuser_pipe.set_joint_mode()
|
||||
assert unidiffuser_pipe.mode == "joint"
|
||||
|
||||
# inputs = self.get_dummy_inputs(device)
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
inputs["data_type"] = 1
|
||||
sample = unidiffuser_pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.5760, 0.6270, 0.6571, 0.4965, 0.4638, 0.5663, 0.5254, 0.5068, 0.5716])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-3
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_text2img_v1(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained("hf-internal-testing/unidiffuser-test-v1")
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'text2img'
|
||||
unidiffuser_pipe.set_text_to_image_mode()
|
||||
assert unidiffuser_pipe.mode == "text2img"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete image for text-conditioned image generation
|
||||
del inputs["image"]
|
||||
image = unidiffuser_pipe(**inputs).images
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_slice = np.array([0.5758, 0.6269, 0.6570, 0.4967, 0.4639, 0.5664, 0.5257, 0.5067, 0.5715])
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
|
||||
|
||||
def test_unidiffuser_default_img2text_v1(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained("hf-internal-testing/unidiffuser-test-v1")
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img2text'
|
||||
unidiffuser_pipe.set_image_to_text_mode()
|
||||
assert unidiffuser_pipe.mode == "img2text"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete text for image-conditioned text generation
|
||||
del inputs["prompt"]
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
expected_text_prefix = " no no no "
|
||||
assert text[0][:10] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_text2img_multiple_images(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'text2img'
|
||||
unidiffuser_pipe.set_text_to_image_mode()
|
||||
assert unidiffuser_pipe.mode == "text2img"
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
# Delete image for text-conditioned image generation
|
||||
del inputs["image"]
|
||||
inputs["num_images_per_prompt"] = 2
|
||||
inputs["num_prompts_per_image"] = 3
|
||||
image = unidiffuser_pipe(**inputs).images
|
||||
assert image.shape == (2, 32, 32, 3)
|
||||
|
||||
def test_unidiffuser_img2text_multiple_prompts(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img2text'
|
||||
unidiffuser_pipe.set_image_to_text_mode()
|
||||
assert unidiffuser_pipe.mode == "img2text"
|
||||
|
||||
inputs = self.get_dummy_inputs(device)
|
||||
# Delete text for image-conditioned text generation
|
||||
del inputs["prompt"]
|
||||
inputs["num_images_per_prompt"] = 2
|
||||
inputs["num_prompts_per_image"] = 3
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
assert len(text) == 3
|
||||
|
||||
def test_unidiffuser_text2img_multiple_images_with_latents(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'text2img'
|
||||
unidiffuser_pipe.set_text_to_image_mode()
|
||||
assert unidiffuser_pipe.mode == "text2img"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete image for text-conditioned image generation
|
||||
del inputs["image"]
|
||||
inputs["num_images_per_prompt"] = 2
|
||||
inputs["num_prompts_per_image"] = 3
|
||||
image = unidiffuser_pipe(**inputs).images
|
||||
assert image.shape == (2, 32, 32, 3)
|
||||
|
||||
def test_unidiffuser_img2text_multiple_prompts_with_latents(self):
|
||||
device = "cpu" # ensure determinism for the device-dependent torch.Generator
|
||||
components = self.get_dummy_components()
|
||||
unidiffuser_pipe = UniDiffuserPipeline(**components)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img2text'
|
||||
unidiffuser_pipe.set_image_to_text_mode()
|
||||
assert unidiffuser_pipe.mode == "img2text"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete text for image-conditioned text generation
|
||||
del inputs["prompt"]
|
||||
inputs["num_images_per_prompt"] = 2
|
||||
inputs["num_prompts_per_image"] = 3
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
assert len(text) == 3
|
||||
|
||||
@require_torch_gpu
|
||||
def test_unidiffuser_default_joint_v1_cuda_fp16(self):
|
||||
device = "cuda"
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-test-v1", torch_dtype=torch.float16
|
||||
)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'joint'
|
||||
unidiffuser_pipe.set_joint_mode()
|
||||
assert unidiffuser_pipe.mode == "joint"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
inputs["data_type"] = 1
|
||||
sample = unidiffuser_pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.5049, 0.5498, 0.5854, 0.3052, 0.4460, 0.6489, 0.5122, 0.4810, 0.6138])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-3
|
||||
|
||||
expected_text_prefix = '" This This'
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
|
||||
@require_torch_gpu
|
||||
def test_unidiffuser_default_text2img_v1_cuda_fp16(self):
|
||||
device = "cuda"
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-test-v1", torch_dtype=torch.float16
|
||||
)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'text2img'
|
||||
unidiffuser_pipe.set_text_to_image_mode()
|
||||
assert unidiffuser_pipe.mode == "text2img"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["image"]
|
||||
inputs["data_type"] = 1
|
||||
sample = unidiffuser_pipe(**inputs)
|
||||
image = sample.images
|
||||
assert image.shape == (1, 32, 32, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.5054, 0.5498, 0.5854, 0.3052, 0.4458, 0.6489, 0.5122, 0.4810, 0.6138])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-3
|
||||
|
||||
@require_torch_gpu
|
||||
def test_unidiffuser_default_img2text_v1_cuda_fp16(self):
|
||||
device = "cuda"
|
||||
unidiffuser_pipe = UniDiffuserPipeline.from_pretrained(
|
||||
"hf-internal-testing/unidiffuser-test-v1", torch_dtype=torch.float16
|
||||
)
|
||||
unidiffuser_pipe = unidiffuser_pipe.to(device)
|
||||
unidiffuser_pipe.set_progress_bar_config(disable=None)
|
||||
|
||||
# Set mode to 'img2text'
|
||||
unidiffuser_pipe.set_image_to_text_mode()
|
||||
assert unidiffuser_pipe.mode == "img2text"
|
||||
|
||||
inputs = self.get_dummy_inputs_with_latents(device)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
inputs["data_type"] = 1
|
||||
text = unidiffuser_pipe(**inputs).text
|
||||
|
||||
expected_text_prefix = '" This This'
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch_gpu
|
||||
class UniDiffuserPipelineSlowTests(unittest.TestCase):
|
||||
def tearDown(self):
|
||||
super().tearDown()
|
||||
gc.collect()
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
def get_inputs(self, device, seed=0, generate_latents=False):
|
||||
generator = torch.manual_seed(seed)
|
||||
image = load_image(
|
||||
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/unidiffuser/unidiffuser_example_image.jpg"
|
||||
)
|
||||
inputs = {
|
||||
"prompt": "an elephant under the sea",
|
||||
"image": image,
|
||||
"generator": generator,
|
||||
"num_inference_steps": 3,
|
||||
"guidance_scale": 8.0,
|
||||
"output_type": "numpy",
|
||||
}
|
||||
if generate_latents:
|
||||
latents = self.get_fixed_latents(device, seed=seed)
|
||||
for latent_name, latent_tensor in latents.items():
|
||||
inputs[latent_name] = latent_tensor
|
||||
return inputs
|
||||
|
||||
def get_fixed_latents(self, device, seed=0):
|
||||
if type(device) == str:
|
||||
device = torch.device(device)
|
||||
latent_device = torch.device("cpu")
|
||||
generator = torch.Generator(device=latent_device).manual_seed(seed)
|
||||
# Hardcode the shapes for now.
|
||||
prompt_latents = randn_tensor((1, 77, 768), generator=generator, device=device, dtype=torch.float32)
|
||||
vae_latents = randn_tensor((1, 4, 64, 64), generator=generator, device=device, dtype=torch.float32)
|
||||
clip_latents = randn_tensor((1, 1, 512), generator=generator, device=device, dtype=torch.float32)
|
||||
|
||||
# Move latents onto desired device.
|
||||
prompt_latents = prompt_latents.to(device)
|
||||
vae_latents = vae_latents.to(device)
|
||||
clip_latents = clip_latents.to(device)
|
||||
|
||||
latents = {
|
||||
"prompt_latents": prompt_latents,
|
||||
"vae_latents": vae_latents,
|
||||
"clip_latents": clip_latents,
|
||||
}
|
||||
return latents
|
||||
|
||||
def test_unidiffuser_default_joint_v1(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1")
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
# inputs = self.get_dummy_inputs(device)
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
sample = pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 512, 512, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.2402, 0.2375, 0.2285, 0.2378, 0.2407, 0.2263, 0.2354, 0.2307, 0.2520])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-1
|
||||
|
||||
expected_text_prefix = "A living room"
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_text2img_v1(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1")
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
del inputs["image"]
|
||||
sample = pipe(**inputs)
|
||||
image = sample.images
|
||||
assert image.shape == (1, 512, 512, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_slice = np.array([0.0242, 0.0103, 0.0022, 0.0129, 0.0000, 0.0090, 0.0376, 0.0508, 0.0005])
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-1
|
||||
|
||||
def test_unidiffuser_default_img2text_v1(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1")
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
del inputs["prompt"]
|
||||
sample = pipe(**inputs)
|
||||
text = sample.text
|
||||
|
||||
expected_text_prefix = "An astronaut"
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_joint_v1_fp16(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1", torch_dtype=torch.float16)
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
# inputs = self.get_dummy_inputs(device)
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
# Delete prompt and image for joint inference.
|
||||
del inputs["prompt"]
|
||||
del inputs["image"]
|
||||
sample = pipe(**inputs)
|
||||
image = sample.images
|
||||
text = sample.text
|
||||
assert image.shape == (1, 512, 512, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_img_slice = np.array([0.2402, 0.2375, 0.2285, 0.2378, 0.2407, 0.2263, 0.2354, 0.2307, 0.2520])
|
||||
assert np.abs(image_slice.flatten() - expected_img_slice).max() < 1e-1
|
||||
|
||||
expected_text_prefix = "A living room"
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
|
||||
def test_unidiffuser_default_text2img_v1_fp16(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1", torch_dtype=torch.float16)
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
del inputs["image"]
|
||||
sample = pipe(**inputs)
|
||||
image = sample.images
|
||||
assert image.shape == (1, 512, 512, 3)
|
||||
|
||||
image_slice = image[0, -3:, -3:, -1]
|
||||
expected_slice = np.array([0.0242, 0.0103, 0.0022, 0.0129, 0.0000, 0.0090, 0.0376, 0.0508, 0.0005])
|
||||
assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-1
|
||||
|
||||
def test_unidiffuser_default_img2text_v1_fp16(self):
|
||||
pipe = UniDiffuserPipeline.from_pretrained("thu-ml/unidiffuser-v1", torch_dtype=torch.float16)
|
||||
pipe.to(torch_device)
|
||||
pipe.set_progress_bar_config(disable=None)
|
||||
pipe.enable_attention_slicing()
|
||||
|
||||
inputs = self.get_inputs(device=torch_device, generate_latents=True)
|
||||
del inputs["prompt"]
|
||||
sample = pipe(**inputs)
|
||||
text = sample.text
|
||||
|
||||
expected_text_prefix = "An astronaut"
|
||||
assert text[0][: len(expected_text_prefix)] == expected_text_prefix
|
||||
Reference in New Issue
Block a user