Commit graph

267 commits

Author SHA1 Message Date
Cédric Deltheil 5ab5d7fd1c import ControlLoraAdapter part of latent_diffusion 2024-02-19 14:11:32 +01:00
Laurent da3c3602fb write StyleAligned e2e test 2024-02-15 15:22:47 +01:00
Laurent 60c0780fe7 write StyleAligned inject/eject tests 2024-02-15 15:22:47 +01:00
limiteinductive 432e32f94f rename Scheduler -> LRScheduler 2024-02-15 11:48:36 +01:00
Laurent 00270604ef fix conversion_script bug + rename control_lora e2e test 2024-02-14 18:20:46 +01:00
Laurent 7fe392298a write ControlLora e2e tests 2024-02-14 18:20:46 +01:00
limiteinductive bec845553f update deprecated validator for field_validator 2024-02-13 18:35:51 +01:00
limiteinductive ab506b4db2 fix bug that was causing double registration 2024-02-13 11:12:13 +01:00
limiteinductive 3488273f50 Enforce correct subtype for the config param in both decorators
Also add a custom ModelConfig for the MockTrainer test

Update src/refiners/training_utils/config.py

Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
2024-02-12 16:21:04 +01:00
limiteinductive cef8a9936c refactor register_model decorator 2024-02-12 16:21:04 +01:00
limiteinductive d6546c9026 add @register_model and @register_callback decorators
Refactor ClockTrainer to include Callback
2024-02-12 10:24:19 +01:00
limiteinductive f541badcb3 Allow optional train ModelConfig + forbid extra input for configs 2024-02-10 16:13:10 +01:00
Pierre Colle 25bfa78907 lr, betas, eps, weight_decay at model level
Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
2024-02-09 12:05:13 +01:00
Cédric Deltheil 9aefc9896c test_trainer: use model_copy instead of copy
The `copy` method has been deprecated.
2024-02-08 20:07:34 +01:00
Colle f4aa0271b8
less than 1 epoch training duration 2024-02-08 19:20:31 +01:00
limiteinductive 41508e0865 change param name of abstract get_item method 2024-02-08 18:52:52 +01:00
Laurent 6d599d53fd beautify EXPECTED_TREE in test_chain.py 2024-02-08 15:09:47 +01:00
limiteinductive 2e526d35d1 Make Dataset part of the trainer 2024-02-07 16:13:01 +01:00
limiteinductive 2ef4982e04 remove wandb from base config 2024-02-07 11:06:59 +01:00
Pierre Chapuis 11da76f7df fix sdxl structural copy 2024-02-07 10:51:26 +01:00
Pierre Chapuis ca9e89b22a cosmetics 2024-02-07 10:51:26 +01:00
limiteinductive ea05f3d327 make device and dtype work in Trainer class 2024-02-06 23:10:10 +01:00
Pierre Chapuis 37425fb609 make LoRA generic 2024-02-06 11:32:18 +01:00
Pierre Chapuis 471ef91d1c make __getattr__ on Module return object, not Any
PyTorch chose to make it Any because they expect its users' code
to be "highly dynamic": https://github.com/pytorch/pytorch/pull/104321

It is not the case for us, in Refiners having untyped code
goes contrary to one of our core principles.

Note that there is currently an open PR in PyTorch to
return `Module | Tensor`, but in practice this is not always
correct either: https://github.com/pytorch/pytorch/pull/115074

I also moved Residuals-related code from SD1 to latent_diffusion
because SDXL should not depend on SD1.
2024-02-06 11:32:18 +01:00
Pierre Chapuis 3de1508b65 increase tolerance on Euler test 2024-02-04 08:58:22 +01:00
Pierre Chapuis 83b478c0ff fix test failure caused by Diffusers 0.26.0 2024-02-04 08:58:22 +01:00
Colle 4a6146bb6c clip text, lda encode batch inputs
* text_encoder([str1, str2])
* lda decode_latents/encode_image image_to_latent/latent_to_image
* images_to_tensor, tensor_to_images
---------
Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
2024-02-01 17:05:28 +01:00
Pierre Chapuis 8a2b955bd0 add a test for SDXL + EulerScheduler (deterministic) 2024-02-01 16:17:07 +01:00
Pierre Chapuis 5ac5373310 add a test for SDXL with sliced attention 2024-02-01 16:17:07 +01:00
Pierre Chapuis 3ddd258d36 add a test for noise schedules 2024-02-01 16:17:07 +01:00
Pierre Chapuis df843f5226 test SAG setter 2024-02-01 16:17:07 +01:00
Pierre Chapuis f4ed7254fa test IP adapter scale setter 2024-02-01 16:17:07 +01:00
Pierre Chapuis 8341d3a74b use float16, save memory 2024-02-01 16:17:07 +01:00
Pierre Chapuis d185711bc5 add tests based on repr for inject / eject 2024-02-01 16:17:07 +01:00
Pierre Chapuis 0e77ef1720 add inject / eject test for concept extender (+ better errors) 2024-02-01 16:17:07 +01:00
Pierre Chapuis 93270ec2d7 add inject / eject test for t2i adapter 2024-02-01 16:17:07 +01:00
Pierre Chapuis f43a530254 add extra tests for Chain 2024-02-01 16:17:07 +01:00
Pierre Chapuis bca50b71f2 test (and fix) basic_attributes 2024-02-01 16:17:07 +01:00
Pierre Chapuis bba478abf2 create test_module 2024-02-01 16:17:07 +01:00
limiteinductive abe50076a4 add NoiseSchedule to solvers __init__ + simplify some import pathing
further improve import pathing
2024-01-31 17:03:52 +01:00
limiteinductive 73f6ccfc98 make Scheduler a fl.Module + Change name Scheduler -> Solver 2024-01-31 17:03:52 +01:00
Pierre Chapuis 7eb8eb4c68 add support for pytorch 2.2 (2.1 is still supported)
also bump all dev dependencies to their latest version
2024-01-31 15:03:06 +01:00
Cédric Deltheil 5634e68fde add end-to-end test for multi-ip adapter 2024-01-31 11:03:49 +01:00
Cédric Deltheil feff4c78ae segment-anything: fix class name typo
Note: weights are impacted
2024-01-30 09:52:40 +01:00
Pierre Chapuis 19bc081658 call GC between each e2e test to avoid OOM 2024-01-29 17:57:40 +01:00
Pierre Chapuis ce0339b4cc add a get_path helper to modules 2024-01-26 19:31:13 +01:00
limiteinductive 0ee2d5e075 Fix warmup steps calculation when gradient_accumulation is used 2024-01-25 12:20:36 +01:00
Bryce 12a5439fc4 refactor: rename noise => predicted_noise
and in euler, `alt_noise` can now be simply `noise`
2024-01-24 18:15:10 +01:00
limiteinductive 3b458f0d8d fix test_names LoraManager test 2024-01-23 14:12:03 +01:00
limiteinductive 421da6a3b6 Load Multiple LoRAs with SDLoraManager 2024-01-23 14:12:03 +01:00
Cédric Deltheil 40c33b9595 rollback to 50 inference steps in IP-Adapter tests
Follow up of 8a36c8c

This is what is used in the official notebook (ip_adapter_demo.ipynb and
ip_adapter-plus_demo.ipynb)
2024-01-22 09:35:31 +01:00
limiteinductive ed3621362f Add load_tensors utils in fluxion 2024-01-21 12:34:33 +01:00
Pierre Colle 91aea9b7ff fix: summarize_tensor(tensor) when tensor.numel() == 0 2024-01-20 14:32:35 +01:00
Pierre Chapuis 8a36c8c279 make the first diffusion step a first class property of LDM & Schedulers 2024-01-19 18:52:45 +01:00
hugojarkoff 42b7749630 Fix references for e2e tests 2024-01-19 15:00:03 +01:00
Pierre Chapuis ce3035923b improve DPM solver test 2024-01-18 19:23:11 +01:00
hugojarkoff 17d9701dde Remove additional noise in final sample of DDIM inference process 2024-01-18 18:43:13 +01:00
limiteinductive a1f50f3f9d refactor Lora LoraAdapter and the latent_diffusion/lora file 2024-01-18 16:27:38 +01:00
limiteinductive 2b977bc69e fix broken self-attention guidance with ip-adapter
The #168 and #177 refactorings caused this regression. A new end-to-end
test has been added for proper coverage.

(This fix will be revisited at some point)
2024-01-16 17:21:24 +01:00
limiteinductive d9ae7ca6a5 cast to float32 before converting to image in tensor_to_image to fix bfloat16 conversion 2024-01-16 11:50:58 +01:00
limiteinductive 7f722029be add basic unit test for training_utils 2024-01-14 22:08:20 +01:00
Colle dba9065229
fix test_debug_print
Follow-up of #173
2024-01-12 18:32:22 +01:00
Colle c141091afc
Make summarize_tensor robust to non-float dtypes (#171) 2024-01-11 09:57:58 +01:00
Cédric Deltheil ce0f9887a3 test_schedulers: fix pyright error
Due to changes in diffusers 0.25.0
2024-01-10 16:53:06 +01:00
Cédric Deltheil 6dbaec3e56 add end-to-end test for euler scheduler
Reference image generated with diffusers [1]

[1]: tests/e2e/test_diffusion_ref/README.md#expected-outputs
2024-01-10 16:53:06 +01:00
Cédric Deltheil 2b2b6740b7 fix or silent pyright issues 2024-01-10 16:53:06 +01:00
Cédric Deltheil 65f19d192f ruff fix 2024-01-10 16:53:06 +01:00
Cédric Deltheil ad143b0867 ruff format 2024-01-10 16:53:06 +01:00
Israfel Salazar 8423c5efa7
feature: Euler scheduler (#138) 2024-01-10 11:32:40 +01:00
limiteinductive c9e973ba41 refactor CrossAttentionAdapter to work with context. 2024-01-08 15:20:23 +01:00
hugojarkoff 00f494efe2 SegmentAnything: add dense mask prompt support 2024-01-05 18:53:25 +01:00
limiteinductive 20c229903f upgrade pyright to 1.1.342 ; improve no_grad typing 2023-12-29 15:09:02 +01:00
Cédric Deltheil 22ce3fd033 sam: wrap high-level methods with no_grad 2023-12-19 21:45:23 +01:00
Cédric Deltheil e7892254eb dinov2: add some coverage for registers
Those are not supported yet in HF: so just compared with a precomputed
norm. Note: in the initial PR [1] the Refiners' implementation has been
tested against the official code using Torch Hub.

[1]:
https://github.com/finegrain-ai/refiners/pull/132#issuecomment-1852021656
2023-12-18 10:29:28 +01:00
Cédric Deltheil 68cc346905 add minimal unit tests for DINOv2
To be completed with tests using image preprocessing, e.g. test cosine
similarity on a relevant pair of images
2023-12-18 10:29:28 +01:00
Benjamin Trom e2f2e33add Update tests/fluxion/layers/test_basics.py
Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
2023-12-13 17:03:28 +01:00
limiteinductive 7d9ceae274 change default behavior of end to None 2023-12-13 17:03:28 +01:00
Cédric Deltheil 82a2aa1ec4 deprecate DDPM step which is unused for now 2023-12-13 15:51:42 +01:00
limiteinductive a7551e0392 Change fl.Slicing API 2023-12-13 09:38:13 +01:00
Cédric Deltheil 315b4ed2e4 test_schedulers: enforce manual seed 2023-12-12 17:26:14 +01:00
Cédric Deltheil 792a0fc3d9 run lint rules using latest isort settings 2023-12-11 11:58:43 +01:00
limiteinductive 86c54977b9 replace poetry by rye for python dependency management
Co-authored-by: Cédric Deltheil <cedric@deltheil.me>
Co-authored-by: Pierre Chapuis <git@catwell.info>
2023-12-08 17:40:10 +01:00
limiteinductive 0dc3a17fbf remove unnecessary test 2023-12-04 15:27:06 +01:00
limiteinductive 8bbcf2048d add README bullet point 2023-12-04 15:27:06 +01:00
limiteinductive 37a74bd549 format test_scheduler file 2023-12-04 15:27:06 +01:00
limiteinductive 90db6ef59d add e2e test for sd15 with karras noise schedule 2023-12-04 15:27:06 +01:00
limiteinductive 6f110ee2b2 fix test_scheduler_utils 2023-12-04 15:27:06 +01:00
Pierre Chapuis f22f969d65 remove Black preview mode
also fix multiline logs in training
2023-12-04 14:15:56 +01:00
Cédric Deltheil b306c7db1b freeu: add one more test for identity scales
It should act as a NOP when [1.0, 1.0] is used for backbone and skip
scales.
2023-12-01 12:48:19 +01:00
Cédric Deltheil ab0915d052 add tests for FreeU 2023-11-18 16:15:44 +01:00
Pierre Chapuis 02f3c46e2e update pyright 2023-10-25 14:56:07 +02:00
Benjamin Trom ea44262a39 unnest Residual subchain by modifying its forward
And replaced the remaining Sum-Identity layers by Residual.

The tolerance used to compare SAM's ViT models has been tweaked: for
some reasons there is a small difference (in float32) in the neck layer
(first conv2D)

Co-authored-by: Cédric Deltheil <cedric@deltheil.me>
2023-10-19 10:34:51 +02:00
Cédric Deltheil 46dd710076 test_converter: use proper exception type
Follow up of #102
2023-10-18 14:39:24 +02:00
Benjamin Trom 6ddd901767 improve image_to_tensor and tensor_to_image utils 2023-10-17 18:08:58 +02:00
limiteinductive 7a62049d54 implement Restart method for latent diffusion 2023-10-12 15:48:43 +02:00
Benjamin Trom 0024191c58 improve debug print for chains 2023-10-10 15:25:09 +02:00
Benjamin Trom a663375dc7 prevent setattr pytorch module to register on the Chain class 2023-10-10 14:46:15 +02:00
Cédric Deltheil d02be0d10e tests: update ref image for SDXL IP-Adapter plus
Note: https://pytorch.org/docs/stable/notes/randomness.html

> Completely reproducible results are not guaranteed across PyTorch
> releases [...]
2023-10-10 14:19:47 +02:00
Cédric Deltheil b80769939d add support for self-attention guidance
See https://arxiv.org/abs/2210.00939
2023-10-09 17:33:15 +02:00
Cédric Deltheil 05126c8f4d make gaussian_blur work with float16 2023-10-07 21:48:38 +02:00
Cédric Deltheil 7d2abf6fbc scheduler: add remove noise
aka original sample prediction (or predict x0)

E.g. useful for methods like self-attention guidance (see equation (2)
in https://arxiv.org/pdf/2210.00939.pdf)
2023-10-05 17:05:15 +02:00
Cédric Deltheil 665bcdc95c add unit tests covering fluxion's gaussian_blur 2023-10-05 16:30:27 +02:00
Cédric Deltheil 338042f332 test_diffusion: remove debug leftovers 2023-09-29 18:54:24 +02:00
Cédric Deltheil 5fc6767a4a add IP-Adapter plus (aka fine-grained features) 2023-09-29 15:23:43 +02:00
Cédric Deltheil 63f5723449 test_concepts: silent static type checker error 2023-09-25 13:54:26 +02:00
Cédric Deltheil f37f25a2e4 add e2e test for T2I-Adapter XL canny 2023-09-25 13:54:26 +02:00
Cédric Deltheil 4301e81eb3 add e2e test for T2I-Adapter depth
Expected output generated with diffusers' StableDiffusionAdapterPipeline
2023-09-25 13:54:26 +02:00
Cédric Deltheil d72e1d3478 chain: add insert_before_type 2023-09-25 13:54:26 +02:00
Doryan Kaced 251277a0a8 Fix module registration in IP-Adapter 2023-09-22 17:34:55 +02:00
Pierre Chapuis cd1fdb5585 fix scheduler device choice 2023-09-21 12:00:19 +02:00
Benjamin Trom 282578ddc0 add Segment Anything (SAM) to foundational models
Note: dense prompts (i.e. masks) support is still partial (see MaskEncoder)

Co-authored-by: Cédric Deltheil <cedric@deltheil.me>
2023-09-21 11:44:30 +02:00
Cédric Deltheil 2faff9f57a ldm: properly resize non-square init image 2023-09-20 10:27:22 +02:00
Benjamin Trom 01aeaf3e36 add unit test for multi_diffusion 2023-09-19 15:30:50 +02:00
Pierre Chapuis fc2390ad1c fix legacy wording for refonly control 2023-09-14 11:21:11 +02:00
Pierre Chapuis 0e0c39b4b5 black 2023-09-13 17:02:47 +02:00
Pierre Chapuis c421cfd56c add a test for IP-Adapter + ControlNet 2023-09-13 14:24:53 +02:00
Pierre Chapuis cf9efb57c8 remove useless torch.no_grad() contexts 2023-09-13 11:14:09 +02:00
Cédric Deltheil eea340c6c4 add support for SDXL IP-Adapter
This only supports the latest SDXL IP-Adapter release (2023.9.8) which
builds upon the ViT-H/14 CLIP image encoder.
2023-09-12 18:00:39 +02:00
Cédric Deltheil 1b4dcebe06 make scheduler an actual abstract base class 2023-09-12 16:47:47 +02:00
Pierre Chapuis 7a32699cc6 add ensure_find and ensure_find_parent helpers 2023-09-12 14:19:10 +02:00
Pierre Chapuis b69dbc4e5c improve CrossAttentionAdapter test 2023-09-12 11:58:24 +02:00
Pierre Chapuis dc2c3e0163 implement CrossAttentionAdapter using chain operations 2023-09-12 11:58:24 +02:00
Pierre Chapuis 3c056e2231 expose lookup_top_adapter 2023-09-12 11:58:24 +02:00
Cédric Deltheil f4e9707297 sdxl test: refreshed reference image
The former one was generated using SDXL 0.9 vs 1.0. The new one has been
generated with diffusers:

    import torch
    from diffusers import StableDiffusionXLPipeline, DDIMScheduler

    noise_scheduler = DDIMScheduler(
        num_train_timesteps=1000,
        beta_start=0.00085,
        beta_end=0.012,
        beta_schedule="scaled_linear",
        clip_sample=False,
        set_alpha_to_one=False,
        steps_offset=1,
    )

    base_model_path = "/path/to/stabilityai/stable-diffusion-xl-base-1.0"

    device = "cuda"
    prompt = "a cute cat, detailed high-quality professional image"
    negative_prompt = "lowres, bad anatomy, bad hands, cropped, worst quality"
    seed = 2

    pipe = StableDiffusionXLPipeline.from_pretrained(base_model_path, scheduler=noise_scheduler, torch_dtype=torch.float16, add_watermarker=False)
    pipe = pipe.to(device)
    generator = torch.Generator(device).manual_seed(seed)
    images = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=30, generator=generator).images
2023-09-12 10:59:26 +02:00
Cédric Deltheil 0e38928c8d sdxl test: add missing torch no_grad 2023-09-12 10:59:26 +02:00
Cédric Deltheil 32cba1afd8 test_sdxl_double_encoder: use proper weights 2023-09-11 21:49:24 +02:00
Pierre Chapuis be54cfc016 fix weight loading for float16 LoRAs 2023-09-11 16:14:19 +02:00
Pierre Chapuis dd0cca5855 use float32 reference for textual inversion (fixes tests on CPU) 2023-09-11 16:11:53 +02:00
Cédric Deltheil e5425e2968 make IP-Adapter generic for SD1 and SDXL 2023-09-08 16:38:01 +02:00
limiteinductive 2786117469 implement SDXL + e2e test on random init 2023-09-07 18:34:42 +02:00
Pierre Chapuis 78e69c7da0 fix typo + skip test if weights are not available 2023-09-07 17:31:20 +02:00
Pierre Chapuis d9a461e9b5 stop relying on SDXL 0.9 weights in test 2023-09-07 12:18:38 +02:00
Pierre Chapuis d54a38ae07 do not hardcode a CUDA device in tests 2023-09-06 19:33:48 +02:00
Cédric Deltheil c55917e293 add IP-Adapter support for SD 1.5
Official repo: https://github.com/tencent-ailab/IP-Adapter
2023-09-06 15:12:48 +02:00
Cédric Deltheil d4dd45fd4d use Module's load_from_safetensors
Instead of manual calls to load_state_dict
2023-09-06 15:06:51 +02:00
Pierre Chapuis 4388968ad3 Update tests/e2e/test_diffusion.py
Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
2023-09-06 11:49:55 +02:00
Pierre Chapuis 547a73e67a clarify the "adapting when a LoRA is injected" issue in tests 2023-09-06 11:49:55 +02:00
Pierre Chapuis 864937a776 support injecting several LoRAs simultaneously 2023-09-06 11:49:55 +02:00
limiteinductive 88efa117bf fix model comparison with custom layers 2023-09-05 12:34:38 +02:00
Cédric Deltheil b933fabf31 unet: get rid of clip_embedding attribute for SD1
It is implicitly defined by the underlying cross-attention layer. This
also makes it consistent with SDXL.
2023-09-01 19:23:33 +02:00
Pierre Chapuis e91e31ebd2 check no two controlnets have the same name 2023-09-01 17:47:29 +02:00
Pierre Chapuis d389d11a06 make basic adapters a part of Fluxion 2023-09-01 17:29:48 +02:00
Pierre Chapuis 31785f2059 scope range adapter in latent diffusion 2023-09-01 17:29:48 +02:00
Pierre Chapuis 73813310d0 rename SelfAttentionInjection to ReferenceOnlyControl and vice-versa 2023-09-01 17:29:48 +02:00
Doryan Kaced 9f6733de8e Add concepts learning via textual inversion 2023-08-31 16:07:53 +02:00
Pierre Chapuis 0f476ea18b make high-level adapters Adapters
This generalizes the Adapter abstraction to higher-level
constructs such as high-level LoRA (targeting e.g. the
SD UNet), ControlNet and Reference-Only Control.

Some adapters now work by adapting child models with
"sub-adapters" that they inject / eject when needed.
2023-08-31 10:57:18 +02:00
Cédric Deltheil 7dc2e93cff tests: add test for clip image encoder
This covers a CLIPImageEncoderH model (Stable Diffusion v2-1-unclip)
specifically
2023-08-30 21:50:01 +02:00
Pierre Chapuis 32c1cfdbb1 add black to CI 2023-08-30 14:50:03 +02:00
Pierre Chapuis 18c84c7b72 shorter import paths 2023-08-29 16:57:40 +02:00
limiteinductive 7ca6bd0ccd implement the ConvertModule class and refactor conversion scripts 2023-08-28 14:39:14 +02:00