Finegrain/refiners
1
0
Fork 0
mirror of https://github.com/finegrain-ai/refiners.git synced 2024-11-23 14:48:45 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
649 commits 35 branches 5 tags 149 MiB
Commit graph

374 commits

Author SHA1 Message Date
limiteinductive f32ccc3474 Remove seed_everything logging because it is too verbose 2024-04-22 18:14:33 +02:00
limiteinductive 5dde281ada Implement EventConfig 2024-04-22 18:14:33 +02:00
limiteinductive 446796da57 Refactor TimeValue 2024-04-18 20:58:47 +02:00
Laurent 17246708b9 Add sample_noise staticmethod and modify add_noise to support batched steps 2024-04-18 12:55:49 +02:00
limiteinductive 7427c171f6 fix training_utils requirements check 2024-04-17 18:10:28 +02:00
Pierre Colle bf7852b88e SAM: image_to_scaled_tensor gray images 2024-04-16 18:45:17 +02:00
Laurent eb4bb34f8b (training_utils) add new ForceCommit callback 2024-04-16 14:43:10 +02:00
limiteinductive be7d065a33 Add DataloadeConfig to Trainer 2024-04-15 20:56:19 +02:00
limiteinductive b9b999ccfe turn scoped_seed into a context manager 2024-04-13 15:03:35 +02:00
Pierre Colle 64692c3b5b TrainerClock: assert dataset_length >= batch_size 2024-04-12 15:05:52 +02:00
Pierre Colle 0ac290f67d SAM: expose sizing helpers 2024-04-12 08:56:23 +02:00
Laurent 06ff2f0a5f add support for dinov2 giant flavors 2024-04-11 14:48:33 +02:00
Laurent 04e59bf3d9 fix GLU Activation docstrings 2024-04-11 14:48:33 +02:00
limiteinductive f26b6ee00a add static typing to __call__ method for latent_diffusion models ; fix multi_diffusion bug that wasn't taking guidance_scale into account 2024-04-11 12:13:30 +02:00
Cédric Deltheil a2ee705783 hq sam: add constructor args to docstring 
Additionally, mark `register_adapter_module` for internal use.
 2024-04-08 11:46:37 +02:00
Pierre Colle d05ebb8dd3 SAM/HQSAMAdapter: docstring examples 2024-04-08 07:12:57 +02:00
hugojarkoff bbb46e3fc7 Fix clock step inconsistencies on batch end 2024-04-05 15:52:43 +02:00
Pierre Chapuis 09af570b23 add DINOv2-FD metric 2024-04-03 16:45:00 +02:00
Laurent 5f07fa9c21 fix dinov2 interpolation, support batching 2024-04-02 18:57:25 +02:00
Pierre Chapuis fd5a15c7e0 update pyright and fix Pillow 10.3 typing issues 2024-04-02 18:15:52 +02:00
Laurent 4f94dfb494 implement dinov2 positional embedding interpolation 2024-04-02 10:02:43 +02:00
Laurent 0336bc78b5 simplify interpolate function and layer 2024-04-02 10:02:43 +02:00
Pierre Colle 6c37e3f933 hq-sam : weights/load_weights 2024-03-29 11:25:43 +01:00
Pierre Chapuis 404a15aad2 tweak auto_attach_loras so debugging is easier when it fails 2024-03-26 16:12:48 +01:00
Laurent a0715806d2 modify ip_adapter's ImageCrossAttention scale getter and setter 
this new version makes it robust in case mulitple Mulitply-s are inside the Chain (e.g. if the Linear layers are LoRA-ified)
 2024-03-26 11:15:04 +01:00
Laurent 7e64ba4011 modify ip_adapter's CrossAttentionAdapters injection logic 2024-03-26 11:15:04 +01:00
Cédric Deltheil df0cc2aeb8 do not call __getattr__ with keyword argument 
Same for __setattr__. Use positional arguments instead. E.g.:

    import torch
    import refiners.fluxion.layers as fl
    m = torch.compile(fl.Linear(1,1))
    m(torch.zeros(1))
    # TypeError: Module.__getattr__() got an unexpected keyword argument 'name'
 2024-03-25 21:46:13 +01:00
Pierre Colle cba83b0558 SAM init with mask_decoder after #325 2024-03-24 20:18:57 +01:00
Pierre Chapuis 364e196874 support no CFG in compute_clip_text_embedding 2024-03-22 17:06:51 +01:00
Pierre Colle 94e8b9c23f SAM MaskDecoder token slicing 2024-03-22 13:11:40 +01:00
hugojarkoff a93ceff752 Add HQ-SAM Adapter 2024-03-21 15:36:55 +01:00
limiteinductive 38c86f59f4 Switch gradient clipping to native torch torch.nn.utils.clip_grad_norm_ 2024-03-19 22:08:48 +01:00
Pierre Colle 68fe725767 Add multimask_output flag to SAM 2024-03-19 17:40:26 +01:00
limiteinductive 6a72943ff7 change TimeValue to a dataclass 2024-03-19 14:49:24 +01:00
Laurent b8fae60d38 make LoRA's weight initialization overridable 2024-03-13 17:32:16 +01:00
Pierre Chapuis e32d8d16f0 LoRA loading: forward exclusions when preprocessing parts of the UNet 2024-03-13 15:25:00 +01:00
limiteinductive ff5341c85c Change weight decay for Optimizer to normal PyTorch default 2024-03-12 15:20:21 +01:00
Pierre Chapuis 975560165c improve docstrings 2024-03-08 15:43:57 +01:00
Pierre Chapuis cd5fa97c20 ability to get LoRA weights in SDLoraManager 2024-03-08 15:43:57 +01:00
Pierre Chapuis fb90b00e75 add_loras_to_unet: add preprocess values as exclusions in last step 2024-03-08 15:43:57 +01:00
Pierre Chapuis 4259261f17 simplify LCM weights loader using new manager features 2024-03-08 15:43:57 +01:00
Pierre Chapuis ccd9414ff1 fix debug map when attaching two LoRAs 
(in that case return the path of the LoraAdapter)
 2024-03-08 15:43:57 +01:00
Pierre Chapuis 8c7fcbc00f LoRA manager: move exclude / include to add_loras call 
Always exclude the TimestepEncoder by default.
This is because some keys include both e.g. `resnet` and `time_emb_proj`.

Preprocess blocks that tend to mix up with others in a separate
auto_attach call.
 2024-03-08 15:43:57 +01:00
Pierre Chapuis 052a20b897 remove add_multiple_loras 2024-03-08 15:43:57 +01:00
Pierre Chapuis ed8ec26e63 allow passing inclusions and exlusions to SDLoraManager 2024-03-08 15:43:57 +01:00
Pierre Chapuis cce2a98fa6 add sanity check to auto_attach_loras 2024-03-08 15:43:57 +01:00
Pierre Chapuis 1eb71077aa use same scale setter / getter interface for all controls 2024-03-08 11:29:28 +01:00
Laurent 5e7986ef08 adding more log messages in training_utils 2024-03-08 10:52:14 +01:00
Pierre Chapuis be2368cf20 ruff 3 formatting (Rye 0.28) 2024-03-08 10:42:05 +01:00
Pierre Chapuis 91d1b46aa9 Add a note that we mitigate non-zero SNR in DDIM. 2024-02-26 12:14:02 +01:00
Pierre Chapuis 7f51d18045 clarify that add_lcm_lora can load SDXL-Lightning 2024-02-26 12:14:02 +01:00
Pierre Chapuis 7e4e0f0650 correctly scale init latents for Euler scheduler 2024-02-26 12:14:02 +01:00
Pierre Chapuis bf0ba58541 refactor solver params, add sample prediction type 2024-02-26 12:14:02 +01:00
Pierre Chapuis ddc1cf8ca7 refactor solvers to support different timestep spacings 2024-02-26 12:14:02 +01:00
Cédric Deltheil 176807740b control_lora: fix adapter set scale 
The adapter set scale did not propagate the scale to the underlying
zero convolutions. The value set at CTOR time was used instead.

Follow up of #285
 2024-02-22 10:01:05 +01:00
Pierre Chapuis 03b79d6d34 rename ResidualBlock to ConditionScaleBlock in LCM 2024-02-21 16:37:27 +01:00
Pierre Chapuis 684e2b9a47 add docstrings for LCM / LCM-LoRA 2024-02-21 16:37:27 +01:00
Pierre Chapuis 12b6829a26 add support for LCM LoRA weights loading 2024-02-21 16:37:27 +01:00
Pierre Chapuis fafe5f8f5a Improve filtering when auto-attaching LoRAs. 
Also support debug output to help diagnose bad mappings.
 2024-02-21 16:37:27 +01:00
Pierre Chapuis f8d55ccb20 add LcmAdapter 
This adds support for the condition scale embedding.
Also updates the UNet converter to support LCM.
 2024-02-21 16:37:27 +01:00
Pierre Chapuis c8c6294550 add LCMSolver (Latent Consistency Models) 2024-02-21 16:37:27 +01:00
Pierre Chapuis 4a619e84f0 support disabling CFG in LatentDiffusionModel 2024-02-21 16:37:27 +01:00
Pierre Colle d199cd4f24 batch sdxl + sd1 + compute_clip_text_embedding 
Co-authored-by: Cédric Deltheil <355031+deltheil@users.noreply.github.com>
 2024-02-21 15:17:11 +01:00
Cédric Deltheil 5ab5d7fd1c import ControlLoraAdapter part of latent_diffusion 2024-02-19 14:11:32 +01:00
Laurent 2a3e353f04 enable StyleAligned related docstrings in mkdocstrings 2024-02-15 15:22:47 +01:00
Laurent efa3988638 implement StyleAlignedAdapter 2024-02-15 15:22:47 +01:00
limiteinductive 432e32f94f rename Scheduler -> LRScheduler 2024-02-15 11:48:36 +01:00
Laurent 684303230d export ControlLora and ControlLoraAdapter in refiners.foundationals.latent_diffusion.stable_diffusion_xl 2024-02-15 11:32:49 +01:00
Laurent 41a5ce2052 implement ControlLora and ControlLoraAdapter 2024-02-14 18:20:46 +01:00
Laurent a54808e757 add context_key getter and setter to RangeAdapter2d 2024-02-14 18:20:46 +01:00
Laurent 35b6e2f7c5 add context_key getter and setter to TimestepEncoder 2024-02-14 18:20:46 +01:00
Laurent 0230971543 simplify is_compatible in lora.py 2024-02-14 18:20:46 +01:00
Pierre Chapuis 35868ba34b Move helper to attach several LoRAs from SD to Fluxion 2024-02-14 13:35:46 +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 0caa72a082 remove deprecated on_checkpoint_save 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 Chapuis 402d3105b4 support multiple IP adapter inputs as tensor 2024-02-09 17:16:17 +01:00
Cédric Deltheil 5a7085bb3a training_utils/config.py: inline type alias 
Follow up of #227
 2024-02-09 14:36:22 +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
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
Cédric Deltheil e36dda63fd fix miscellaneous typos 2024-02-07 17:51:25 +01:00
Pierre Chapuis 396d166564 make pad method private 2024-02-07 17:47:14 +01:00
Pierre Chapuis 4d85918336 Update src/refiners/foundationals/latent_diffusion/lora.py 
Co-authored-by: Laureηt <laurent@lagon.tech>
 2024-02-07 17:47:14 +01:00
Pierre Chapuis b1c200c63a Update src/refiners/foundationals/latent_diffusion/lora.py 
Co-authored-by: Laureηt <laurent@lagon.tech>
 2024-02-07 17:47:14 +01:00
Pierre Chapuis eb9abefe07 add a few comments in SDLoraManager 2024-02-07 17:47:14 +01:00
Benjamin Trom bbe0759151 fix docstring 2024-02-07 16:13:01 +01:00
limiteinductive 2e526d35d1 Make Dataset part of the trainer 2024-02-07 16:13:01 +01:00
Laurent 9883f24f9a (fluxion/layers) remove View layer 
+ replace existing `View` layers by `Reshape`
 2024-02-07 12:06:07 +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 98fce82853 fix 37425fb609 
Things to understand:

- subscripted generic basic types (e.g. `list[int]`) are types.GenericAlias;
- subscripted generic classes are `typing._GenericAlias`;
- neither can be used with `isinstance()`;
- get_origin is the cleanest way to check for this.
 2024-02-06 13:49:37 +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