rockerBOO
4f27c6a0c9
Add BPO, CPO, DDO, SDPO, SimPO
...
Refactor Preference Optimization
Refactor preference dataset
Add iterator support for ImageInfo and ImageSetInfo
- Supporting iterating through either ImageInfo or ImageSetInfo to
clean up preference dataset implementation and support 2 or more
images more cleanly without needing to duplicate code
Add tests for all PO functions
Add metrics for process_batch
Add losses for gradient manipulation of loss parts
Add normalizing gradient for stabilizing gradients
Args added:
mapo_beta = 0.05
cpo_beta = 0.1
bpo_beta = 0.1
bpo_lambda = 0.2
sdpo_beta = 0.02
simpo_gamma_beta_ratio = 0.25
simpo_beta = 2.0
simpo_smoothing = 0.0
simpo_loss_type = "sigmoid"
ddo_alpha = 4.0
ddo_beta = 0.05
2025-06-03 15:09:48 -04:00
rockerBOO
d8716a9cb9
Rework DDO loss
2025-05-02 02:07:53 -04:00
rockerBOO
22447ebc76
Use mean, use ddo_loss
2025-04-30 19:46:44 -04:00
rockerBOO
9a2101a040
Add DDO loss
2025-04-30 03:34:19 -04:00
rockerBOO
8e8243a423
Add DDO preference optimization
2025-04-28 22:37:44 -04:00
rockerBOO
d22c827544
Update PO cached latents, move out functions, update calls
2025-04-27 17:38:50 -04:00
rockerBOO
74529743d4
Merge branch 'sd3' into po
2025-04-14 17:30:54 -04:00
Kohya S
5a18a03ffc
Merge branch 'dev' into sd3
2025-04-07 21:55:17 +09:00
Kohya S
d0b5c0e5cf
chore: formatting, add TODO comment
2025-03-30 21:15:37 +09:00
Kohya S.
59d98e45a9
Merge pull request #1974 from rockerBOO/lora-ggpo
...
Add LoRA-GGPO for Flux
2025-03-30 21:07:31 +09:00
Kohya S.
93a4efabb5
Merge branch 'sd3' into resize-interpolation
2025-03-30 19:30:56 +09:00
DKnight54
381303d64f
Update train_network.py
2025-03-29 02:26:18 +08:00
rockerBOO
0181b7a042
Remove progress bar avg norms
2025-03-27 03:28:33 -04:00
rockerBOO
3647d065b5
Cache weight norms estimate on initialization. Move to update norms every step
2025-03-18 14:25:09 -04:00
rockerBOO
ea53290f62
Add LoRA-GGPO for Flux
2025-03-06 00:00:38 -05:00
Kohya S
1fcac98280
Merge branch 'sd3' into val-loss-improvement
2025-02-26 21:09:10 +09:00
Kohya S.
6e90c0f86c
Merge pull request #1909 from rockerBOO/progress_bar
...
Move progress bar to account for sampling image first
2025-02-24 18:57:44 +09:00
Kohya S
efb2a128cd
fix wandb val logging
2025-02-21 22:07:35 +09:00
rockerBOO
ca1c129ffd
Fix metadata
2025-02-19 14:20:40 -05:00
rockerBOO
7729c4c8f9
Add metadata
2025-02-19 14:20:40 -05:00
Kohya S
4a36996134
modify log step calculation
2025-02-18 22:05:08 +09:00
Kohya S
dc7d5fb459
Merge branch 'sd3' into val-loss-improvement
2025-02-18 21:34:30 +09:00
rockerBOO
4671e23778
Fix validation epoch loss to check epoch average
2025-02-16 01:42:44 -05:00
Kohya S
63337d9fe4
Merge branch 'sd3' into val-loss-improvement
2025-02-15 21:41:07 +09:00
rockerBOO
ab88b431b0
Fix validation epoch divergence
2025-02-14 11:14:38 -05:00
Kohya S
76b761943b
fix: simplify validation step condition in NetworkTrainer
2025-02-11 21:53:57 +09:00
Kohya S
177203818a
fix: unpause training progress bar after vaidation
2025-02-11 21:42:46 +09:00
Kohya S
344845b429
fix: validation with block swap
2025-02-09 21:25:40 +09:00
Kohya S
0911683717
set python random state
2025-02-09 20:53:49 +09:00
Kohya S
c5b803ce94
rng state management: Implement functions to get and set RNG states for consistent validation
2025-02-04 21:59:09 +09:00
rockerBOO
de830b8941
Move progress bar to account for sampling image first
2025-01-29 00:02:45 -05:00
rockerBOO
9637ce0c23
Preference optimization with MaPO and Diffusion-DPO
2025-01-28 13:31:42 -05:00
Kohya S
45ec02b2a8
use same noise for every validation
2025-01-27 22:10:38 +09:00
Kohya S
0778dd9b1d
fix Text Encoder only LoRA training
2025-01-27 22:03:42 +09:00
Kohya S
0750859133
validation: Implement timestep-based validation processing
2025-01-27 21:56:59 +09:00
Kohya S
29f31d005f
add network.train()/eval() for validation
2025-01-27 21:35:43 +09:00
Kohya S
b6a3093216
call optimizer eval/train fn before/after validation
2025-01-27 21:22:11 +09:00
Kohya S
86a2f3fd26
Fix gradient handling when Text Encoders are trained
2025-01-27 21:10:52 +09:00
Kohya S
532f5c58a6
formatting
2025-01-27 20:50:42 +09:00
rockerBOO
c04e5dfe92
Fix loss recorder on 0. Fix validation for cached runs. Assert on validation dataset
2025-01-23 09:57:24 -05:00
rockerBOO
25929dd0d7
Remove Validating... print to fix output layout
2025-01-12 15:38:57 -05:00
rockerBOO
ee9265cf26
Fix validate_every_n_steps for gradient accumulation
2025-01-12 14:56:35 -05:00
rockerBOO
0456858992
Fix validate_every_n_steps always running first step
2025-01-12 14:47:49 -05:00
rockerBOO
2bbb40ce51
Fix regularization images with validation
...
Adding metadata recording for validation arguments
Add comments about the validation split for clarity of intention
2025-01-12 14:29:50 -05:00
rockerBOO
4c61adc996
Add divergence to logs
...
Divergence is the difference between training and validation to
allow a clear value to indicate the difference between the two
in the logs.
2025-01-12 13:18:26 -05:00
rockerBOO
1e61392cf2
Revert bucket_reso_steps to correct 64
2025-01-08 18:43:26 -05:00
rockerBOO
556f3f1696
Fix documentation, remove unused function, fix bucket reso for sd1.5, fix multiple datasets
2025-01-08 13:41:15 -05:00
rockerBOO
1231f5114c
Remove unused train_util code, fix accelerate.log for wandb, add init_trackers library code
2025-01-07 22:31:41 -05:00
rockerBOO
742bee9738
Set validation steps in multiple lines for readability
2025-01-06 17:34:23 -05:00
rockerBOO
fcb2ff010c
Clean up some validation help documentation
2025-01-06 11:39:32 -05:00
