Commit message (Collapse) | Author | Age | Files | Lines | ||
---|---|---|---|---|---|---|
... | ||||||
* | Unified training script structure | Volpeon | 2023-01-13 | 2 | -130/+84 | |
| | ||||||
* | Reverted modularization mostly | Volpeon | 2023-01-13 | 7 | -613/+458 | |
| | ||||||
* | More modularization | Volpeon | 2023-01-13 | 9 | -653/+677 | |
| | ||||||
* | Simplified step calculations | Volpeon | 2023-01-13 | 2 | -33/+33 | |
| | ||||||
* | Removed PromptProcessor, modularized training loop | Volpeon | 2023-01-13 | 9 | -293/+334 | |
| | ||||||
* | Added TI decay start offset | Volpeon | 2023-01-13 | 2 | -3/+9 | |
| | ||||||
* | Code deduplication | Volpeon | 2023-01-13 | 6 | -146/+149 | |
| | ||||||
* | Update | Volpeon | 2023-01-12 | 3 | -34/+63 | |
| | ||||||
* | Fixed TI decay | Volpeon | 2023-01-12 | 2 | -9/+12 | |
| | ||||||
* | Disable Adam weight decay | Volpeon | 2023-01-12 | 1 | -1/+1 | |
| | ||||||
* | Fix | Volpeon | 2023-01-11 | 2 | -5/+5 | |
| | ||||||
* | Heck | Volpeon | 2023-01-11 | 1 | -1/+1 | |
| | ||||||
* | TI: Use grad clipping from LoRA #104 | Volpeon | 2023-01-11 | 4 | -12/+15 | |
| | ||||||
* | Better defaults | Volpeon | 2023-01-10 | 2 | -7/+6 | |
| | ||||||
* | Fix | Volpeon | 2023-01-10 | 2 | -2/+2 | |
| | ||||||
* | Added arg to disable tag shuffling | Volpeon | 2023-01-10 | 4 | -18/+37 | |
| | ||||||
* | Enable buckets for validation, fixed vaildation repeat arg | Volpeon | 2023-01-09 | 3 | -10/+5 | |
| | ||||||
* | Add --valid_set_repeat | Volpeon | 2023-01-09 | 3 | -1/+37 | |
| | ||||||
* | No cache after all | Volpeon | 2023-01-08 | 1 | -17/+7 | |
| | ||||||
* | Cache token IDs in dataset | Volpeon | 2023-01-08 | 1 | -8/+20 | |
| | ||||||
* | Fix | Volpeon | 2023-01-08 | 2 | -6/+6 | |
| | ||||||
* | Improved aspect ratio bucketing | Volpeon | 2023-01-08 | 3 | -4/+61 | |
| | ||||||
* | Fixed aspect ratio bucketing | Volpeon | 2023-01-08 | 1 | -3/+5 | |
| | ||||||
* | Cleanup | Volpeon | 2023-01-08 | 1 | -15/+25 | |
| | ||||||
* | Fixed aspect ratio bucketing; allow passing token IDs to pipeline | Volpeon | 2023-01-08 | 5 | -68/+102 | |
| | ||||||
* | Improved aspect ratio bucketing | Volpeon | 2023-01-08 | 4 | -223/+237 | |
| | ||||||
* | Cleanup | Volpeon | 2023-01-07 | 3 | -178/+201 | |
| | ||||||
* | Made aspect ratio bucketing configurable | Volpeon | 2023-01-07 | 3 | -18/+37 | |
| | ||||||
* | Added progressive aspect ratio bucketing | Volpeon | 2023-01-07 | 5 | -126/+151 | |
| | ||||||
* | Update | Volpeon | 2023-01-07 | 3 | -11/+6 | |
| | ||||||
* | Fix | Volpeon | 2023-01-06 | 1 | -2/+4 | |
| | ||||||
* | Relaxed ema_power for TI | Volpeon | 2023-01-06 | 1 | -1/+1 | |
| | ||||||
* | Update | Volpeon | 2023-01-06 | 4 | -33/+26 | |
| | ||||||
* | Use context manager for EMA, on_train/eval hooks | Volpeon | 2023-01-06 | 3 | -81/+92 | |
| | ||||||
* | Package update | Volpeon | 2023-01-06 | 1 | -1/+1 | |
| | ||||||
* | Add prompt template argument to inference | Volpeon | 2023-01-06 | 1 | -0/+7 | |
| | ||||||
* | Add contextmanager to EMAModel to apply weights temporarily | Volpeon | 2023-01-06 | 2 | -33/+36 | |
| | ||||||
* | Log EMA decay | Volpeon | 2023-01-05 | 1 | -1/+2 | |
| | ||||||
* | Added EMA to TI | Volpeon | 2023-01-05 | 4 | -10/+157 | |
| | ||||||
* | Fix LR finder | Volpeon | 2023-01-05 | 1 | -7/+23 | |
| | ||||||
* | Update | Volpeon | 2023-01-05 | 1 | -1/+1 | |
| | ||||||
* | Update | Volpeon | 2023-01-05 | 7 | -51/+75 | |
| | ||||||
* | Fix | Volpeon | 2023-01-05 | 2 | -10/+13 | |
| | ||||||
* | Various cleanups | Volpeon | 2023-01-05 | 8 | -1091/+133 | |
| | ||||||
* | Update | Volpeon | 2023-01-04 | 6 | -9/+15 | |
| | ||||||
* | Various updates | Volpeon | 2023-01-04 | 5 | -34/+87 | |
| | ||||||
* | Better eval generator | Volpeon | 2023-01-04 | 3 | -13/+13 | |
| | ||||||
* | Fixed reproducibility, more consistant validation | Volpeon | 2023-01-04 | 5 | -43/+113 | |
| | ||||||
* | Don't use vector_dropout by default | Volpeon | 2023-01-03 | 2 | -2/+2 | |
| | ||||||
* | Added vector dropout | Volpeon | 2023-01-03 | 4 | -15/+69 | |
| |