summaryrefslogtreecommitdiffstats
Commit message (Collapse)AuthorAgeFilesLines
* Fixed Lora trainingVolpeon2023-02-084-37/+35
|
* Fix Lora memory usageVolpeon2023-02-075-11/+5
|
* Add LoraVolpeon2023-02-0710-93/+819
|
* Restored LR finderVolpeon2023-01-209-397/+111
|
* Move Accelerator preparation into strategyVolpeon2023-01-194-19/+51
|
* UpdateVolpeon2023-01-175-22/+25
|
* Optimized embedding normalizationVolpeon2023-01-171-5/+2
|
* Smaller emb decayVolpeon2023-01-171-1/+1
|
* FixVolpeon2023-01-171-4/+5
|
* FixVolpeon2023-01-171-1/+0
|
* Make embedding decay work like Adam decayVolpeon2023-01-172-21/+9
|
* UpdateVolpeon2023-01-174-9/+12
|
* UpdateVolpeon2023-01-176-73/+104
|
* Training updateVolpeon2023-01-165-16/+25
|
* If valid set size is 0, re-use one image from train setVolpeon2023-01-162-6/+2
|
* Moved multi-TI code from Dreambooth to TI scriptVolpeon2023-01-164-244/+131
|
* More training adjustmentsVolpeon2023-01-166-43/+101
|
* Pad dataset if len(items) < batch_sizeVolpeon2023-01-162-20/+23
|
* Handle empty validation datasetVolpeon2023-01-166-76/+91
|
* Extended Dreambooth: Train TI tokens separatelyVolpeon2023-01-163-71/+84
|
* Implemented extended Dreambooth trainingVolpeon2023-01-164-372/+200
|
* Added Dreambooth strategyVolpeon2023-01-152-23/+206
|
* Restored functional trainerVolpeon2023-01-155-104/+112
|
* Fixed Conda envVolpeon2023-01-151-2/+4
|
* UpdateVolpeon2023-01-155-162/+106
|
* Removed unused code, put training callbacks in dataclassVolpeon2023-01-157-1470/+40
|
* Added functional TI strategyVolpeon2023-01-153-78/+312
|
* Added functional trainerVolpeon2023-01-153-37/+101
|
* UpdateVolpeon2023-01-146-127/+33
|
* UpdateVolpeon2023-01-144-15/+15
|
* WIP: Modularization ("free(): invalid pointer" my ass)Volpeon2023-01-1411-286/+1541
|
* TI: Prepare UNet with Accelerate as wellVolpeon2023-01-143-39/+41
|
* FixVolpeon2023-01-144-6/+6
|
* CleanupVolpeon2023-01-147-131/+103
|
* Unified training script structureVolpeon2023-01-132-130/+84
|
* Reverted modularization mostlyVolpeon2023-01-137-613/+458
|
* More modularizationVolpeon2023-01-139-653/+677
|
* Simplified step calculationsVolpeon2023-01-132-33/+33
|
* Removed PromptProcessor, modularized training loopVolpeon2023-01-139-293/+334
|
* Added TI decay start offsetVolpeon2023-01-132-3/+9
|
* Code deduplicationVolpeon2023-01-136-146/+149
|
* UpdateVolpeon2023-01-123-34/+63
|
* Fixed TI decayVolpeon2023-01-122-9/+12
|
* Disable Adam weight decayVolpeon2023-01-121-1/+1
|
* FixVolpeon2023-01-112-5/+5
|
* HeckVolpeon2023-01-111-1/+1
|
* TI: Use grad clipping from LoRA #104Volpeon2023-01-114-12/+15
|
* Better defaultsVolpeon2023-01-102-7/+6
|
* FixVolpeon2023-01-102-2/+2
|
* Added arg to disable tag shufflingVolpeon2023-01-104-18/+37
|