summaryrefslogtreecommitdiffstats
path: root/training
Commit message (Expand)AuthorAgeFilesLines
...
* FixVolpeon2023-03-221-1/+1
* Log DAdam/DAdan dVolpeon2023-03-211-0/+14
* Fixed SNR weighting, re-enabled xformersVolpeon2023-03-212-21/+84
* Restore min SNRVolpeon2023-03-191-13/+12
* New loss weighting from arxiv.org:2204.00227Volpeon2023-03-181-2/+5
* Better SNR weightingVolpeon2023-03-181-3/+3
* Fixed snr weight calculationVolpeon2023-03-171-1/+4
* Fix loss=nanVolpeon2023-03-171-2/+2
* Test: https://arxiv.org/pdf/2303.09556.pdfVolpeon2023-03-171-3/+8
* UpdateVolpeon2023-03-072-15/+16
* UpdateVolpeon2023-03-061-16/+0
* Added Perlin noise to trainingVolpeon2023-03-041-0/+17
* Removed offset noise from training, added init offset to pipelineVolpeon2023-03-031-10/+2
* Implemented different noise offsetVolpeon2023-03-032-22/+10
* Low freq noise with randomized strengthVolpeon2023-03-031-1/+8
* Better low freq noiseVolpeon2023-03-021-1/+1
* Changed low freq noiseVolpeon2023-03-011-23/+10
* UpdateVolpeon2023-03-014-28/+32
* Fixed TI normalization orderVolpeon2023-02-213-15/+19
* FixVolpeon2023-02-211-6/+3
* Don't rely on Accelerate for gradient accumulationVolpeon2023-02-212-30/+29
* Embedding normalization: Ignore tensors with grad = 0Volpeon2023-02-212-6/+16
* UpdateVolpeon2023-02-181-7/+14
* Added Lion optimizerVolpeon2023-02-171-4/+5
* Remove xformers, switch to Pytorch NightlyVolpeon2023-02-174-9/+8
* FixVolpeon2023-02-161-4/+2
* Integrated WIP UniPC schedulerVolpeon2023-02-161-8/+22
* UpdateVolpeon2023-02-151-1/+1
* Made low-freq noise configurableVolpeon2023-02-141-6/+11
* Better noise generation during training: https://www.crosslabs.org/blog/diffu...Volpeon2023-02-131-0/+7
* UpdateVolpeon2023-02-133-3/+3
* Fixed Lora trainingVolpeon2023-02-081-18/+5
* Fix Lora memory usageVolpeon2023-02-074-9/+3
* Add LoraVolpeon2023-02-074-37/+214
* Restored LR finderVolpeon2023-01-206-393/+82
* Move Accelerator preparation into strategyVolpeon2023-01-193-16/+48
* UpdateVolpeon2023-01-174-14/+19
* FixVolpeon2023-01-171-4/+5
* FixVolpeon2023-01-171-1/+0
* Make embedding decay work like Adam decayVolpeon2023-01-171-9/+5
* UpdateVolpeon2023-01-172-8/+8
* UpdateVolpeon2023-01-174-21/+38
* Training updateVolpeon2023-01-163-12/+15
* Moved multi-TI code from Dreambooth to TI scriptVolpeon2023-01-161-3/+14
* More training adjustmentsVolpeon2023-01-163-8/+9
* Handle empty validation datasetVolpeon2023-01-163-47/+58
* Extended Dreambooth: Train TI tokens separatelyVolpeon2023-01-161-0/+1
* Implemented extended Dreambooth trainingVolpeon2023-01-161-3/+4
* Added Dreambooth strategyVolpeon2023-01-151-0/+183
* Restored functional trainerVolpeon2023-01-152-27/+83