summaryrefslogtreecommitdiffstats
Commit message (Expand)AuthorAgeFilesLines
* TI: Delta learningVolpeon2023-04-034-68/+46
* TI: No tag dropout by defaultVolpeon2023-04-031-1/+1
* Bring back Lion optimizerVolpeon2023-04-034-9/+82
* Fix memory leakVolpeon2023-04-032-5/+5
* Update dataset format: Separate prompt and keywordsVolpeon2023-04-025-48/+46
* UpdateVolpeon2023-04-021-3/+3
* Lora: Only register params with grad to optimizerVolpeon2023-04-023-8/+7
* RevertVolpeon2023-04-014-64/+132
* FixVolpeon2023-04-011-1/+3
* Combined TI with embedding and LoRAVolpeon2023-04-013-87/+38
* Experimental: TI via LoRAVolpeon2023-04-013-63/+44
* UpdateVolpeon2023-04-015-10/+18
* Add support for Adafactor, add TI initializer noiseVolpeon2023-04-016-6/+67
* UpdateVolpeon2023-03-313-3/+9
* UpdateVolpeon2023-03-315-24/+32
* FixVolpeon2023-03-313-5/+6
* FixVolpeon2023-03-311-1/+1
* Support Dadaptation d0, adjust sample freq when steps instead of epochs are usedVolpeon2023-03-313-12/+33
* FixVolpeon2023-03-314-6/+9
* FixVolpeon2023-03-283-3/+3
* Support num_train_steps arg againVolpeon2023-03-284-23/+49
* Fix TIVolpeon2023-03-272-17/+33
* Fix TIVolpeon2023-03-272-26/+19
* Fix TIVolpeon2023-03-272-9/+11
* Revert to regular embeddingsVolpeon2023-03-271-19/+15
* Sparse TI embeddings without sparse tensorsVolpeon2023-03-272-27/+31
* Fix TI embeddings initVolpeon2023-03-261-1/+1
* Improved TI embeddingsVolpeon2023-03-262-9/+24
* Improved inverted tokensVolpeon2023-03-265-34/+70
* Fix training with guidanceVolpeon2023-03-252-6/+14
* UpdateVolpeon2023-03-258-987/+99
* UpdateVolpeon2023-03-244-9/+22
* Fixed Lora training perf issueVolpeon2023-03-242-9/+9
* Lora fix: Save config JSON, tooVolpeon2023-03-241-0/+3
* Refactoring, fixed Lora trainingVolpeon2023-03-245-66/+104
* Bring back Perlin offset noiseVolpeon2023-03-232-1/+21
* UpdateVolpeon2023-03-2310-30/+29
* FixVolpeon2023-03-223-6/+2
* Log DAdam/DAdan dVolpeon2023-03-214-6/+20
* Added dadaptationVolpeon2023-03-214-1/+85
* Fixed SNR weighting, re-enabled xformersVolpeon2023-03-215-50/+97
* 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.pdfVolpeon2023-03-173-39/+56
* UpdateVolpeon2023-03-075-27/+37
* UpdateVolpeon2023-03-063-25/+13
* More flexible pipeline wrt init noiseVolpeon2023-03-041-13/+44