summaryrefslogtreecommitdiffstats
Commit message (Expand)AuthorAgeFilesLines
* UpdateVolpeon2023-01-064-33/+26
* Use context manager for EMA, on_train/eval hooksVolpeon2023-01-063-81/+92
* Package updateVolpeon2023-01-061-1/+1
* Add prompt template argument to inferenceVolpeon2023-01-061-0/+7
* Add contextmanager to EMAModel to apply weights temporarilyVolpeon2023-01-062-33/+36
* Log EMA decayVolpeon2023-01-051-1/+2
* Added EMA to TIVolpeon2023-01-054-10/+157
* Fix LR finderVolpeon2023-01-051-7/+23
* UpdateVolpeon2023-01-051-1/+1
* UpdateVolpeon2023-01-057-51/+75
* FixVolpeon2023-01-052-10/+13
* Various cleanupsVolpeon2023-01-058-1091/+133
* UpdateVolpeon2023-01-046-9/+15
* Various updatesVolpeon2023-01-045-34/+87
* Better eval generatorVolpeon2023-01-043-13/+13
* Fixed reproducibility, more consistant validationVolpeon2023-01-045-43/+113
* Don't use vector_dropout by defaultVolpeon2023-01-032-2/+2
* Added vector dropoutVolpeon2023-01-034-15/+69
* Fixed LR finderVolpeon2023-01-023-22/+26
* UpdateVolpeon2023-01-023-9/+55
* FixVolpeon2023-01-022-8/+2
* Save args before training, tooVolpeon2023-01-022-0/+8
* FixVolpeon2023-01-022-4/+2
* UpdateVolpeon2023-01-022-13/+12
* FixVolpeon2023-01-021-1/+1
* Improved one cycle schedulerVolpeon2023-01-022-31/+63
* UpdateVolpeon2023-01-013-15/+15
* CleanupVolpeon2023-01-013-14/+5
* Fix MultiCLIPTokenizer (forgot to override encode)Volpeon2023-01-011-13/+20
* UpdatesVolpeon2023-01-016-147/+227
* Fixed accuracy calc, other improvementsVolpeon2023-01-015-60/+74
* FixVolpeon2023-01-011-1/+1
* Better token shufflingVolpeon2023-01-011-2/+3
* FixVolpeon2022-12-311-5/+7
* UpdateVolpeon2022-12-314-26/+35
* Bugfixes for multi-vector token handlingVolpeon2022-12-314-27/+53
* Simplified multi-vector embedding codeVolpeon2022-12-313-17/+14
* FixesVolpeon2022-12-312-3/+2
* Added multi-vector embeddingsVolpeon2022-12-318-81/+299
* Misc improvementsVolpeon2022-12-303-78/+56
* Training script improvementsVolpeon2022-12-305-25/+89
* UpdateVolpeon2022-12-292-11/+16
* Training improvementsVolpeon2022-12-293-29/+42
* Updated 1-cycle schedulerVolpeon2022-12-282-9/+15
* Integrated updates from diffusersVolpeon2022-12-285-39/+59
* Improved learning rate finderVolpeon2022-12-273-23/+30
* Added validation phase to learn rate finderVolpeon2022-12-272-15/+31
* Added learning rate finderVolpeon2022-12-274-162/+257
* Set default dimensions to 768; add config inheritanceVolpeon2022-12-267-26/+36
* Code simplifications, avoid autocastVolpeon2022-12-254-82/+92