Commit Graph

135 Commits

Author SHA1 Message Date
Damian Stewart 1ab27a9059 better check for null manual_data_root 2023-06-17 11:04:14 +02:00
Damian Stewart dd6b37840d permit empty manual_data_root 2023-06-17 11:04:14 +02:00
Damian Stewart 6c60b76fb6 log combined loss if there are >1 val subsets 2023-06-17 11:04:14 +02:00
Damian Stewart 403f7ddf07 fix chunked_shuffle crash with empty list 2023-06-17 11:04:14 +02:00
Damian Stewart ded73f088f cleanup 2023-06-17 11:04:14 +02:00
Damian Stewart 8885d58efd create bucket if it doesn't exist 2023-06-17 11:04:14 +02:00
Damian Stewart bd89ad96d2 typo 2023-06-17 11:04:14 +02:00
Damian Stewart e4872fdc0c improve shuffle and runt handling for named buckets 2023-06-17 11:04:14 +02:00
Damian Stewart a047294676 assert an important precondition 2023-06-08 11:01:16 +02:00
Damian Stewart 4f98f0bcc9 ensure predictable shuffle behaviour and further cleanup 2023-06-08 10:50:51 +02:00
Damian Stewart 1874a38663 cleanup 2023-06-08 10:39:32 +02:00
Damian Stewart b3c5d656e3 refactor for clarity 2023-06-08 10:31:10 +02:00
damian 86f80a8776 allow empty default_batch 2023-06-07 18:39:13 +02:00
damian 59fc9891d4 shuffle named batches while respecting and accounting for grad_accum 2023-06-07 18:07:37 +02:00
Damian Stewart ba95b8c6d1 simplify runt handling 2023-06-05 01:04:21 +02:00
Damian Stewart 53d0686086 add a batch_id.txt file to subfolders or a batch_id key to local yaml to force images for that folder to be processed in the same batch 2023-06-05 01:02:27 +02:00
Victor Hall a96c6e2166 dadapt stuff 2023-06-03 11:26:53 -04:00
Victor Hall 9ee2effacd add 320 res 2023-06-01 21:19:20 -04:00
Victor Hall 97f1160496 correct image cropping 2023-06-01 21:06:55 -04:00
Victor Hall c0e7c4adf9 errant print 2023-06-01 19:23:43 -04:00
Victor Hall 615fa929e5 errant print 2023-06-01 19:17:43 -04:00
Victor Hall 0e0b546ef7 help debug crop error someone got 2023-06-01 19:11:22 -04:00
Victor Hall 56deb26a59 early work on shuffle_tags.txt and add try around trimming 2023-06-01 16:18:21 -04:00
MFAlex 7dcfa7acbf Speed up preloading by not loading pixel data
PIL lazy loads pixel data, so image size can be accessed without loading the full image.
This commit makes it so only image size and EXIF metadata are fetched from disk, speeding up the preload stage while still supporting transposing images.
2023-05-31 20:48:48 -04:00
Victor Hall 5c98cdee70 update to torch2, xformers 20, bnb 381 2023-05-30 22:15:02 -04:00
Damian Stewart f3468fe7e7 Merge remote-tracking branch 'upstream/main' into val_partial_epochs 2023-05-07 02:11:58 +02:00
damian 8e28a74386 add migration code for 2023-04-29 16:55:02 -04:00
damian 360eaa5a42 don't print, using logging 2023-04-29 16:55:02 -04:00
damian 413f981512 implement extra manual validation splits 2023-04-29 16:55:02 -04:00
Damian Stewart aad00eab2e switch up errored item logic flow 2023-04-29 13:03:10 -04:00
Damian Stewart ce85ce30ae warn on chronically underfilled aspect ratio buckets 2023-04-29 13:03:10 -04:00
Victor Hall 630ac5e46a
Merge pull request #152 from qslug/yaml-tag-str
Always convert YAML tags to strings
2023-04-22 19:03:04 -04:00
Augusto de la Torre 2e8056174d Always convert YAML tags to strings 2023-04-23 00:36:32 +02:00
Victor Hall e3e30a5599 fix exif issues 2023-04-16 14:53:34 -04:00
Victor Hall 3697b16344 put exif transpose back in preloading to fix bug 2023-04-16 02:19:27 -04:00
tyler b8f637873f removing the convert_rgb flag for a simpler design 2023-04-14 20:21:48 -05:00
tyler a839180199 set # of data loaders by the min of batch size or cpu count, do not do an rg b conversion when only loading image metadata 2023-04-14 14:59:28 -05:00
Augusto de la Torre 2bb35eaa0a Use filenames for caption if no main prompt in yaml 2023-04-14 00:59:26 +02:00
Augusto de la Torre dd98ebe080 Add error handling for bad exif, and rotate before sizing 2023-03-28 19:19:08 +02:00
Victor Hall 56256ab9ef attempt to catch some errors reported on github 2023-03-26 11:52:49 -04:00
Victor Hall 35d52b56e0 added some resolutions, option for val-loss pos-neg, fix wandb 2023-03-25 20:09:06 -04:00
Augusto de la Torre cdafa2dc43 Avoid unnecessary iteration over files in folder 2023-03-21 12:09:52 +01:00
Augusto de la Torre 161e0a563c Prioritize tags, `image > local > global`, but respect weights 2023-03-21 00:15:53 +01:00
Augusto de la Torre fae0b3c535 Retain original tag order when parsing captions 2023-03-19 23:30:42 +01:00
Victor Hall 5afd75fd98 try-except around preloading items to help troubleshoot issues 2023-03-18 22:27:09 -04:00
Victor Hall f01a7354f0 remove sort from dataset due to slowdown on large sets, add contribution readme 2023-03-18 22:24:03 -04:00
Augusto de la Torre 48f132554c Assign default value for MAX_CAPTION_LENGTH 2023-03-15 21:52:32 +01:00
Victor Hall ba687de8b4 add pbar back to preloading, remove cruft from testing loss stuff 2023-03-15 12:06:29 -04:00
Victor Hall 605716a646 conf 2023-03-15 11:22:40 -04:00
Victor Hall da3c183cc5 autofix exif orientation 2023-03-15 11:12:59 -04:00