Commit Graph

1473 Commits

Author SHA1 Message Date
Aleksandras Kostarevas
5ecdafd60b Ensure conv graph measurement never undermeasures memory with dynamic audio ctx 2024-03-30 18:45:54 -05:00
Aleksandras Kostarevas
601d6df6b3 Cancel native inference early 2024-03-21 18:32:54 -05:00
Aleksandras Kostarevas
434a751d63 Fix modified utf-8 errors when returning strings 2024-03-21 16:49:45 -05:00
Aleksandras Kostarevas
c101772317 Add suppressNonSpeechTokens 2024-03-18 16:24:16 -05:00
Aleksandras Kostarevas
38055fae65 Add other workaround 2024-03-13 16:11:52 -05:00
Aleksandras Kostarevas
350b8e8fcf Add bad word filtering and blacklisting 2024-03-13 13:31:51 -05:00
Aleksandras Kostarevas
9fed68c03a Fix segfault when no results / only 1 result 2024-03-07 14:56:21 +02:00
Aleksandras Kostarevas
69649256f8 Fix comma key to open settings menu 2024-03-05 15:54:18 +02:00
Aleksandras Kostarevas
c57a3d83af Add personal dictionary glossary for voice input and keyboard 2024-03-05 15:24:30 +02:00
Aleksandras Kostarevas
42ac255a81 Sync whisper.cpp changes from voice input 2024-03-05 11:06:24 +02:00
Aleksandras Kostarevas
e4d41567b0 Update hyperparameter 2024-02-20 21:04:07 +02:00
Aleksandras Kostarevas
6453c15a21 Merge branch 'lm-2-finetuning-whisperggml' into 'model-metadata'
Add autocorrect threshold to model-metadata branch

See merge request alex/latinime!6
2024-02-03 15:18:27 +00:00
Aleksandras Kostarevas
c7113297fb Add radio selection for threshold 2024-02-01 21:55:56 +02:00
Aleksandras Kostarevas
a111164bb8 Improve algorithm in a few ways:
* If the first letter is capital, only capitalized first tokens will be sampled. If the whole text is capitalized, then only fully capital tokens will be sampled for the whole word
* If a word is an exact match, it gets boosted relative to others
* Probability threshold for autocorrect is now 18.0
* Add "clueless" threshold, if it's less than 1.3 then just show the user's typed word in the middle instead.
2024-01-30 20:30:44 +02:00
Aleksandras Kostarevas
f888ba3353 Fix finetuning, add a finetuning screen, handle errors during importing model, update metadata format, add model exporting 2024-01-30 17:14:02 +02:00
Aleksandras Kostarevas
5bf4492634 Fix embedded tokenizer loading, implement new model management methods, implement model info loading, model importing 2024-01-28 22:40:39 +02:00
Aleksandras Kostarevas
0021b6aa04 Model metadata and manager component 2024-01-24 01:03:16 +02:00
Aleksandras Kostarevas
7aea41eede Add decomposeTapPosition fallback for nonintersecting cases 2024-01-22 08:21:27 +02:00
Aleksandras Kostarevas
5e0722c984 Fix issue with apostrophe token being banned 2024-01-22 08:20:55 +02:00
Aleksandras Kostarevas
dbad61d2e6 Fix non-English dictionary prediction 2024-01-16 21:02:55 +02:00
Aleksandras Kostarevas
55d5959f54 Skip non-alphabetic characters during mixing 2024-01-09 18:25:14 +02:00
Aleksandras Kostarevas
ebb70b9c12 Fix build, disable gesture input pending model update 2023-12-19 20:28:58 +02:00
Aleksandras Kostarevas
4e9e86d871 Implement multimodal position encoder 2023-12-19 20:02:20 +02:00
Aleksandras Kostarevas
314cf8c84c Type out whisper.cpp result 2023-12-05 18:06:12 +00:00
Aleksandras Kostarevas
7075c22179 Add key embedding mixing 2023-12-04 20:09:51 +00:00
Aleksandras Kostarevas
4f15ff4a73 Add experimental swipe typing 2023-11-28 17:01:58 +00:00
Aleksandras Kostarevas
854e1295cc Fix problem with n_tokens==0 2023-11-28 16:20:33 +00:00
abb128
ca9c9d5a9a ggml backend v2 2023-11-25 09:39:04 +02:00
abb128
f31db527d6 Add whisper.cpp 2023-11-25 09:13:50 +02:00
Aleksandras Kostarevas
001f1eb442 Disable debug flag 2023-11-21 20:33:21 +02:00
Aleksandras Kostarevas
cb2edca601 Update training hyperparameters 2023-11-21 17:07:43 +02:00
Aleksandras Kostarevas
14fcb55565 Save LoRA-merged model after training 2023-11-14 20:40:00 +02:00
Aleksandras Kostarevas
2409eecef5 Loss/progress training callbacks 2023-11-14 18:11:00 +02:00
Aleksandras Kostarevas
b53a46b18d Move training to CoroutineWorker 2023-11-14 17:23:08 +02:00
Aleksandras Kostarevas
38b06d7909 History logging and training based on log 2023-11-14 11:43:36 +02:00
Aleksandras Kostarevas
0e0876f06c Revise training 2023-11-13 16:42:01 +02:00
Aleksandras Kostarevas
ee8a81f12c Initial fine-tuning 2023-11-07 16:48:48 +02:00
Aleksandras Kostarevas
5778cd15a0 Update ggml and llama.cpp 2023-11-06 13:41:25 +02:00
Aleksandras Kostarevas
7c4531e32d Fix crashes related to too large context 2023-10-16 18:24:00 +03:00
Aleksandras Kostarevas
c73fe16ddc Add q8 model 2023-10-16 17:31:43 +03:00
Aleksandras Kostarevas
92480fd460 Adjust space probability and mustNotAutocorrect 2023-10-13 18:44:38 +03:00
Aleksandras Kostarevas
c34a411989 Fix infinite prediction loop 2023-10-13 18:34:49 +03:00
Aleksandras Kostarevas
b8539ce88a Initial batched inference using llama_batch 2023-10-10 22:34:04 +03:00
Aleksandras Kostarevas
6d17f00296 Update llama.cpp 2023-10-10 22:32:25 +03:00
Aleksandras Kostarevas
16fdb3629d Add LanguageModel class 2023-09-28 19:42:29 +03:00
Aleksandras Kostarevas
ea0af67ecc Add ggml, sentencepiece and dependencies 2023-09-28 19:41:05 +03:00
abb128
434f8b6b27 Initial working build of fork 2023-07-06 21:57:49 +03:00
Jing Mike
03eef94a8d Remove unused variables
Since some variables with module LatinIME are defined but not used,
when compiled with build combination "sdk_pc_x86_64-userdebug" and
build command "mmm packages/inputmethods/LatinIME", the following
code lines will be reported that "variable 'XXX' set but not used".
(should be similar for all the other build combinations)

Repeated 10 times for each:
terminal_position_lookup_table.cpp:74:9 removedEntryCount
terminal_position_lookup_table.cpp:85:9 removedEntryCount
proximity_info_state_utils.cpp:493:9 tempTime
trie_map.cpp:56:9 unusedRegionSize
suggestion_results.cpp💯9 index

Repeated 80+ times:
proximity_info_utils.h:75:25 proximityChar

With this patch we are removing some of the unused variables and
putting the C++ 17 attribute [[maybe_unused]] to the others which
are used for logging. Then all the related build warnings have been
eliminated.

Test: mmm packages/inputmethods/LatinIME, presubmit check.

Change-Id: Ia66766322d6ae8a010b1cb55cc22993fbc6d012c
Signed-off-by: Jing Mike <jingyangliu@eswincomputing.com>
2023-03-19 10:00:01 +00:00
Bob Badour
f3d9532a32 [LSC] Add LOCAL_LICENSE_KINDS to packages/inputmethods/LatinIME
Added SPDX-license-identifier-Apache-2.0 to:
  Android.bp
  common/Android.bp
  java/Android.bp
  native/dicttoolkit/Android.bp
  native/jni/Android.bp
  tests/Android.bp
  tools/EditTextVariations/Android.bp
  tools/dicttool/Android.bp
  tools/make-keyboard-text/Android.bp

Bug: 68860345
Bug: 151177513
Bug: 151953481

Test: m all

Exempt-From-Owner-Approval: janitorial work
Change-Id: I440008bffac5c97a2497970af377a9d03262b6d8
2021-02-17 09:46:27 -08:00
Treehugger Robot
e04480b68b Merge "Mark liblatinime_unittests as unit_test:true to run in presubmit in CI" 2021-02-17 10:24:20 +00:00