Commit Graph

28 Commits

Author SHA1 Message Date
Reuben Morais
d2d46c3aee
Merge pull request #3113 from mozilla/generate-package-cpp
Rewrite generate_package.py in C++ to avoid training dependencies
2020-07-02 23:13:25 +02:00
Reuben Morais
8f6106b35d Update docs to refer to new generate_scorer_package 2020-06-30 16:47:41 +02:00
Tilman Kamp
a48ebdfde8 Reverb augmentation: Workaround for import problem in scikit-learn dependency of librosa 2020-06-30 14:13:12 +02:00
Reuben Morais
a84abf813c Deduplicate Alphabet implementations, use C++ one everywhere 2020-06-30 09:52:45 +02:00
Tilman Kamp
eebf12134e Warp augmentation 2020-06-29 16:22:31 +02:00
Tilman Kamp
da96d14eaa Fix #3089 - Recreate overlay queue on augmentation restart 2020-06-19 11:10:09 +02:00
Tilman Kamp
4c6245d155
Merge pull request #3055 from tilmankamp/augext
Refactoring of TF based augmentations
2020-06-18 16:13:46 +02:00
Tilman Kamp
a28df45192 Respect None case for augmentations list 2020-06-16 15:46:28 +02:00
Tilman Kamp
5b6de213d8 Follow-up on PR comments; removed warp augmentation; split pitch_and_tempo augmentation 2020-06-16 11:07:57 +02:00
Tilman Kamp
ea21c7d24e
Apply suggestions from code review
Co-authored-by: Reuben Morais <reuben.morais@gmail.com>
2020-06-16 10:22:45 +02:00
Tilman Kamp
d94db7ca43 Refactoring of TF based augmentations 2020-06-10 13:42:45 +02:00
Alexandre Lissy
cfc79799ec Report imported vs total audio time 2020-06-10 13:12:15 +02:00
Reuben Morais
e069b6d61f Add read only validation metrics
For now this is just CTC loss like a validation set, but without affecting
best validation checkpoint tracking logic. Eventually this could compute WER
on a smaller set, for example.
2020-06-08 15:26:37 +02:00
Reuben Morais
28c7f4c35d Only use drop_remainder in dataset for train phase 2020-06-08 10:47:13 +02:00
Marek Grzegorek
8a87d1d100
Fix arguments order in pcm_to_np call
pcm_to_np takes segment buffer as its first argument and audio format as a second one. The wrong order cause "bytes object has no attribute 'channels'' ArgumentError.
2020-06-03 18:05:52 +02:00
Reuben Morais
ab2ba41c7f Convert path to str to fix Python 3.5 compat 2020-05-30 15:16:30 +02:00
Tilman Kamp
96caa2d115 Follow up on PR comments 2020-05-18 14:23:02 +02:00
Tilman Kamp
a5303ccca6 Renamed prepare_samples to augment_samples 2020-05-14 16:50:18 +02:00
Tilman Kamp
c5ceee26dd Live audio augmentation 2020-05-12 10:18:21 +02:00
Tilman Kamp
927859728f Named tuple AudioFormat, parameter re-ordering in util.audio and NP to PCM conversion support 2020-05-12 10:09:15 +02:00
Reuben Morais
b283aadae6 Don't escape non-ASCII characters in test_output_file 2020-04-29 15:45:09 +02:00
Reuben Morais
6f9fcf3029 Embed flag definitions 2020-04-28 13:33:45 +02:00
kdavis-mozilla
f3b4943e18 Updated alpha and beta 2020-04-23 07:19:37 +02:00
Reuben Morais
0c6e90868e Split --load into two to avoid unexpected behavior at evaluation time 2020-04-07 14:24:05 +02:00
Reuben Morais
cc7a0ada46 Only allow graph/layer initialization at start of training 2020-04-07 14:24:05 +02:00
Reuben Morais
c428acf478 Automatically install ds_ctcdecoder in setup.py 2020-03-31 18:47:45 +02:00
Reuben Morais
20b0ab17ea Remove unused GPU usage tools 2020-03-31 13:42:41 +02:00
Reuben Morais
a05baa35c9 Package training code to avoid sys.path hacks 2020-03-25 21:34:50 +01:00