Commit Graph

307 Commits

Author SHA1 Message Date
Reuben Morais
a05baa35c9 Package training code to avoid sys.path hacks 2020-03-25 21:34:50 +01:00
Tilman Kamp
41da7b2870 Fix #2830 - Support for unlabeled samples 2020-03-24 16:53:02 +01:00
Norman Koch
6ec19b9498 Corrected typo in flags 2020-03-22 20:39:41 +01:00
Alexandre Lissy
7b2a409f9f Converting importers from multiprocessing.dummy to multiprocessing
Fixes #2817
2020-03-18 11:04:36 +01:00
Alexandre Lissy
ce59228824 Localizeable validate_label
Fixes #2804
2020-03-18 11:04:36 +01:00
Alexandre Lissy
f9e05fe0c3 Share argparser amongst importers 2020-03-18 11:04:36 +01:00
Reuben Morais
29a2ac37f0
Merge pull request #2779 from reuben/export-metadata
Write model metadata to export folder unconditionally
2020-03-17 12:08:35 -03:00
Kelly Davis
b57eaa19d6
Merge pull request #2783 from mozilla/optuna
Added optimizer for lm_alpha + lm_beta
2020-03-17 09:58:01 +01:00
Reuben Morais
94cca3c651
Correctly handle non stable versions in --branch 2020-03-12 13:43:27 -03:00
Tilman Kamp
63bc695600 Process pool for audio preparation 2020-03-12 14:34:24 +01:00
Tilman Kamp
6b1d6773de SDB support 2020-03-10 10:32:58 +01:00
Alexandre Lissy
763ed38bae Enforce proper line ending removal when reading alphabet
Fixes #2611
2020-03-06 15:19:56 +01:00
lissyx
0f8291df71
Proper arguments ordering 2020-03-06 10:15:18 +01:00
Alexandre Lissy
61fa1ad428 Show actual alphabet path in error message 2020-03-06 10:13:18 +01:00
Alexandre Lissy
639a68d2ae Ensure sample rate comparison with proper types
Fixes #2798
2020-02-28 11:16:34 +01:00
kdavis-mozilla
561131a05c Added optimizer for lm_alpha + lm_beta 2020-02-26 11:41:22 +01:00
Reuben Morais
669aa497cc Address review comments 2020-02-25 11:52:48 +01:00
Reuben Morais
1f1f5a98e4
Merge pull request #2781 from rhamnett/patch-2
Add flag to force reinitialisation of learning rate after lr_plateau
2020-02-24 16:08:13 +01:00
Richard Hamnett
a3268545ab
Update flags.py
change flag datatype to boolean
2020-02-21 19:24:13 +00:00
Reuben Morais
46e7993075
Merge pull request #2771 from reuben/warn-sample-rate
Warn if --audio_sample_rate does not match training sample
2020-02-21 19:46:05 +01:00
Richard Hamnett
5e1f54ae4f
Reset learning rate if force set 2020-02-21 18:33:43 +00:00
Richard Hamnett
0de9e4bf80
Add force_initialize_learning_rate
Ability to reset learning rate which has been reduced by reduce_lr_on_plateau
2020-02-21 18:32:03 +00:00
Reuben Morais
48178005a2 Write model metadata to export folder unconditionally 2020-02-21 12:46:44 +01:00
Reuben Morais
4291db7309 Handle graph without learning rate variable for export case 2020-02-20 15:40:35 +01:00
Reuben Morais
0b82c751db
Merge pull request #2770 from reuben/lr_reduction_rebased
Reduce learning rate on plateau
2020-02-18 22:20:54 +01:00
Reuben Morais
1178215423 Warn if --audio_sample_rate does not match training sample
In PR #2688, we started specifying the upper frequency limit when computing Mfccs.
This value was computed as half of the --audio_sample_rate value. Despite accepting
a variable sample rate input for the Mfcc computation, the TensorFlow OP only takes
a constant upper frequency limit, so we can't pass a dynamic value computed from each
sample to the op.

This means we lost the ability to transparently train on data with multiple sample
rates. This commit adds a warning message in case a training sample does not match
the --audio_sample_rate flag.
2020-02-18 18:15:01 +01:00
Reuben Morais
78e8dfdf38 Disable early stopping and LR reduction on plateau by default 2020-02-18 16:18:37 +01:00
Reuben Morais
6e12b7caed Allow missing learning rate variable in older checkpoints 2020-02-18 16:18:31 +01:00
Daniel
17ddc5600e Reduce learning rate on plateau. 2020-02-18 16:17:51 +01:00
Alexandre Lissy
020619fa97 Automagically decompress GZipped artifacts
Fixes #2760
2020-02-18 13:59:30 +01:00
Reuben Morais
0e2f34f8cf Synchronize TensorFlow logging with --log_level flag 2020-02-17 12:46:00 +01:00
Reuben Morais
f32fd7a33f Add transfer learning test 2020-02-17 08:29:10 +01:00
JRMeyer
5bba9ea5d1 Transfer-learning support 2020-02-17 08:29:10 +01:00
PedroDKE
50c5e5b23a
Make the webrtcvad dependency optional (#2754) 2020-02-13 17:16:37 +01:00
Reuben Morais
3637f88c06 Fix CI errors, address comments, update examples 2020-02-12 10:13:02 +01:00
Reuben Morais
8e9b6ef7b3 Embed default beam width into exported graph and remove param from DS_CreateModel 2020-02-11 21:23:05 +01:00
Reuben Morais
1e2eb96248 Update all API consumers 2020-02-11 19:44:36 +01:00
Reuben Morais
49fe54e0e3
Merge pull request #2688 from mozilla/mfcc-upper-frequency
Specify upper frequency limit when computing Mfccs
2020-02-11 10:26:09 +01:00
Daniel
726cc20586 Rename dataset param. 2020-02-06 14:47:59 +01:00
Daniel
8cc91fafb2 Moved summary printing to samples printing. 2020-02-06 14:44:41 +01:00
Daniel
de92142986 Named example sections. 2020-02-06 13:31:30 +01:00
Daniel
4186cbef88 Reverse ordered loss again. 2020-02-06 12:50:20 +01:00
Daniel
9ec88b7f28 Add whitespace again. 2020-02-06 11:41:40 +01:00
Daniel
63a07e6834 Added summary to evaluate_tflite.py and moved method to evaluate_tools.py. 2020-02-06 11:39:31 +01:00
Daniel
369e3c9fc3 Revert linebreak. 2020-02-06 11:01:37 +01:00
Daniel
272ed99d24 Add median examples. Fix sorting. 2020-02-06 10:55:48 +01:00
Daniel
a2f05ccabe Print best and worst results in a WER report. 2020-02-05 17:52:10 +01:00
Alexandre Lissy
ff401732a3 Enforce CTC decoder version check
Fix #2710
2020-02-03 10:24:21 +01:00
Reuben Morais
50830d7022 Fix whitespace 2020-01-27 15:59:39 +01:00
PedroDKE
3e349497ed added an argument to choose the final export model name 2020-01-25 12:30:07 +01:00