b1d7a7ff5d
Show percentage when calculating vectors
2018-08-09 23:59:26 -05:00
7bded13de4
Change vectorizer to Sonopy
...
Also adds vectorizer parameter to choose between mels and mfccs
2018-08-09 23:58:21 -05:00
d71ea3fee0
Add visual element to precise-listen
...
The previous default `precise-listen` is not `precise-listen -l`
2018-08-09 11:36:10 -05:00
8cfbd41d8c
Fix Python 3.4 compatibility in train-incremental
2018-08-07 16:27:56 -05:00
fd2b71b069
Replace Python 3.5+ syntax
...
Otherwise it errors on Python 3.4
2018-08-06 16:54:26 -05:00
7b802be7b8
Move Trainer into precise.scripts.train
...
Not having it there was inconsistent because the other trainer classes are in the scripts package within the respective scripts
2018-07-26 13:19:10 -05:00
9642fcc07f
Inject params in precise-simulate
2018-07-25 10:06:14 -05:00
81a8b18986
Removed redundant code ( #20 )
...
Removed a redundant line of code. print("False positive: ", false_positives * 100, "%") is already printed so print(false_positives) isn't needed.
2018-07-16 15:22:00 -05:00
75e0769026
Changed the maximum possible amount of units ( #19 )
...
Changed the maximum possible amount of units for the optimization to return to 70 units.
2018-07-16 11:57:43 -05:00
d6db6c8ec2
Make train_optimize inherit from Trainer
2018-07-12 21:22:08 -05:00
7f915f939d
Remove unused variables
2018-07-12 17:21:54 -05:00
3549765a84
Move module level code into main function
2018-07-12 17:21:54 -05:00
4d1f212c2b
Remove environment specific definition
2018-07-12 17:21:54 -05:00
c5dd7c891d
Autoformat and remove unused imports
2018-07-12 17:21:54 -05:00
5818a101da
Add license header
2018-07-12 17:21:54 -05:00
4f5df1b997
Make imports absolute
2018-07-12 17:21:54 -05:00
ac43e4b6bb
Add .py extension to train_optimize
...
Also rename to train_optimize to be consistent with verbiage
2018-07-12 17:21:54 -05:00
01dc86896d
Created train_optimizer.py
2018-07-12 17:21:54 -05:00
7d97cd2331
Add simulate script
2018-07-12 16:57:02 -05:00
d32fb6fee3
Fix extra comma on assignment
2018-07-12 14:21:54 -05:00
36804075b3
Fix Nan in false_neg and false_pos metrics
...
This was caused when small batches with no positive samples caused a divide-by-zero
2018-07-11 19:34:22 -05:00
d8dd0f3240
Shuffle training and testing data before training
...
This fixes issues with small batch sizes and metrics
2018-07-11 19:32:39 -05:00
06d9669e79
Add sampled training support
2018-07-11 12:45:41 -05:00
1f5e17ba0a
Formatting
2018-07-11 12:45:41 -05:00
6a8b7b8ea8
Add delta vector support
2018-07-11 12:45:41 -05:00
ebd5e09feb
Rework train script to use new Trainer class
...
This allows IncrementalTrainer to share similar behavior
2018-07-11 12:45:41 -05:00
4da65d1f37
Quiet tensorflow when performing predictions
2018-07-11 12:45:41 -05:00
a946de21ce
Make ListenerParams proper class via attr
2018-07-11 12:45:41 -05:00
779ddd2d02
Use vectorize_raw instead of rewriting mfcc call
2018-07-11 12:45:41 -05:00
0da11b33bd
Make mfcc generation use mfcc_size instead of feature_size
...
feature_size could be different.
2018-07-11 12:45:41 -05:00
603385291f
Fix up docstring for create_model
2018-07-11 12:45:41 -05:00
252abbeb15
Make loss bias use module level variable
2018-07-11 12:45:41 -05:00
1f15f9da92
Increase default batch size in train-incremental
2018-06-25 12:37:56 -05:00
c663719cce
Increment version to 0.2.0
2018-06-20 16:49:50 -05:00
523eb719f1
listen: not-wake-up: increase the random value's range
...
the old range is too narrow and therefore make data override often happen.
Signed-off-by: Wu Zhangjin <wuzhangjin@gmail.com>
2018-06-20 16:49:50 -05:00
3add0c320e
activate: add a running directory independent noitfier
...
Signed-off-by: Wu Zhangjin <wuzhangjin@gmail.com>
2018-06-20 16:49:50 -05:00
b14c4b34b8
activate: use a right audio player for mac osx
...
Signed-off-by: Wu Zhangjin <wuzhangjin@gmail.com>
2018-06-13 14:33:58 -05:00
e3e15f65e7
Don't output params warning when it doesn't make sense
2018-06-12 13:58:31 -05:00
989e29bb83
Support pb files in precise-test
2018-04-26 19:04:05 -05:00
0cbb98fa78
Increase default weight of weighted_log_loss to 0.7
...
No point in using it by default if it isn't weighted. This makes it reduce false positives by default
2018-04-18 23:11:54 -05:00
a3274b8d33
Fix major bug in precise-train-incremental
...
Previously it wouldn't update its own buffer so the files it would save would be mostly silence
2018-04-18 17:52:09 -05:00
755af134f6
Use K.mean instead of K.sum in loss
...
This makes the loss actually reasonable rather than something like 132.3
2018-04-18 17:26:14 -05:00
476728b138
Add warning if loading from folder with many wavs without specifying tags file
2018-04-18 16:09:46 -05:00
45c72a80c2
Add rename folder and tags arguments
...
Everything is now referred to as tags rather than db since it's a set of tags not a database. It also switches the positional argument to refer to the regular structured data folder, with a separate tags-folder option to override where to load the file ids from the tags from
2018-04-18 15:59:02 -05:00
19bf49a4ae
Rename precise-record to precise-listen
...
This just makes a lot more sense. It's not recording you; it's listening to you
2018-04-18 14:48:43 -05:00
bb4cb4badd
Fix double slash when loading from folder ending in slash
2018-04-17 20:04:02 -05:00
cb00e03d2d
Fix chunk size parameter in precise-record
2018-04-17 18:33:08 -05:00
9418990ff1
Don't train when activation subsides in incremental training
...
This allows batching up larger amounts of data before retraining
2018-04-16 17:31:53 -05:00
fbdca29949
Fix database groups file when in subdirectory and formatting
2018-04-12 18:21:24 -05:00
befd296128
FIx vectorizing invalid files
2018-03-21 16:34:53 -05:00
1ee402fd54
Fix database files not loading properly
2018-03-21 16:34:53 -05:00
dcad96fc4e
Fix invalid chunk size in PreciseRunner and fix importing from db
2018-03-21 16:34:53 -05:00
0fe1bceede
Change database to simple text file
2018-03-21 16:34:53 -05:00
09e8b91ad2
Add try except clauses to make more robust
2018-03-21 16:34:53 -05:00
b5c38ee71a
Clean up setup script
...
Now uses a configurable VENV location used for parent projects
2018-03-21 16:34:53 -05:00
2878e8f963
Refactor test statistics, change buffer_size to mean bytes, and
...
add pocketsphinx calculation to precise-eval
2018-03-21 16:34:53 -05:00
32a001511e
Refactor test statistics, change buffer_size to mean bytes, and
...
add pocketsphinx calculation to precise-eval
2018-03-07 13:37:03 -06:00
98e0e7161c
Output model being processed in precise-eval
2018-03-06 15:15:03 -06:00
09c3ad927b
Change license to Apache 2.0
2018-02-28 20:49:34 -06:00
b254fe1358
Rename precise-stream to precise-engine
...
It makes way more sense
2018-02-26 21:19:30 -06:00
8ced324465
Rename Tensorflow to TensorFlow
2018-02-23 15:23:20 -06:00
88173fef95
Recursive glob data/random
2018-02-23 14:45:31 -06:00
988a4fe60c
Add extra option to precise-collect
2018-02-23 14:38:04 -06:00
a3d434de08
Fix default channels
2018-02-23 14:23:49 -06:00
ce5f93369c
New precise-collect script and minor fixes
...
- Fix precise-convert script
- Fix pyinstaller spec
- Change keyword folder to wake-word. This is more consistent with other labels
2018-02-23 13:47:56 -06:00
cc32959e29
Remove system path modifications
...
To run a script during development, instead ./setup.sh followed by precise-<script_name> should be used
This ensures all the proper paths are already setup (such as for PreciseRunner)
2018-02-22 01:10:41 -06:00
d622eef3ec
Support only loading test data without train data
2018-02-22 01:02:50 -06:00
13c3d81993
Add precise-eval script
2018-02-21 23:16:54 -06:00
484a90f9bc
Refacetor and source code standardization
2018-02-21 13:15:51 -06:00
2a3ff7dc26
Standardize imports and type hints
2018-02-15 14:54:08 -06:00
d2a5469bf3
Improvements
2018-02-15 03:03:52 -06:00
e3cb1f0bf4
Make training fail gracefully if dimensions are invalid
2017-12-18 18:51:03 -06:00
b8f393fc53
Add sound on activation to precise-record
2017-12-18 18:49:19 -06:00
5d925dae59
Make record.py use precise wrapper
2017-11-28 14:58:37 -06:00
cb31c79e86
Fix bytearray bug
2017-11-28 14:57:54 -06:00
9ceb23e34e
Add copyright header
2017-11-28 14:57:54 -06:00
76a36ac712
Clean up imports
2017-11-21 16:51:37 -06:00
9463b0babf
Add type hints
2017-11-21 16:30:22 -06:00
cd9a08a605
Add weighted log loss option
2017-11-21 14:12:02 -06:00
2741b7afc3
Add more arguments to train_feedback.py
2017-11-14 19:28:24 -06:00
5e3b2594d4
Increase dropout and change loss to crossentropy
...
These changes have been found to improve the quality of the trained models
2017-11-14 19:27:48 -06:00
de8037dffa
Add functions to generate inhibitory data to reduce false positives
...
In practice, this actually increases false positives
2017-11-14 19:26:45 -06:00
61be0f6884
Increment version to 0.1.1
2017-11-10 16:41:44 -06:00
6b2e0cc303
Fix typo preventing params from loading properly
2017-11-10 16:40:45 -06:00
7622194691
Decrease chunk size to increase accuracy
2017-11-10 16:40:15 -06:00
d970a41879
Add argument parser to precise-record
2017-11-10 16:39:50 -06:00
1b01d16386
Standardize arguments
2017-11-08 21:47:00 -06:00
5ce56ff7e7
Add version to precise-stream and fix inconsistencies
2017-11-03 14:39:52 -05:00
98da7805b6
Increase compatibility
...
glob recursive isn't available in Python 3.4 and JSONDecodeError is only on some json implementations and inherits from ValueError
2017-11-03 14:10:36 -05:00
431843a016
Add RNN wakeword listener
2017-11-03 00:50:53 -05:00
Matthew Scholefield
Mikhail Dmitrienko
Wu Zhangjin
Matthew D. Scholefield