shithub: opus

Info  •  Files  •  Log  •  Branches

054d984b – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/16 13:20

Freeze LPCs when quantizing e2e models

8cdc8081 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/10/08 23:20

Fix non-128 batch sizes

37c9bd8d – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/04 23:00

Making sure we recompile when updating the model

2a131ff6 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/04 11:59

Add lpcnet_compute_single_frame_features() to header

be4f70bc – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/04 11:51

split model downloading away from autogen.sh

82c31b4c – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/03 23:24

update instructions

2275853a – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/03 23:14

Merge branch 'exp_quant_grub2'

444b4370 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/03 23:12

New model with quantized GRU B

4a7ce81f – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/03 23:04

Re-enabling quantized feature dump

c5a17a07 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/10/03 22:53

Hard quantization for training

3b8d64d7 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/09/15 12:02

single-frame features for -features option

e4b4613d – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/21 18:35

Fix signed-unsigned biases

51ef273e – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/21 12:38

Using 8-bit recurrent weights for GRU B

8783ef00 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/09/01 08:31

Same as 47, without the sign augmentation

9776e8e8 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/16 11:20

Refactoring frame rate network

2c9b8474 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/17 14:17

noise tuning, flip sign of speech signal

80751bad – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/13 22:06

frame-wise features

4c6a5e0e – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/13 11:36

oops, fix single-frame pitch

6b4e3c56 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/13 08:43

WIP: single-frame inference

66c29fb6 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/12 20:20

Remove some useless buffers

adc50cab – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/04 10:56

dump_lpcnet.py should work the same for end2end

7331e17e – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/04 10:03

Don't remove silence from training data

c5364153 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/04 10:02

Add more training options

ab9a0926 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/24 14:09

Sharing conditioning network with LPC

c1532559 – Krishna Subramani <subramani.krishna97@gmail.com> authored and Jean-Marc Valin <jmvalin@amazon.com> committed on 2021/07/28 23:36

Adds end-to-end LPC training

cba0ecd4 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/02 14:37

Fix warnings about ignoring fread() return value

6ea726d4 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored and Jean-Marc Valin <jmvalin@amazon.com> committed on 2021/07/24 14:25

Avoiding feature copies

4f2caa35 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/08/02 15:01

Update model

65858432 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored and Jean-Marc Valin <jmvalin@amazon.com> committed on 2021/07/22 23:26

Removing the unused features

b90729b8 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/20 13:01

dump_lpcnet.py now checks the size of GRU B

4322c163 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/20 11:36

Oops, actually use the size of GRU B for training

346a96fa – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/19 14:25

Training options for sparse GRU B

c76756e1 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/17 22:24

Adding sparse training for GRU B inputs

8bdbbfa1 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/15 23:07

Support for sparse GRU B input matrices

4c0e2248 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/15 12:12

Model update

c74330e8 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/15 12:06

Pre-compute GRU B conditioning

0d53fad5 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/14 09:47

Using np.memmap() to load the training data

5a51e2ee – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/12 23:09

Adding command-line options to training script

1edf5d79 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/07/10 23:46

README.md update

4298f2f9 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/07/10 23:36

Adding support for SSE2 and SSSE3

116bcb38 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/07/10 10:08

Adding SSE 4.1 for older platforms

3e223e60 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/09 19:16

Fixes Python inference for the binary probability tree

f8f12e7f – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/08 12:38

NEON float->char conversion (same as the AVX2 version)

a1079c2c – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/08 09:20

Again, same conversion as 3206cec, for NEON

7d8b00f1 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/07 23:33

Sampling directly from the logit

e8f70128 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 15:02

same conversion cleanup as 3206cec for sgemv_accum8x4()

7cef98ec – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 14:45

Minor optimization: merging all 3 embeddings

714380e7 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 14:26

More manual unrolling

00655603 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 14:17

Cleaning up the sparse GRU

44fe0556 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 13:59

cleanup float<->int conversions

60d6eab6 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 13:33

Doing a bit of unrolling to speed things up

3e7ab9ff – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 13:31

update model

54abdb6f – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/06 13:05

Sparse matrix indexing optimization

2681822c – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/05 09:31

update model

d3321008 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/07/03 10:54

Representing output pdf as binary probability tree

c151fc18 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/30 14:56

Merge branch 'exp_improved_simd2'

f0ce4338 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/30 14:54

Update test_lpcnet.py, remove old TF1 code

d428b0d3 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/30 14:27

Update model

8c4b88cf – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/30 14:14

Using a bisection search for sampling

e35441f2 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/29 00:05

Faster activation functions for AVX

5571ef1b – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/25 21:27

minor optimization: removing some copying

d61f7e00 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/25 09:43

Fix missing transpose in the sparity code

ca0a43be – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/06/24 13:47

Update README.md

c1535c8c – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/06/24 13:31

Adding option to disable int8 dot products

0b9f6bab – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/06/20 21:34

Remove unnecessary mask in exp() approximation

ae2ae5ea – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/06/20 21:30

Remove useless multiply by one

c7ba313a – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/18 13:39

Adding extra constraint to avoid saturation for SSE/AVX2

237245f8 – Jean-Marc Valin <jmvalin@amazon.com> authored on 2021/06/12 23:50

Support for multi-GPU training

ebc9483b – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/31 20:07

update model

79980b20 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/17 21:13

Minor update to training scripts

20fea538 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/17 16:39

more reasonable noise

b9c230b3 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/01 21:07

Add NEON intrinsics

b214e684 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/01 20:09

Neon WIP: Compiles but very slow

8c3fe6f3 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/30 11:04

Cleaning up float version

40b9fd0a – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/30 10:57

Fix some quantization issues

83657d0e – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/29 22:07

Dot product AVX2 code for non-sparse multiply

1707b960 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/29 20:11

cleanup, add signed-unsigned biases

40b309d9 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/28 21:35

WIP: 8-bit SIMD for GRU B

e695355b – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/26 21:00

some cleanup

06489b42 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/26 13:14

oops, fix number of columns

d87f9744 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/26 13:10

Vectorizing conversion

6b582edb – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/25 23:01

WIP: remove scalar code from AVX2 code

be392e38 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/25 22:58

WIP: Got some AVX2 code working

2b4652f9 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/25 22:20

WIP: cleanup

bce77988 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/25 21:15

WIP: signed*unsigned arithmetic

11736ca9 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/23 22:47

WIP: 8-bit mul

1657bae0 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/23 21:50

WIP: Adding a constraint

c045702e – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/23 21:46

Add non-dot-product AVX code

73a05f55 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/22 11:52

wip 8x4

cc285186 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/19 14:25

wip 8x4 sparseness

8e405b44 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/15 20:05

Improve accuracy of AVX sigmoid

56d9f13e – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2021/01/09 22:15

Fix quantization bug where pitch can get too low

f0df3e82 – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/18 19:16

Update to model trained with Tensorflow 2

078d90cb – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/12/18 19:20

Merge Tensorflow 2 code

90fec91b – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/08/19 10:27

Convert training code to Tensorflow 2

171b1ba0 – Marcus Asteborg <maastebo@microsoft.com> authored and Jean-Marc Valin <jmvalin@jmvalin.ca> committed on 2020/08/02 18:09

Initialize excitation memory value to unquantized 0

88a7878f – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/08/14 22:18

Fix pitch-related bugs

9a7f3e4a – Marcus Asteborg <maastebo@microsoft.com> authored and Jean-Marc Valin <jmvalin@jmvalin.ca> committed on 2020/08/01 18:20

Open files in binary mode to enable dump_data to work properly on Windows

eeb7615b – Marcus Asteborg <maastebo@microsoft.com> authored and Jean-Marc Valin <jmvalin@jmvalin.ca> committed on 2020/07/30 10:25

Add conda env file with working tensorflow and keras version for LPCNet

14fb264a – Jean-Marc Valin <jmvalin@jmvalin.ca> authored on 2020/06/20 19:47

Fix sampling bug for 16-bit rand()