Jean-Marc Valin
7f7b2a1c66
Smaller version of fargan
...
800k parameters, 600 MFLOPS, with a receptive field of 3 feature vectors
2023-11-16 02:06:14 -05:00
Jean-Marc Valin
b0620c0bf9
Using sparse GRUs in DRED decoder
...
Saves ~270 kB of weights in the decoder
2023-11-15 04:08:50 -05:00
Jean-Marc Valin
77594bf158
Dumping RDOVAE stats from XML
2023-11-08 17:32:43 -05:00
Jean-Marc Valin
222662dac8
DRED: quantize scale and dead zone to 8 bits
2023-11-07 18:10:50 -05:00
Jan Buethe
4e104555e9
added weight export script for LACE/NoLACE
2023-11-07 15:12:12 +01:00
Jan Buethe
8af5c6b4a1
added transposed 1d convolutions to wexchange
2023-11-07 11:54:22 +01:00
Jean-Marc Valin
0ab0640d4a
Split stats in two and remove useless dimensions
2023-11-07 00:07:14 -05:00
Jan Buethe
2386a60ec6
updated moc to match results in ietf118 presentation
2023-11-06 17:50:48 +01:00
Jean-Marc Valin
544b3e576c
DRED: quantize r and p0 parameters with 8 bits
...
Only code non-degenerate symbols, which makes the encoder faster
2023-11-06 03:16:43 -05:00
Jan Buethe
da60266f6e
updated moc method
2023-11-02 16:52:50 +01:00
Jean-Marc Valin
ddd5669e79
Pitch and fargan model updates
...
Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.
2023-10-28 23:33:47 -04:00
Jean-Marc Valin
4598fe5409
Quantizing pitchdnn and rdovae weights
2023-10-20 12:54:13 -04:00
Jan Buethe
290be25b98
added 16kHz version of opus_compare in python
2023-10-20 14:24:27 +02:00
Jan Buethe
1accd2472e
finalized quantization option in export_rdovae_weights.py
2023-10-20 14:14:31 +02:00
Jean-Marc Valin
d720955d61
Marking RDOVAE layers to quantize
2023-10-19 16:06:52 -04:00
Jan Buethe
60ac1c6c99
prepared quantization implementation for DRED
2023-10-19 21:54:39 +02:00
Jan Buethe
2192e85b91
restructured osce readme
2023-10-19 21:45:45 +02:00
Jan Buethe
055c683018
added LACE/NoLACE checkpoint URL
2023-10-19 21:34:13 +02:00
Jean-Marc Valin
35cb8d7f66
C implementation of FARGAN
2023-10-14 16:42:33 -04:00
Jean-Marc Valin
9e76a7bfb8
update fargan to match version 45
2023-10-10 00:51:57 -04:00
Jan Buethe
0563d71b25
updated osce readme
2023-10-07 18:52:38 +02:00
Jean-Marc Valin
f0ec990dba
Switching to neural pitch estimator
...
Remove old pitch estimator and retrain all models
2023-10-06 03:14:56 -04:00
Jean-Marc Valin
27663d3641
Using a DenseNet for DRED
2023-10-02 01:43:44 -04:00
Jean-Marc Valin
8e8edf71bd
Remove unneeded (I think) tanh at the end
2023-10-01 21:34:58 -04:00
Jean-Marc Valin
33adba02c7
First version of pitch DNN C code
...
Totally untested -- most likely doesn't work
2023-10-01 03:59:17 -04:00
Jean-Marc Valin
f3b86f9414
Fix model saving
2023-09-30 02:48:26 -04:00
Jan Buethe
0459a572f5
updated PitchDNN export script
2023-09-29 15:34:59 +02:00
Jan Buethe
ce28695844
refactoring and cleanup
2023-09-29 15:31:45 +02:00
Jan Buethe
4901445490
fixed type in error message
2023-09-29 14:34:11 +02:00
Jan Buethe
c5c214df1b
added rudimentary support for dumping nn.Conv2d layers
2023-09-29 14:25:26 +02:00
Jean-Marc Valin
9a7bb764d4
No features skip needed to align pitch features
2023-09-27 13:01:21 -04:00
Jean-Marc Valin
a6b4fe375a
Script to compute the groundtruth data using CREPE
2023-09-27 13:00:12 -04:00
Jan Buethe
ae0a140c40
updated default values in adv_train_fargan.py
2023-09-26 21:42:01 +02:00
Jan Buethe
41a4c9515d
changed checkpoint format
2023-09-26 12:12:47 -04:00
Jean-Marc Valin
733a095ba2
Adapting to new data format/model
2023-09-26 12:12:47 -04:00
Krishna Subramani
f38b4a317f
Python code for neural pitch
2023-09-26 12:12:47 -04:00
Jean-Marc Valin
c4b83ae62d
RDOVAE model update
2023-09-26 12:11:25 -04:00
Jan Buethe
00580a63aa
bugfix
2023-09-22 11:39:22 +02:00
Jan Buethe
aad74fdfd6
pulse-related bugfix
2023-09-22 11:18:41 +02:00
Jan Buethe
0a92bc5eaa
more lavoce stuff
2023-09-21 15:01:11 +02:00
Jean-Marc Valin
b88644b9c7
Quantizing initial state with rdovae too
...
More efficient than PVQ
2023-09-20 18:04:08 -04:00
Jan Buethe
2ec31cc5cc
added FARGAN adversarial training script
2023-09-20 19:06:10 +02:00
Jan Buethe
82f48d368b
removed trailing whitespace in fargan
...
Signed-off-by: Jan Buethe <jbuethe@amazon.de>
2023-09-13 16:57:28 +02:00
Jan Buethe
e7beaec3fb
integrated JM's FFT ada conv
...
Signed-off-by: Jan Buethe <jbuethe@amazon.de>
2023-09-13 16:31:29 +02:00
Jean-Marc Valin
b24c7b433a
Remove --has-gain and --passthrough-size args
...
Couldn't be used anymore
2023-09-12 22:58:34 -04:00
Jean-Marc Valin
2f8b36d691
Add conditioning interpolation, fwconv layer
2023-09-12 22:50:48 -04:00
Jean-Marc Valin
72c5ea4129
Only use one frame of pre-loading
2023-09-12 22:50:48 -04:00
Jean-Marc Valin
108b75c4b1
Randomly double the training sequence length
...
Helps with stability with little cost in training speed
2023-09-12 22:50:48 -04:00
Jean-Marc Valin
d54b9fb49a
Adds skip connections
2023-09-12 22:50:47 -04:00
Jean-Marc Valin
fb570ed8bb
5-tap pitch predictor
2023-09-12 22:50:47 -04:00
