FreeDV
/
ampnn
mirror of https://github.com/drowe67/ampnn.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Commit Graph

  • f529237fd2 first pass output from eband_vq_am, 3/4 test samples reasonable master David 2021-01-03 20:20:32 +1030
  • 9e5bcfa042 corrected mean removal on test at end, setting up for model saving David 2021-01-03 17:54:07 +1030
  • 19464deed2 eband->Am experiments using timpulse David 2021-01-03 09:07:47 +1030
  • 42959c762a script to drive test with synthetic impulse David 2021-01-02 18:01:42 +1030
  • 90aebb6438 Merge branch 'master' of github.com:drowe67/ampnn David 2021-01-02 18:00:44 +1030
  • bf3bdbcc92 through synthetic signal unit test, determined that eband->Am was a function of Wo, so introduced an Wo parameter to the model. Improved perf significantly David 2021-01-02 18:00:26 +1030
  • 4c1e9a0003 working on per chunk mean removal David 2021-01-01 15:34:05 +1030
  • 374a645837 removing low energy vectors David 2021-01-01 13:56:55 +1030
  • fa1b6a7b10 conversion to sparse vectors as pre-prcoessing step to save time David 2021-01-01 12:37:05 +1030
  • 72087671bb first pass at combined VQVAE and rate K->L, running but reults poor David 2021-01-01 11:51:52 +1030
  • f1a8fb2b71 better pager David 2020-12-31 13:49:56 +1030
  • 51a41e4489 first pass at vqvae->eband, muffled forr males David 2020-12-31 12:37:46 +1030
  • 24eca42413 preserving original mean energy PDF info David 2020-12-31 08:14:25 +1030
  • 6e69615306 trying striedes=2 rather than maxpooling David 2020-12-29 17:42:39 +1030
  • 39e4e409ff helper script for processing eband stuff David 2020-12-29 16:09:49 +1030
  • 6af5029db9 vectors in dB, adam optimizer, working quite well David 2020-12-29 15:49:49 +1030
  • d45dcdfa5f slightly faster as just reads records once David 2020-12-29 15:45:57 +1030
  • f5b3b774fb weighting vector option - but didn't help much David 2020-12-29 13:16:11 +1030
  • 90d6a49e9d phase0 output file David 2020-12-29 13:15:44 +1030
  • 290bfed45b
         Update README.md drowe67 2020-12-29 06:36:40 +1030
  • ff5b6502ac weighted loss function - didn't improve speech of males David 2020-12-27 20:13:02 +1030
  • 4c954d8dea refactoring into vqvae_util.py David 2020-12-27 15:52:38 +1030
  • 1c7fb55cc0 updated README David 2020-12-23 16:54:07 +1030
  • 2859384ed0 setting lower energy limit and handy synth script David 2020-12-23 15:46:51 +1030
  • 6d9f8810a2 first speech out! Pretty poor speech, but it vorks :-) David 2020-12-23 12:00:21 +1030
  • 96d86b3c1c first pass at vq_vae_kmeans_conv1d_out decoder script David 2020-12-23 11:20:50 +1030
  • 27a377bc97 saving and loading NN weights and VQ David 2020-12-23 10:44:13 +1030
  • ba34d1d311 make sure we only train VQ in training phase David 2020-12-23 09:45:22 +1030
  • 063e2f0291 moving vqvae_twostage to a separate file that can be shared with prediction script David 2020-12-23 09:44:57 +1030
  • b3f4038c0c two stage VQ, various mean options David 2020-12-23 07:46:38 +1030
  • 42125f4c4b removing means from chunks, optional threshold David 2020-12-22 20:23:52 +1030
  • 8c9ef7be27 using conv1d valid, less sensitive to nb_timesteps David 2020-12-22 17:56:07 +1030
  • fae9050eea options for mean David 2020-12-22 16:07:31 +1030
  • 8c648930b7 VQ working for larger codebooks, protection against unused vectors and NaNs David 2020-12-22 13:10:26 +1030
  • d07280e39f starting to get sensible results David 2020-12-22 09:09:11 +1030
  • 09d80356d2 first pass vq_vae_kmeans_conv1d.py training David 2020-12-22 07:48:14 +1030
  • c0179d2ed3 ported eband_train and eband_out to TF2 David 2020-12-21 10:46:30 +1030
  • a9dfd2f48a updated README David 2020-11-30 07:35:56 +1030
  • 5b458f156c added dense layer at input, plot encoder space David 2020-11-30 07:25:09 +1030
  • 8a91b95961 added comittment loss David 2020-11-28 06:08:52 +1030
  • 66db32ca97 shorter names David 2020-11-27 18:43:45 +1030
  • b821dde612 modified to remove mean energy, might change back David 2020-11-27 18:43:32 +1030
  • 160c301a8d shorter names David 2020-11-27 18:42:18 +1030
  • df04d50454 using custom layer to get gradient over top of VQ David 2020-11-27 18:42:03 +1030
  • bea6cd8b92 doc tweaks David 2020-11-24 08:36:44 +1030
  • 51d3a4650a moved vq_ewma_kmeans to a file we can share David 2020-11-24 08:31:19 +1030
  • ef41fe9f5c initialising VQ using a numpy matrix David 2020-11-24 07:37:59 +1030
  • b626e38d74 some refactoring David 2020-11-24 06:55:19 +1030
  • ec900cd069 fixed instability Centroid needed it's own variable David 2020-11-24 06:32:25 +1030
  • a1fe6bd5a4 found correct function for assignment I think, still a WIP David 2020-11-23 20:48:51 +1030
  • b668686a16 building up kmeans David 2020-11-23 20:06:31 +1030
  • 0ef94b3acf building up kmeans David 2020-11-23 20:05:58 +1030
  • 77c69cc982 typo David 2020-11-23 17:48:09 +1030
  • 7cd0db4790 setting VQ entry externally David 2020-11-23 17:47:44 +1030
  • 24407ae8df better variable name, tested decoder David 2020-11-19 20:46:35 +1030
  • 77238c5437 rotated weights to be more sensible for my VQ work David 2020-11-18 07:19:14 +1030
  • 9fc20cc302 about to rotate weights David 2020-11-18 06:47:47 +1030
  • bfa1a04e81 building up EWMA David 2020-11-18 06:34:43 +1030
  • 6d05f77956 ported vq_vae_demo to tf.keras but had to disable eager execution David 2020-11-09 06:19:05 +1030
  • f9b8387133 visualising first conv1D which didn't prove very useful! David 2020-11-01 16:58:02 +1030
  • bee466e8eb README edits David 2020-11-01 15:09:58 +1030
  • 6faead1b96 clean up and pager improvements David 2020-10-31 10:28:08 +1030
  • 4451a0b103 cleaned up repo and README David 2020-10-31 07:45:39 +1030
  • a704caf56a building up pager David 2020-10-31 07:45:20 +1030
  • 7f26b71d6b running on TF 2 David 2020-10-31 07:37:29 +1030
  • 5357121037 plotting 2D PCA of encoder and VQ space, looks cool David 2020-10-30 15:58:25 +1030
  • 7b99076f9e modified for TF2, possible bug in e_latent_loss formation for 2 stages David 2020-10-29 19:06:47 +1030
  • ec9a01a361 update description of repo David 2020-03-22 12:08:44 +1030
  • e4f5e783a6 counting how many times each vector is used, a test of VQ training David 2020-03-22 12:04:34 +1030
  • a3ac134982 counting how many times each vector is used, a test of VQ trarining David 2020-03-22 12:04:05 +1030
  • 55ae42b814 typo David 2020-03-22 11:58:00 +1030
  • 06ddf8d408 checking VQ index use with non-backend VQ search David 2020-03-22 11:57:24 +1030
  • e167ca2719 cleanup David 2020-03-19 11:06:09 +1030
  • f600fddb98 cleaned up unused code David 2020-03-18 14:06:00 +1030
  • 6af7b4ef76 first pass running, printing tensors David 2020-03-18 13:50:17 +1030
  • b78b7c1b5e 6.5dB*dB 11*4/0.08=550 bits/s David 2020-03-14 13:51:45 +1030
  • bc230e144a wip, 12dB*dB with 20 bits/80ms David 2020-03-14 10:48:47 +1030
  • 6f76e25daf two stage VQ conv1D working, but not getting sensible results yet David 2020-03-14 09:54:51 +1030
  • 74d18e6ffd two stage VQ demo working David 2020-03-14 09:13:02 +1030
  • 8d4582a777 simplified David 2020-03-14 08:13:33 +1030
  • d255df9f97 renamed some variables David 2020-03-14 08:09:11 +1030
  • 134119a049 first stage only David 2020-03-14 08:04:29 +1030
  • b239f26969 gaussian test case David 2020-03-14 06:31:20 +1030
  • 430f46293c conv1D VQ-VAE stuck at about 15 dB*dB David 2020-03-10 15:48:05 +1030
  • bc18a6e296 VQVAE with conv1d running but silly results David 2020-03-09 12:54:48 +1030
  • 51fde46af0 plotting VQ entry scatter on every epoch, to see VQ evolve David 2020-03-09 11:17:53 +1030
  • e52b2d7f41 added features to investgate worst frames David 2020-03-09 09:26:29 +1030
  • 682fef0f34 added some options for analysing worstcase David 2020-03-08 16:43:33 +1030
  • 61eaf2b418 typo in name David 2020-03-08 07:56:04 +1030
  • bed6ca506c 8 frames down to 2 x 16 element vectors, quite reasonable David 2020-02-28 20:46:47 +1030
  • f3faf129d1 first pass using conv1D doing something sensible, 2.5dB mse David 2020-02-28 18:53:32 +1030
  • 5a8b47c2e7 4D test giving MSE results around 6.5dB*dB with and without Dense output layer. Suggests Dense output layer isn't breaking anything. Need to run for just 1 epoch, train_scale=0.125 David 2020-02-22 15:09:51 +1030
  • 78f5934d00 scaling features to +/-1 David 2020-02-21 17:05:03 +1030
  • 4e991e5500 doing sensible things with 2 and 4d vectors David 2020-02-15 12:21:39 +1030
  • 867c4e932f some more plots, attempting to get working on lower dimensional data David 2020-02-15 10:29:47 +1030
  • 8ce1c9c1d3 first pass VQ-VAE running, but poor SDcresults David 2020-02-14 12:58:28 +1030
  • 3d9a4cf3ce 4 point demo cleaned up David 2020-02-13 18:19:24 +1030
  • 3e22fd6284 4 point demo cleaned up David 2020-02-13 18:18:48 +1030
  • 0c4d2dbe51 4 point demo working David 2020-02-13 17:46:28 +1030
  • 62ce3a4474 simple VQ example David 2020-02-13 17:35:31 +1030