Commit Graph

165 Commits (788c8100ba65a54120defdbe168676b7fed705a8)

Author SHA1 Message Date
Eren Golge bb2b705e01 small bug fixes 2019-05-14 13:53:26 +02:00
Eren Golge 2b60f9a731 Fix trans agent implementation in relation to the paper. Use query vector insteadd of processed_query 2019-05-12 17:39:43 +02:00
Eren Golge 6331bccefc make dropout oprional #2 2019-05-12 17:35:31 +02:00
Eren Golge 820d18c922 make dropout at prenet optional 2019-05-12 17:34:57 +02:00
Eren Golge afb5a17221 bug fix 2019-04-30 10:59:29 +02:00
Eren Golge e2439fde9a make location attention optional and keep all attention weights in attention class 2019-04-29 11:37:01 +02:00
Eren Golge 38213dffe9 bug fix #2 2019-04-18 18:55:37 +02:00
Eren Golge 9ba13b2d2f fix forward attention 2019-04-18 18:36:01 +02:00
Eren Golge f450fe3571 use stop token again 2019-04-18 15:20:19 +02:00
Eren Golge 3c2d500f53 Changesat windowing and some comments 2019-04-12 16:13:40 +02:00
Eren Golge 312a539a0e Enable optional forward attention with transition agent 2019-04-10 16:41:30 +02:00
Eren Golge e3647fa7b3 bug fix for prenet setup 2019-04-08 18:28:19 +02:00
Eren Golge 961af0f5cd setup_model externally based on model selection. Make forward attention and prenet type configurable in config.json 2019-04-05 17:49:18 +02:00
Eren Golge 043e49f367 active windowing 2019-04-05 08:40:09 +02:00
Eren Golge 6e8b66aa89 smaller windowing range 2019-04-02 10:35:13 +02:00
Eren Golge 68f8ef730d stop conditioning with padding for inference_truncated 2019-04-01 14:10:38 +02:00
Eren Golge 5212a11836 longer stop_token padding 2019-04-01 12:07:24 +02:00
Eren Golge e1cd253d65 change stop conditioning 2019-03-31 16:44:17 +02:00
Eren Golge fdca8402c7 config updates 2019-03-26 15:46:26 +01:00
Eren Golge 0a92c6d5a7 Set attention norm method by config.json 2019-03-26 00:48:12 +01:00
Eren Golge 786510cd6a loss functions updates 2019-03-23 17:33:47 +01:00
Eren Golge 82cde95cfa add bias to attention v 2019-03-19 12:21:36 +01:00
Eren Golge 1b68d3cb4e control synthesis lenght as an additional stop condition 2019-03-15 14:01:43 +01:00
Eren Golge 4f89029577 Merge branch 'state-pass' into dev-tacotron2 2019-03-12 09:52:15 +01:00
Eren Golge 3128378bdf bug fix for stop token prediciton 2019-03-12 00:20:57 +01:00
Eren Golge 527567d7ce renaming 2019-03-12 00:20:43 +01:00
Eren Golge b9b79fcf0f inference truncated NEED TO BE TESTED 2019-03-11 17:40:09 +01:00
Eren Golge 4ffda89c42 reshape input vectors before and after bn layer 2019-03-11 13:03:43 +01:00
Eren Golge a144acf466 use BN for prenet 2019-03-07 16:28:50 +01:00
Eren Golge 4b116a2a88 Look for the last two attention values for stop condition and attend to the first encoder verctor if it is the first decoder iteration 2019-03-06 23:46:02 +01:00
Eren Golge b031a65677 compute sequence mask in model, add tacotron2 relatedfiles 2019-03-06 13:14:58 +01:00
Eren Golge a4474abd83 tacotron parse output bug fix 2019-03-06 13:10:54 +01:00
Eren Golge 4326582bb1 TTSDataset formatting and batch sorting to use pytorch pack for rnns 2019-03-06 13:10:05 +01:00
Eren Golge 44c66c6e3e remove comments 2019-03-05 13:34:33 +01:00
Eren Golge 1e8fdec084 Modularize functions in Tacotron 2019-03-05 13:25:50 +01:00
Eren Golge 1c99be2ffd Change window size for attention 2019-02-18 13:06:26 +01:00
Eren Golge 6ea31e47df Constant queue size for autoregression window 2019-02-16 03:18:49 +01:00
Eren Golge 90f0cd640b memoru queueing 2019-02-12 15:27:42 +01:00
Eren Golge c5b6227848 init with embedding lyaers 2019-02-06 16:54:33 +01:00
Eren Golge d28bbe09fb Attention bias setting Revert to old 2019-02-06 16:23:01 +01:00
Eren Golge e12bbc2a5c init decoder states with a function 2019-01-22 18:25:55 +01:00
Eren Golge 66f8d0e260 Attention biased chaged 2019-01-22 18:18:21 +01:00
Eren Golge 562d73d3d1 Soem bug fixes 2019-01-17 15:48:22 +01:00
Eren Golge 4431e04b48 use sigmoid for attention 2019-01-16 16:26:05 +01:00
Eren Golge 7e020d4084 Bug fixes 2019-01-16 16:23:04 +01:00
Eren Golge af22bed149 set bias 2019-01-16 15:53:24 +01:00
Eren Golge f4fa155cd3 Make attn windowing optional 2019-01-16 12:33:07 +01:00
Eren Golge 8969d59902 Use the last attention value as a threshold to stop decoding. since stoptoken prediction is not precies enough to synthsis at the right time. 2019-01-16 12:32:40 +01:00
Eren Golge ed1f648b83 Enalbe attention windowing and make in configurable at model level. 2019-01-16 12:32:40 +01:00
Eren Golge 916f5df5f9 config update and increase dropout p 0.5 2019-01-16 12:14:58 +01:00