25 Commits

Author	SHA1	Message	Date
Eren Gölge	7a382a5c2b	stowed aligntts commit and small refactoring with feed_forward layers	2021-03-30 14:39:16 +02:00
Eren Gölge	65533f33e9	fix #374	2021-03-18 13:33:00 +01:00
Eren Gölge	d0454461de	Merge branch 'pr/gerazov/650-2' into dev	2021-02-17 13:40:45 +00:00
gerazov	61c88beb94	refactored keep_all_best	2021-02-15 18:40:17 +01:00
Eren Gölge	e774f68aee	save used model characters to the checkpoints	2021-02-12 12:03:42 +00:00
gerazov	310d18325e	brushed up printing model load path and best loss path	2021-02-12 10:55:45 +01:00
gerazov	af46727517	loading last checkpoint/best_model works, deleting last best models options added, loading last best_loss added	2021-02-12 02:12:00 +01:00
gerazov	ad17dc9e76	final fixes	2021-02-06 23:05:01 +01:00
gerazov	8fdd08ea15	updated to current dev	2021-02-06 22:59:52 +01:00
gerazov	2705d27b28	changed train scripts	2021-02-06 22:29:30 +01:00
Eren Gölge	d49757faaa	linter updates	2021-02-05 13:10:43 +00:00
Alexander Korolev	ca28e05ed7	update fixed stopnet_pos_weight parameter ... config parameter c.stopnet_pos_weight has currently no effect as it is not used.	2021-01-27 16:33:25 +01:00
Alexander Korolev	f251dc8c0e	Update train_tacotron.py ... When attempting to fine-tune a model with "prenet_type": "bn" that was originally trained with "prenet_type": "original", a RuntimeError is thrown that stops the training. By catching the RuntimeError, the required layers can be partially restored and the training will continue without any problems.	2021-01-21 21:16:30 +01:00
erogol	d5a0190c4b	update copy_config_file to copy_model_files	2021-01-06 13:19:40 +01:00
erogol	29b17c0808	bug fix for gradual training	2021-01-06 13:19:40 +01:00
erogol	6478d552dc	tacotron training bug fix	2021-01-06 13:19:40 +01:00
erogol	1dd086577a	tacotron training bug fix	2021-01-06 13:18:41 +01:00
erogol	affe1c1138	setup training scripts for computing phonemes before training optionally. And define data_loaders before starting training and re-use them instead of re-define for every train and eval calls. This is to enable better instance filtering based on input length.	2020-12-07 11:26:57 +01:00
erogol	7c3cdced1a	make speaker_mapping a global variable to prevent reload. Fix glow-tts training	2020-12-01 03:23:25 +01:00
erogol	d8c1b5b73d	print max lengths in tacotron training	2020-11-25 14:49:07 +01:00
erogol	1229554c42	use native amp	2020-11-25 14:48:54 +01:00
erogol	aa2b31a1b0	use 'enabled' argument to control autocast	2020-11-17 14:22:01 +01:00
erogol	d8511efa8f	use native amp for tacotron training	2020-11-14 12:59:28 +01:00
erogol	a7aefd5c50	use pytorch amp for mixed precision training for Tacotron	2020-11-12 12:51:56 +01:00
erogol	73581cd94c	renaming train scripts and updating tests	2020-10-29 16:50:07 +01:00