54 Commits

Author	SHA1	Message	Date
Daniel Povey	a6bf32c3e0	Fast code seems to pass tests.	2022-06-18 12:56:16 +08:00
Daniel Povey	ea5cd69e3b	Possibly fix bug RE learning rate	2022-06-17 20:50:00 +08:00
Daniel Povey	cbccc1dd91	Fix bias_correction2, power of 0.5	2022-06-17 20:20:16 +08:00
Daniel Povey	2d13b682fe	Revert param_rel_eps to old value	2022-06-17 18:22:27 +08:00
Daniel Povey	92fb1c52c8	Reduce debug statement frequency	2022-06-17 16:03:50 +08:00
Daniel Povey	d9a3eba370	Fix bug that caused it to use too much memory	2022-06-17 16:02:47 +08:00
Daniel Povey	6873aa1a13	Bug fix to print statement	2022-06-17 15:35:07 +08:00
Daniel Povey	90622f88ad	Decrease param_rel_eps default from 1e-4 to 1e-2	2022-06-17 15:10:46 +08:00
Daniel Povey	7f0756e156	Implement structured version of conformer	2022-06-17 15:10:21 +08:00
Daniel Povey	2fe4af8c99	Prevent crash due to error in C_diff	2022-06-17 12:35:55 +08:00
Daniel Povey	6c499dcd66	Augment debug output	2022-06-17 12:32:09 +08:00
Daniel Povey	5c3e4f506f	Reduce debug frequency	2022-06-17 12:17:48 +08:00
Daniel Povey	47df144253	Bug fix for scalar case	2022-06-17 12:16:12 +08:00
Daniel Povey	9e92d13a33	Some bug fixes.. seems to be working.	2022-06-17 12:04:21 +08:00
Daniel Povey	827a37c7fc	Closer to finished	2022-06-17 11:10:43 +08:00
Daniel Povey	94e48ff07e	Some code improvements...	2022-06-16 10:43:26 +08:00
Daniel Povey	f9bb745cf3	Use exp_avg_sq as ref_exp_avg_sq	2022-06-15 14:03:26 +08:00
Daniel Povey	860322bf30	Make it less verbose; fix scale_speed setting; testing min_diag_smooth=1.0 for debug	2022-06-15 13:02:04 +08:00
Daniel Povey	bc5d72b0f3	Fix potential bug re scale change	2022-06-15 12:49:47 +08:00
Daniel Povey	57957cc049	Add diagnostics	2022-06-15 12:39:16 +08:00
Daniel Povey	0679b363b0	Some debug stuff, modify cond_eps for param_cov	2022-06-14 22:38:51 +08:00
Daniel Povey	932cedce59	Bug fix RE eps*eps; add/tune diagnostics	2022-06-14 21:43:27 +08:00
Daniel Povey	2615e48779	Some cleanup and configuration changes.	2022-06-14 21:17:50 +08:00
Daniel Povey	c6f62dfc08	Version that is passing the tests	2022-06-14 20:45:57 +08:00
Daniel Povey	e597b68160	Improve roundoff properties of algorithm for getting C, use randomized testing.	2022-06-14 12:59:41 +08:00
Daniel Povey	d729e970e1	Some changes for stability	2022-06-14 10:25:17 +08:00
Daniel Povey	d1e96afce2	Improve speed by messing with configuration, removing asserts.	2022-06-13 17:27:34 +08:00
Daniel Povey	a41c4b6c9b	Bug fixes	2022-06-13 17:20:46 +08:00
Daniel Povey	41b0e582df	Bug fix to typo in optim.py	2022-06-13 16:10:49 +08:00
Daniel Povey	ca7cffcb42	Remove Decorrelate() class	2022-06-13 16:08:32 +08:00
Daniel Povey	7338c60296	Remove Decorrelate()	2022-06-13 16:07:15 +08:00
Daniel Povey	c1f487e36d	Move optim2.py to optim.py; use this optimizer in train.py	2022-06-13 16:05:46 +08:00
Daniel Povey	4e0d8c45f4	Move optim2.py to optim.py; use this optimizer in train.py	2022-06-13 16:05:22 +08:00
Daniel Povey	d5b90f278e	Adding draft of new optimizer	2022-06-13 15:01:10 +08:00
Daniel Povey	8c4b5454e7	Remove unnecessary line	2022-06-11 13:14:18 +08:00
Daniel Povey	73fd2fcd77	Updates to cain to new param_rms per element	2022-06-11 13:08:08 +08:00
Daniel Povey	d301f8ac6c	Merge Decorrelate work, and simplification to RandomCombine, into pruned_transducer_stateless7	2022-06-11 11:07:07 +08:00
Daniel Povey	a9a172aa69	Multiply lr by 10; simplify Cain.	2022-06-04 15:48:33 +08:00
Daniel Povey	679972b905	Fix bug; make epsilon work both ways (small+large); increase epsilon to 0.1	2022-06-03 19:37:48 +08:00
Daniel Povey	8085ed6ef9	Turn off natural gradient update for biases.	2022-06-03 18:40:14 +08:00
Daniel Povey	3fff0c75bb	Code cleanup	2022-06-03 11:54:12 +08:00
Daniel Povey	d6e65a0e7f	Remove decompose=True	2022-06-03 11:48:45 +08:00
Daniel Povey	a66a0d84d5	Natural gradient, with power -0.5 (halfway; -1 would be NG)	2022-06-02 14:01:03 +08:00
Daniel Povey	b1f6797af1	Remove some rebalancing code that I am now not going to use.	2022-06-01 22:19:28 +08:00
Daniel Povey	0c73664aef	Reduce threshold to 1024	2022-06-01 14:42:56 +08:00
Daniel Povey	ca09b9798f	Remove decomposition code from checkpoint.py; restore double precision model_avg	2022-06-01 14:01:58 +08:00
Daniel Povey	03e07e80ce	More drafts for rebalancing code	2022-06-01 13:58:42 +08:00
Daniel Povey	9c9bf4f1e3	Some drafts of rebalancing code in optim.py	2022-06-01 11:34:19 +08:00
Daniel Povey	bc5c782294	Limit magnitude of linear_pos	2022-06-01 10:40:54 +08:00
Daniel Povey	61619c031e	Add activation balancer to stop activations in self_attn from getting too large	2022-06-01 00:40:45 +08:00