56 Commits

Author	SHA1	Message	Date
Daniel Povey	2eef001d39	Fix balancer code	2022-03-21 23:59:26 +08:00
Daniel Povey	87c92efbfe	Changes from upstream/master	2022-03-16 21:49:15 +08:00
Daniel Povey	e838c192ef	Cosmetic changes/renaming things	2022-03-16 19:27:45 +08:00
Daniel Povey	dfc75752c4	Remove some dead code.	2022-03-16 18:06:01 +08:00
Daniel Povey	c82db4184a	Remove xscale from pos_embedding	2022-03-16 15:50:11 +08:00
Daniel Povey	0e9cad3f1f	Modifying initialization from normal->uniform; add initial_scale when initializing	2022-03-16 14:42:53 +08:00
Daniel Povey	00be56c7a0	Remove dead code	2022-03-16 12:49:00 +08:00
Daniel Povey	633213424d	Rework of initialization	2022-03-16 12:42:59 +08:00
Daniel Povey	261d7602a7	Draft of 0mean changes..	2022-03-15 23:46:53 +08:00
Daniel Povey	fc873cc50d	Make epsilon in BasicNorm learnable, optionally.	2022-03-15 17:00:17 +08:00
Daniel Povey	86e5dcba11	Remove max-positive constraint in deriv-balancing; add second DerivBalancer in conv module.	2022-03-15 13:10:35 +08:00
Daniel Povey	a23010fc10	Add warmup mode	2022-03-14 23:04:51 +08:00
Daniel Povey	8d17a05dd2	Reduce constraints from deriv-balancer in ConvModule.	2022-03-14 19:23:33 +08:00
Daniel Povey	437e8b2083	Reduce max-abs limit from 1000 to 100; introduce 2 DerivBalancer modules in conv layer.	2022-03-13 23:31:08 +08:00
Daniel Povey	f351777e9c	Remove ExpScale in feedforward layes.	2022-03-13 17:29:39 +08:00
Daniel Povey	5d69acb25b	Add max-abs-value	2022-03-13 13:15:20 +08:00
Daniel Povey	db7a3b6eea	Reduce initial_scale.	2022-03-12 18:50:02 +08:00
Daniel Povey	b7b2d8970b	Cosmetic change	2022-03-12 17:47:35 +08:00
Daniel Povey	a392cb9fbc	Reduce initial scaling of modules	2022-03-12 16:53:03 +08:00
Daniel Povey	ca8cf2a73b	Another rework, use scales on linear/conv	2022-03-12 15:38:13 +08:00
Daniel Povey	0abba9e7a2	Fix self.post-scale-mha	2022-03-12 11:20:44 +08:00
Daniel Povey	76a2b9d362	Add learnable post-scale for mha	2022-03-12 11:19:49 +08:00
Daniel Povey	7eb5a84cbe	Add identity pre_norm_final for diagnostics.	2022-03-11 21:00:43 +08:00
Daniel Povey	cc558faf26	Fix scale from 0.5 to 2.0 as I really intended..	2022-03-11 19:11:50 +08:00
Daniel Povey	98156711ef	Introduce in_scale=0.5 for SwishExpScale	2022-03-11 19:07:34 +08:00
Daniel Povey	5eafccb369	Change how scales are applied; fix residual bug	2022-03-11 17:46:33 +08:00
Daniel Povey	bcf417fce2	Change max_factor in DerivBalancer from 0.025 to 0.01; fix scaling code.	2022-03-11 14:47:46 +08:00
Daniel Povey	2940d3106f	Fix q*scaling logic	2022-03-11 14:44:13 +08:00
Daniel Povey	ab9a17413a	Scale up pos_bias_u and pos_bias_v before use.	2022-03-11 14:37:52 +08:00
Daniel Povey	2fa9c636a4	use nonzero threshold in DerivBalancer	2022-03-10 23:24:55 +08:00
Daniel Povey	425e274c82	Replace norm in ConvolutionModule with a scaling factor.	2022-03-10 16:01:53 +08:00
Daniel Povey	b55472bb42	Replace most normalizations with scales (still have norm in conv)	2022-03-10 14:43:54 +08:00
Daniel Povey	a37d98463a	Restore ConvolutionModule to state before changes; change all Swish,Swish(Swish) to SwishOffset.	2022-03-06 11:55:02 +08:00
Daniel Povey	8a8b81cd18	Replace relu with swish-squared.	2022-03-05 22:21:42 +08:00
Fangjun Kuang	1603744469	Refactor conformer. (#237)	2022-03-05 19:26:06 +08:00
Daniel Povey	5f2c0a09b7	Convert swish nonlinearities to ReLU	2022-03-05 16:28:24 +08:00
Daniel Povey	65b09dd5f2	Double the threshold in brelu; slightly increase max_factor.	2022-03-05 00:07:14 +08:00
Daniel Povey	6252282fd0	Add deriv-balancing code	2022-03-04 20:19:11 +08:00
Daniel Povey	eb3ed54202	Reduce scale from 50 to 20	2022-03-04 15:56:45 +08:00
Daniel Povey	9cc5999829	Fix duplicate Swish; replace norm+swish with swish+exp-scale in convolution module	2022-03-04 15:50:51 +08:00
Daniel Povey	7e88999641	Increase scale from 20 to 50.	2022-03-04 14:31:29 +08:00
Daniel Povey	3207bd98a9	Increase scale on Scale from 4 to 20	2022-03-04 13:16:40 +08:00
Daniel Povey	cd216f50b6	Add import	2022-03-04 11:03:01 +08:00
Daniel Povey	bc6c720e25	Combine ExpScale and swish for memory reduction	2022-03-04 10:52:05 +08:00
Daniel Povey	23b3aa233c	Double learning rate of exp-scale units	2022-03-04 00:42:37 +08:00
Daniel Povey	5c177fc52b	pelu_base->expscale, add 2xExpScale in subsampling, and in feedforward units.	2022-03-03 23:52:03 +08:00
Daniel Povey	3fb559d2f0	Add baseline for the PeLU expt, keeping only the small normalization-related changes.	2022-03-02 18:27:08 +08:00
Daniel Povey	9d1b4ae046	Add pelu to this good-performing setup..	2022-03-02 16:33:27 +08:00
Daniel Povey	c1063def95	First version of rand-combine iterated-training-like idea.	2022-02-27 17:34:58 +08:00
Daniel Povey	63d8d935d4	Refactor/simplify ConformerEncoder	2022-02-27 13:56:15 +08:00