Daniel Povey
6caaa4e9c6
Bug fix in caching_eval, may make no difference.
2022-12-15 23:32:29 +08:00
Daniel Povey
f5d4fb092d
Bug fix in caching_eval
2022-12-15 23:24:36 +08:00
Daniel Povey
d26ee2bf81
Try to implement caching evaluation for memory efficient training
2022-12-15 23:06:40 +08:00
Daniel Povey
f66c1600f4
Bug fix to printing code
2022-12-15 21:55:23 +08:00
Daniel Povey
2d0fe7637c
Memory fix in WithLoss
2022-12-11 17:20:26 +08:00
Daniel Povey
0fc646f281
Merge branch 'scaled_adam_exp663' into scaled_adam_exp665
2022-12-10 00:07:37 +08:00
Daniel Povey
d35eb7a3a6
Add cosmetic/diagnostics changes from scaled_adam_exp656.
2022-12-09 22:02:42 +08:00
Daniel Povey
5c0957d950
Fix memory issue in ActivationBalancer
2022-12-09 18:11:27 +08:00
Daniel Povey
2ef0228db0
Make the ActivationBalancer relative to the mean, limited to -min_abs..max_abs
2022-12-09 17:59:00 +08:00
Daniel Povey
3f82ee0783
Merge dropout schedule, 0.3 ... 0.1 over 20k batches
2022-12-08 18:18:46 +08:00
Daniel Povey
22617da725
Make dropout a schedule starting at 0.3.
2022-12-05 23:39:24 +08:00
Daniel Povey
178eca1c0e
Revert scaling, scale only grad.
2022-12-05 17:53:23 +08:00
Daniel Povey
b93cf0676a
Initialize Conv2dSubsampling with scale.
2022-12-05 17:31:56 +08:00
Daniel Povey
12fb2081b1
Fix deriv code
2022-12-04 21:22:06 +08:00
Daniel Povey
c57eaf7979
Change x coeff from -0.1 to -0.08, as in 608.
2022-12-04 21:15:49 +08:00
Daniel Povey
7b1f093077
Use Swoosh-R in the Conv and Swoosh-L in the feedforward.
2022-12-04 19:18:16 +08:00
Daniel Povey
67812276ed
Change Swoosh formula so left crossing is near zero; change min_positive, max_positive of ActivationBalancer.
2022-12-03 15:10:03 +08:00
Daniel Povey
b8e3091e04
Increase scale_gain_factor to 0.04.
2022-12-03 00:48:19 +08:00
Daniel Povey
bd1b1dd7e3
Simplify formula for Swoosh and make it pass through 0; make max_abs of ConvolutionModule a constant.
2022-12-03 00:13:09 +08:00
Daniel Povey
84f51ab1b1
Bug fix in scripting mode
2022-12-02 20:28:17 +08:00
Daniel Povey
9a2a58e20d
Fix bug one versus zero
2022-12-02 19:12:18 +08:00
Daniel Povey
2bfc38207c
Fix constants in SwooshFunction.
2022-12-02 18:37:23 +08:00
Daniel Povey
14267a5194
Use Swoosh not DoubleSwish in zipformer; fix constants in Swoosh
2022-12-02 16:58:31 +08:00
Daniel Povey
ec10573edc
First version of swoosh
2022-12-02 16:34:53 +08:00
Daniel Povey
d260b54177
Subtract, not add, 0.025.
2022-12-02 15:55:48 +08:00
Daniel Povey
9a71406a46
Reduce offset from 0.075 to 0.025.
2022-12-02 15:40:21 +08:00
Daniel Povey
c71a3c6098
Change offset
2022-12-02 15:20:37 +08:00
Daniel Povey
f0f204552d
Add -0.05 to DoubleSwish.
2022-12-02 15:17:41 +08:00
Daniel Povey
983a690c63
Change DoubleSwish formulation, add alpha*x only for x.abs() > 0.15.
2022-12-01 17:20:56 +08:00
Daniel Povey
d682ecc246
Introduce alpha for DoubleSwish, set it to -0.05.
2022-11-30 18:58:25 +08:00
Daniel Povey
0bfd81d721
fix bug RE dims_to_mean
2022-11-28 10:42:06 +08:00
Daniel Povey
109825cafb
Fix problem with mean offset in LinearWithAuxLoss.
2022-11-28 09:46:01 +08:00
Daniel Povey
9e7add6be8
Work out alpha (scale on z) in LinearWithAuxLossFunction
2022-11-27 23:48:26 +08:00
Daniel Povey
a610011c3c
Partially revert sign_gain_factor
2022-11-27 17:18:33 +08:00
Daniel Povey
30d0bc6ad7
Make gain factor 4 times larger, for constraining the sign in ActivationBalancer.
2022-11-27 17:17:11 +08:00
Daniel Povey
ff361a7495
Change default prob on limit_param_value from 0.2 to 0.6.
2022-11-27 14:00:59 +08:00
Daniel Povey
a96b92fb54
Make alpha for LinearWithAuxLossFunction be in log space; simplify/rework NonlinAttentionModule, setup more like ConvModule now.
2022-11-26 19:38:29 +08:00
Daniel Povey
110c2601ab
Changes for speed
2022-11-26 14:38:16 +08:00
Daniel Povey
c653c66413
Undo cast in autocast mode.
2022-11-26 14:29:49 +08:00
Daniel Povey
d1ee1f2d98
Try to save memory in autocast mode.
2022-11-26 14:25:27 +08:00
Daniel Povey
7b5c0382f9
Fix to LinearWithAuxLoss for bias=False case
2022-11-26 14:16:53 +08:00
Daniel Povey
1ebc3dd158
Bug fixes to LinearWithAuxLoss
2022-11-25 16:20:28 +08:00
Daniel Povey
0a997d64c4
Fixes for half precision
2022-11-25 16:07:47 +08:00
Daniel Povey
6a91f343e9
Use LinearWithAuxLoss in squeeze-attention module
2022-11-25 16:04:51 +08:00
Daniel Povey
ee61ec63b3
Introduce schedules for whitening.
2022-11-23 19:49:34 +08:00
Daniel Povey
1d0252d420
Merge branch 'scaled_adam_exp466' into scaled_adam_exp472.
...
Below is a more complete list of the changes I am making, although some of
these may be counted in the last
numbers XXX below correspond to branches numbered scaled_adam_expXXX.
- from 412/413 (cherry-picked): dropout for attention in attention_squeeze and nonlin_attention modules,
but simplified this a little to use the same dropout schedule and drop them out all together
also have all 3 submodules use separate heads.
- from 460->461, which is in the history of 464, revert the part about balancing output out attention_squeeze module.
- merge from 462->467, about using TanSwish not tanh.
- merge 462->465, remove whitening in self-attention module
- merge the part of 465->466 that was about diagnostics (name in Whiten module)
2022-11-23 14:41:09 +08:00
Daniel Povey
066f1e4658
Implement TanSwish(), use it as activation in AttentionSqueeze module.
2022-11-22 23:34:11 +08:00
Daniel Povey
6c5763fbb3
Implement subtracted momentum [0.33,0.66], and print name in Whiten module.
2022-11-22 21:57:48 +08:00
Daniel Povey
4e21db07f6
Remove activation in AttentionSqueeze; add balancers; fix bugs RE balancers.
2022-11-19 22:05:10 +08:00
Daniel Povey
6ea1706e11
Fix potential/theoretical issue in backward of LimitParamValue
2022-11-14 23:31:00 +08:00
