Adding dropout schedule option to nnet3 #1248

vimalmanohar · 2016-12-05T18:37:44Z

This PR is a work in progress implementing dropout schedule in nnet3 training. See issue #1247.

vimalmanohar · 2016-12-07T18:47:40Z

@GaofengCheng @pegahgh Please review this PR and test this in your experiments.

pegahgh · 2016-12-07T19:48:16Z

Thanks. I will test it now.

On Wed, Dec 7, 2016 at 1:47 PM, Vimal Manohar ***@***.***> wrote: @GaofengCheng <https://github.com/GaofengCheng> @pegahgh <https://github.com/pegahgh> Please review this PR and test this in your experiments. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AI-PtDdRGEjFvzxdIBoTsyYNEl78rZWcks5rFv9QgaJpZM4LEkLv> .

GaofengCheng · 2016-12-07T22:22:45Z

@vimalmanohar Thanks Vimal, I'm preparing to PR tensorflow+kaldi right now, after this I will continue my experiments based on your PR and review.

Conflicts: egs/wsj/s5/steps/libs/nnet3/train/common.py

pegahgh · 2016-12-08T23:03:20Z

egs/wsj/s5/steps/libs/nnet3/train/chain_objf/acoustic_model.py

+
+    logger.info("On iteration {0}, learning rate is {1}"
+                "{dropout_info}{shrink_info}.".format(
+        iter, learning_rate(iter, num_jobs,


change this to learning_rate!

danpovey · 2016-12-08T23:05:42Z

egs/wsj/s5/steps/libs/nnet3/train/chain_objf/acoustic_model.py

+            dropout_proportions, raw_model_string)
+        dropout_info_str = ', {0}'.format(", ".join(dropout_info))
+
+    shrink_info_str = ' and shrink value is {0}'.format(shrinkage_value)


unrelated, but can we just not print the shrink info if it's 1.0?

pegahgh · 2016-12-09T20:14:43Z

I am testing new dropout schedule method on one of my old models with dropout component and it is still running and the training looks fine.

GaofengCheng · 2016-12-11T12:51:07Z

egs/wsj/s5/steps/libs/nnet3/train/common.py

@@ -530,6 +731,30 @@ def __init__(self):
                                 Note: we implemented it in such a way that it
                                 doesn't increase the effective learning
                                 rate.""")
+        self.parser.add_argument("--trainer.dropout-schedule", type=str,
+                                 dest='dropout_schedule', default='',


default value here better be None, when run steps/nnet3/chain/train.py, your dropout default type is None:

if args.dropout_schedule is not None: dropout_schedule = common_train_lib.parse_dropout_option( num_archives_to_process, args.dropout_schedule)

not null string, this will conflict.

I fixed this.

…dule

GaofengCheng · 2016-12-12T02:54:46Z

@vimalmanohar in the old kaldi version, we'd remove dropout components before the max_combine_iter, do you do this under the xconfigs? or just adding dropout all the way to the end of training?

danpovey · 2016-12-12T02:57:11Z

We don't do that.. it is possible, but I'd have to think about how to do it the right way. It's not really important since the dropout scale is zero by the end of training in the setups we envisage.

…

On Sun, Dec 11, 2016 at 6:54 PM, Gaofeng Cheng ***@***.***> wrote: @vimalmanohar <https://github.com/vimalmanohar> in the old kaldi version, we'd remove dropout components before the max_combine_iter, do you do this under the xconfigs? or just adding dropout all the way to the end of training? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#1248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu6Ppv7bIyJ9erlYqvzk7cXz3pfOBks5rHLd4gaJpZM4LEkLv> .

GaofengCheng · 2016-12-12T03:30:38Z

@danpovey I'm testing on TDNN+LSTM AMI sdm, hoping to get good results

GaofengCheng · 2016-12-14T13:45:00Z

@vijayaditya @danpovey
AMI SMD1 TDNN+LSTM with frame dropout, schedule:0,0@0.20,0.5@0.50,0@0.50,0, I have revised some scripts to simulate my earlier experiments:
results as follow:
%WER 36.7 | 14655 94484 | 66.8 17.7 15.6 3.5 36.7 64.2 | 0.642 | exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ihmali_ld5/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys
%WER 39.9 | 14069 89978 | 63.9 21.0 15.1 3.8 39.9 63.5 | 0.630 | exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ihmali_ld5/decode_eval/ascore_9/eval_hires_o4.ctm.filt.sys
``
exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ihmali_ld5: num-iters=87 nj=2..12 num-params=43.4M dim=40+100->3741 combine=-0.15->-0.13 xent:train/valid[57,86,final]=(-2.33,-1.55,-1.54/-2.52,-2.11,-2.10) logprob:train/valid[57,86,final]=(-0.218,-0.123,-0.118/-0.276,-0.248,-0.246)
baseline: `37.6 40.9` from RESULTS_SDM

vijayaditya · 2016-12-14T16:40:37Z

Nice results. Any idea how it compares to element wise dropout from before ? Vijay

…

On Dec 14, 2016 05:45, "Gaofeng Cheng" ***@***.***> wrote: @vijayaditya <https://github.com/vijayaditya> @danpovey <https://github.com/danpovey> AMI SMD1 TDNN+LSTM with frame dropout, ***@***.******@***.******@***.***,0, I have revised some scripts to simulate my earlier experiments: results as follow: %WER 36.7 | 14655 94484 | 66.8 17.7 15.6 3.5 36.7 64.2 | 0.642 | exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ ihmali_ld5/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys %WER 39.9 | 14069 89978 | 63.9 21.0 15.1 3.8 39.9 63.5 | 0.630 | exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ ihmali_ld5/decode_eval/ascore_9/eval_hires_o4.ctm.filt.sys `` exp/sdm1/chain_cleaned/tdnn_lstm1i_4epoch_dp_test21_sp_bi_ihmali_ld5: num-iters=87 nj=2..12 num-params=43.4M dim=40+100->3741 combine=-0.15->-0.13 xent:train/valid[57,86,final]= (-2.33,-1.55,-1.54/-2.52,-2.11,-2.10) logprob:train/valid[57,86, final]=(-0.218,-0.123,-0.118/-0.276,-0.248,-0.246) baseline: `37.6 40.9` from RESULTS_SDM — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADtwoKOO3KWaXEPf5IW_6sECdiX8Cktgks5rH_LfgaJpZM4LEkLv> .

GaofengCheng · 2016-12-16T10:50:56Z

egs/wsj/s5/steps/libs/nnet3/train/common.py

+                    "order of data fractions.", value_x_pair)
+                raise ValueError
+
+            dropout_values.append(num_archives, float(dropout_proportion))


@vimalmanohar there should be dropout_values.append((num_archives, float(dropout_proportion)))

danpovey · 2017-01-08T21:18:08Z

I'd like to get the parts of this dropout stuff that we know we'll want, checked in, while leaving the uncertain parts till later.
Vimal, this means that I want to check in your code and script changes that allow a dropout schedule to be set (but not the example script or the changes to lstm.py).
@GaofengCheng, can you please submit a separate PR with just your changes to the DropoutComponent, that enable the per-frame dropout? Let's not delay on this.

GaofengCheng · 2017-01-09T01:30:20Z

egs/wsj/s5/steps/libs/nnet3/train/common.py

+            value_x_pair = parts[i].split('@')
+            if len(value_x_pair) == 1:
+                # Dropout proportion at half of training
+                dropout_proportion = float(value_x_pair)


change this into dropout_proportion = float(value_x_pair[0]) to avoid crash

danpovey · 2017-01-14T05:58:39Z

egs/wsj/s5/steps/libs/nnet3/train/common.py

+    in the option.
+
+    Arguments:
+        dropout_option: may have different rules for different component-name


You are starting with the most complex case here and not describing the most simple, common case where there are no names.
This documentation probably requires an example.

danpovey

This is about the python code for the dropout...
Right now it's rather unclear and I think the organization could be improved.

In the top-level training function you know how many archives are to be processed, and you know how many archives have been processed.
You could define a function like this:

def get_dropout_edit_string(proportion_processed,
                     dropout_opt):
   """This function returns a command that will (as part
        of a pipe) edit a raw nnet3 model to set the
       dropout values appropriately.  If dropout_opt is
       empty it will return the empty string.  Otherwise
       it will return a string like:
       "| nnet3-copy --edits='...'", where the ... contains
       one or more commands of type set-dropout-proportion
        that set the dropout proportion for different components.  
        Please see documentation for --trainer.dropout-schedule for more info."""

Then all the top-level trainer code has to do is to call get_dropout_edit_string(num_archives_processed * 1.0 / num_archives_to_process, args.dropout_schedule)
and pass the resulting string into train_one_iteration.
That substantially simplifies the code that has to interact with the dropout schedule, and it means that there are no hard-to-explain interfaces (which your code currently has).
The script has to parse it each time but it doesn't matter, speed is not a concern here.

danpovey · 2017-01-17T07:34:34Z

@vimalmanohar, I'm sure you're busy but don't lose track of this issue... also there was another comment in another thread needing your attention, search your email for "deprecated"; and of course some comments to address in the PR on discriminative training [RE deleting models]... I know this is in addition to your work on diarization etc....
Re the issue here about refactoring the dropout, it may be possible to reuse most of your existing code, just refactor a little.

vimalmanohar · 2017-01-17T18:19:28Z

get_dropout_edit_string will have to go through what these functions do anyways, i.e. call all these functions to parse the pattern-function pairs, parse each function for each pattern and then get the dropout value for a particular fraction. It can be simplified a little maybe if the functions are modified to use the data proportions directly instead of requiring a num_archives argument.

On Sat, Jan 14, 2017 at 1:13 AM Daniel Povey ***@***.***> wrote: ***@***.**** commented on this pull request. This is about the python code for the dropout... Right now it's rather unclear and I think the organization could be improved. In the top-level training function you know how many archives are to be processed, and you know how many archives have been processed. You could define a function like this: def get_dropout_edit_string(proportion_processed, dropout_opt): """This function returns a command that will (as part of a pipe) edit a raw nnet3 model to set the dropout values appropriately. If dropout_opt is empty it will return the empty string. Otherwise it will return a string like: "| nnet3-copy --edits='...'", where the ... contains one or more commands of type set-dropout-proportion that set the dropout proportion for different components. Please see documentation for --trainer.dropout-schedule for more info.""" Then all the top-level trainer code has to do is to call get_dropout_edit_string(num_archives_processed * 1.0 / num_archives_to_process, args.dropout_schedule) and pass the resulting string into train_one_iteration. That substantially simplifies the code that has to interact with the dropout schedule, and it means that there are no hard-to-explain interfaces (which your code currently has). The script has to parse it each time but it doesn't matter, speed is not a concern here. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1248 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AEATVx8ak2_G0-TxTi4TVJqIXIBQBNQOks5rSGdsgaJpZM4LEkLv> .

-- Vimal Manohar PhD Student Electrical & Computer Engineering Johns Hopkins University

danpovey · 2017-01-17T18:37:55Z

Yes, the only real simplification in terms of the quantity of code is to use the proportions directly But the point is that you are no longer exposing a complex and hard-to-document interface to the rest of the program, and there is no longer a need to document so clearly what all those quantities are, since they are segregated behind a single simpler interface. They then behave more like the internal workings of a function. They should still be documented to some extent, of course. On Tue, Jan 17, 2017 at 1:19 PM, Vimal Manohar <notifications@github.com> wrote:

…

get_dropout_edit_string will have to go through what these functions do anyways, i.e. call all these functions to parse the pattern-function pairs, parse each function for each pattern and then get the dropout value for a particular fraction. It can be simplified a little maybe if the functions are modified to use the data proportions directly instead of requiring a num_archives argument. On Sat, Jan 14, 2017 at 1:13 AM Daniel Povey ***@***.***> wrote: > ***@***.**** commented on this pull request. > > This is about the python code for the dropout... > Right now it's rather unclear and I think the organization could be > improved. > > In the top-level training function you know how many archives are to be > processed, and you know how many archives have been processed. > You could define a function like this: > > def get_dropout_edit_string(proportion_processed, > dropout_opt): > """This function returns a command that will (as part > of a pipe) edit a raw nnet3 model to set the > dropout values appropriately. If dropout_opt is > empty it will return the empty string. Otherwise > it will return a string like: > "| nnet3-copy --edits='...'", where the ... contains > one or more commands of type set-dropout-proportion > that set the dropout proportion for different components. > Please see documentation for --trainer.dropout-schedule for more info.""" > > Then all the top-level trainer code has to do is to call > get_dropout_edit_string(num_archives_processed * 1.0 / > num_archives_to_process, args.dropout_schedule) > and pass the resulting string into train_one_iteration. > That substantially simplifies the code that has to interact with the > dropout schedule, and it means that there are no hard-to-explain interfaces > (which your code currently has). > The script has to parse it each time but it doesn't matter, speed is not a > concern here. > > — > You are receiving this because you were mentioned. > Reply to this email directly, view it on GitHub > <#1248 (review) >, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AEATVx8ak2_G0- TxTi4TVJqIXIBQBNQOks5rSGdsgaJpZM4LEkLv> > . > -- Vimal Manohar PhD Student Electrical & Computer Engineering Johns Hopkins University — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1248 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu5H1oDYK8ui5MB0thmabLCm9Wvjbks5rTQYygaJpZM4LEkLv> .

danpovey · 2017-01-21T00:59:49Z

Vimal, I see that you've changed the code, but it still isn't the way I was asking for it to be.
The problem is that you are still exposing to the rest of the program this "dropout_proportions" quantity, which you are not documenting. And just documenting it won't completely solve the issue which is that the way you implemented it has a large, messy interface and not a small, simple interface. You may be concerned about efficiency, but efficiency is not an issue here.

What I was asking for is that if you just pass around the dropout option itself, and have an interface that says "here's the dropout option (maybe an empty string) and here's the proportion of data that we saw; give us the command to edit the nnet (or maybe an empty string)". Then the interface is just one function taking a string and a float, and returning one string; and as far as the rest of the program is concerned there is nothing else that needs to be documented.
You can even use your current functions if you want, just make them internal and don't expose them to the rest of the program. [But in any case you should always document the return-type of functions, and provide examples.]

danpovey · 2017-01-21T01:02:20Z

... also, I notice that this PR contains an example script, and its naming is not right (and it may depend on other stuff that's not in this PR, like per-frame dropout). I'd prefer to merge your changes RE dropout-proportion first, and later worry about Gaofeng's changes.

danpovey · 2017-01-21T01:09:24Z

Actually, I think it would be helpful to put all the internal functions needed for this in a separate module, say dropout_schedule.py, and then have just the top-level function be imported [and hence re-exported, I imagine] by common_lib.py, e.g.

from dropout_scheule.py import get_dropout_command

this will stop common_lib.py from blowing up in size too much.

vimalmanohar

I made the suggested changes.

danpovey · 2017-01-21T05:24:08Z

egs/wsj/s5/steps/libs/nnet3/train/common.py

@@ -17,12 +17,14 @@
 import shutil

 import libs.common as common_lib
+import libs.nnet3.train.dropout_schedule as dropout_schedule
+from dropout_schedule import *


I think it would be better if you just imported get_dropout_edit_string, because that's the only function we need from there, and if you just import the one function it's clear that that's the only one that's the real interface. You could rename all the others with underscores at the start of their names (assuming they really are internal to the module and assuming that's what the Google style guide recommends in such circumstances).

danpovey

Thanks for doing this, I know you have many things to do... just some small comments to address and then we'll be good to go.

danpovey · 2017-01-21T18:35:44Z

egs/wsj/s5/steps/libs/nnet3/train/dropout_schedule.py

+    """Returns dropout proportions based on the dropout_schedule for the
+    fraction of data seen at this stage of training.
+    Returns None if dropout_schedule is None.
+


Can you please give a couple of examples of what this function might return for different inputs, covering different types of input? e.g. (and this will be wrong):

e.g.: _get_dropout_proportions('0.0,0.5,0.0', 0.75) = [ ('*', 0.75) ] _get_dropout_proportions('*=0.0,0.5,0.0,lstm.*=0.0,0.3@0.75,0.0', 0.75) = \ [ ('*', 0.75), ('lstm.*', 0.3) ]

IMO it's always a good idea for this type of code to give such examples, it will
make maintenance much easier. Please give examples for other functions in this
module; and remember to cover trivial cases such as where the input is the
empty string; there may be situations where 3 or 4 examples are needed to
demonstrate the function's range of behavior (but of course we'll
assume the reader is smart enough to extrapolate things).

... actually, here's an idea (this is similar to something I did in the xconfig code),
How about having a function called _self_test(), that will actually test all of these examples, e.g.

def _self_test(): assert _get_dropout_proportions('*=0.0,0.5,0.0,lstm.*=0.0,0.3@0.75,0.0', 0.75) == \ [ ('*', 0.75), ('lstm.*', 0.3) ]

and have it called directly from __main__ so we can check that it works.
Then the documentation for the function can just say 'see _self_test() for examples'.
That way we will have confidence that the examples are actually correct.

Ok I added self_test. Should it be called every time the module is imported on only when run?

danpovey · 2017-01-21T18:37:54Z

src/nnet3bin/nnet3-copy.cc

-    if (dropout > 0)
-      SetDropoutProportion(dropout, &nnet);
+    if (dropout >= 0)
+      KALDI_ERR << "--dropout option is deprecated. "


You can just remove the dropout option from the code; it's not used in any example scripts. Only people here at JHU were probably using it and they can easily ask around or check the git history.

danpovey · 2017-01-23T03:26:15Z

Only when run from the command line as a standalone script.

…

On Sun, Jan 22, 2017 at 10:24 PM, Vimal Manohar ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In egs/wsj/s5/steps/libs/nnet3/train/dropout_schedule.py <#1248>: > + return initial_dropout + + assert (data_fraction >= initial_data_fraction + and data_fraction < final_data_fraction) + + return ((data_fraction - initial_data_fraction) + * (final_dropout - initial_dropout) + / (final_data_fraction - initial_data_fraction) + + initial_dropout) + + +def _get_dropout_proportions(dropout_schedule, data_fraction): + """Returns dropout proportions based on the dropout_schedule for the + fraction of data seen at this stage of training. + Returns None if dropout_schedule is None. + Ok I added self_test. Should it be called every time the module is imported on only when run? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1248>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVuytEBIKct1lbS3E7bxmD7jToZMRfks5rVB1VgaJpZM4LEkLv> .

vimalmanohar force-pushed the dropout_schedule branch 3 times, most recently from 9571832 to d055533 Compare December 6, 2016 19:06

vimalmanohar added 2 commits December 6, 2016 14:41

dropout_schedule: Adding dropout schedule to scripts

e97df65

dropout_schedule: Add set-dropout-proportion in nnet3 utils

8d26ce0

vimalmanohar force-pushed the dropout_schedule branch from d055533 to 8d26ce0 Compare December 6, 2016 19:41

vimalmanohar added 2 commits December 7, 2016 09:02

Changing option

1424c57

dropout_schedule: Print dropout info

818d495

dropout_schedule: Adding more comments and fixing bug

3342dd8

dropout_schedule: Bug fix

f17b0fc

Conflicts: egs/wsj/s5/steps/libs/nnet3/train/common.py

pegahgh reviewed Dec 8, 2016

View reviewed changes

dropout_schedule: Fixed bug

5a6a9b1

danpovey reviewed Dec 8, 2016

View reviewed changes

vimalmanohar added 2 commits December 9, 2016 15:06

dropout_schedule: Fixing logging

4ece089

dropout_schedule: Not printing shrinkage when its 1.0

0dd66c1

GaofengCheng reviewed Dec 11, 2016

View reviewed changes

vimalmanohar added 2 commits December 11, 2016 13:47

Merging

f6d25a2

Merge branch 'master' of github.com:kaldi-asr/kaldi into dropout_sche…

635bb6e

…dule

change dropout_parser strategy

7109c43

adding frame level dropout to TDNN+LSTM on AMI SDM1 kaldi-asr#1248

5435f23

GaofengCheng mentioned this pull request Dec 14, 2016

adding frame level dropout to TDNN+LSTM on AMI SDM1 #1264

Closed

dropout_schedule: Sorting models to combine for easy reading of values

d7ebc31

GaofengCheng reviewed Dec 16, 2016

View reviewed changes

dropout_schedule: Merging from master

8484c58

GaofengCheng reviewed Jan 9, 2017

View reviewed changes

dropout: Minor bug fix

a01ed13

danpovey reviewed Jan 14, 2017

View reviewed changes

dropout_schedule: Simplying dropout in script

e6d886a

vimalmanohar added 3 commits January 20, 2017 23:34

dropout_schedule: Simplified dropout schedule functions

4e8960b

dropout_schedule: removing example script

a6b9389

dropout_schedule: fixing minor errors

df7e7b6

vimalmanohar commented Jan 21, 2017

View reviewed changes

danpovey reviewed Jan 21, 2017

View reviewed changes

dropout_schedule: Made functions internal

c978be3

danpovey reviewed Jan 21, 2017

View reviewed changes

vimalmanohar added 3 commits January 22, 2017 22:03

dropout_schedule: Added self test

e9d498b

dropout_schedule: removing dropout option

d8adee9

dropout_schedule: Add more examples

09cc27b

dropout_schedule: Made self_test to not run on import

2e94018

danpovey changed the title ~~WIP: Adding dropout schedule option to nnet3~~ Adding dropout schedule option to nnet3 Jan 23, 2017

danpovey merged commit 0440417 into kaldi-asr:master Jan 23, 2017

GaofengCheng mentioned this pull request Jan 24, 2017

About By-frame Dropout scripts Level PR #1372

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding dropout schedule option to nnet3 #1248

Adding dropout schedule option to nnet3 #1248

vimalmanohar commented Dec 5, 2016

vimalmanohar commented Dec 7, 2016

pegahgh commented Dec 7, 2016 via email

GaofengCheng commented Dec 7, 2016

pegahgh Dec 8, 2016

danpovey Dec 8, 2016

pegahgh commented Dec 9, 2016

GaofengCheng Dec 11, 2016

vimalmanohar Dec 11, 2016

GaofengCheng commented Dec 12, 2016

danpovey commented Dec 12, 2016 via email

GaofengCheng commented Dec 12, 2016

GaofengCheng commented Dec 14, 2016

vijayaditya commented Dec 14, 2016 via email

GaofengCheng Dec 16, 2016

danpovey commented Jan 8, 2017

GaofengCheng Jan 9, 2017

danpovey Jan 14, 2017

danpovey left a comment •

edited

Loading

danpovey commented Jan 17, 2017

vimalmanohar commented Jan 17, 2017 via email

danpovey commented Jan 17, 2017 via email

danpovey commented Jan 21, 2017 •

edited

Loading

danpovey commented Jan 21, 2017

danpovey commented Jan 21, 2017

vimalmanohar left a comment

danpovey Jan 21, 2017

danpovey left a comment

danpovey Jan 21, 2017

danpovey Jan 21, 2017 •

edited

Loading

vimalmanohar Jan 23, 2017

danpovey Jan 21, 2017

danpovey commented Jan 23, 2017 via email

Adding dropout schedule option to nnet3 #1248

Adding dropout schedule option to nnet3 #1248

Conversation

vimalmanohar commented Dec 5, 2016

vimalmanohar commented Dec 7, 2016

pegahgh commented Dec 7, 2016 via email

GaofengCheng commented Dec 7, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

pegahgh commented Dec 9, 2016

Choose a reason for hiding this comment

Choose a reason for hiding this comment

GaofengCheng commented Dec 12, 2016

danpovey commented Dec 12, 2016 via email

GaofengCheng commented Dec 12, 2016

GaofengCheng commented Dec 14, 2016

vijayaditya commented Dec 14, 2016 via email

Choose a reason for hiding this comment

danpovey commented Jan 8, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danpovey left a comment • edited Loading

Choose a reason for hiding this comment

danpovey commented Jan 17, 2017

vimalmanohar commented Jan 17, 2017 via email

danpovey commented Jan 17, 2017 via email

danpovey commented Jan 21, 2017 • edited Loading

danpovey commented Jan 21, 2017

danpovey commented Jan 21, 2017

vimalmanohar left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danpovey left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danpovey Jan 21, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

danpovey commented Jan 23, 2017 via email

danpovey left a comment •

edited

Loading

danpovey commented Jan 21, 2017 •

edited

Loading

danpovey Jan 21, 2017 •

edited

Loading