-
Notifications
You must be signed in to change notification settings - Fork 15
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Changes from vijay to the xconfig branch. #30
base: xconfig
Are you sure you want to change the base?
Conversation
vijayaditya
commented
Nov 8, 2016
- Moved library files to a sub-directory as we will have more library files in PR 1066
If you move them like that, perhaps you could rename them to e.g. On Mon, Nov 7, 2016 at 7:45 PM, Vijayaditya Peddinti <
|
@vimalmanohar I have created the package structure according to our discussion. |
@freewym Would you be able to help in this PR ? You might be maintaining these scripts in the future so it would be good to familiarize yourself. Could you look at the LSTMP xconfig and write one for the normal LSTM (i.e., without the projection layers) ? |
OK. |
@vijayaditya I made a PR to your repo to add an LSTM's xconfig layer (w/o projection) |
@danpovey could you check the latest commit and let me know if you agree with the Nones class that I am using. This was a hacky way to avoid explicitly specifying the type of config variable when we don't know what the default has to be and want the C++ code to use the default it wants. |
Regarding the Nones stuff... I think negative values would normally be easier to follow. The python could just, say, not print out the config value in the config file if the source was -1, or unset. |
The negative values would not work when we would like to specify things like means (we support bias-mean in NaturalGradientAffineComponent), so I thought this was a consistent way to just say we don't what this value is. |
Sorry, I think it's going to make the code harder for newbies to understand. I'd rather keep it obvious... e.g. in that case of the bias-mean [which it's not clear we would ever need to expose at the xconfig level], you could just leave the default at zero. For cases where we want to, say, set obscure options like natural-gradient options, we could have a config like Also, remember that the script prints out the 'xconfig.expanded', and your nones would appear there and would look rather ugly. I'd rather have more human-interpretable unset values. [and note: if we eventually allow strings with = in them, they'd have to be printed quoted in xconfig.expanded.] |
OK got it. |
7763f2b
to
dbfe938
Compare
…g-fix Fix nnet3 endpointing to correctly use frame subsampling factor (kaldi-asr#1184)
… for computation in nnet3 setup. For more details see Issue kaldi-asr#1190 (kaldi-asr#1194)
It would be nice to see your example scripts that you are using for testing. |
BTW, if you make this a PR to the Kaldi repo it would be better, then others (not just me) will see it. |
Fix an asymmetry in how the derivatives were truncated outside the chunk for BLSTM training. [Caution: may change BLSTM results.] In the nnet3 training code there is a mechanism --{min,max}-deriv-time to stop processing derivatives outside of a particular time range, which can be used to stop wasteful and possibly harmful computation in the e.g. +-40-frame context outside of the chunk boundaries where the supervision lies. [E.g. the gradients may blow up there.] Due to a previous oversight, this was previously only applied on the left, i.e. the python script set the --min-deriv-time not the --max-deriv-time. This commit fixes that, and also tunes the time values used in the scripts, to limit the derivatives to +-10 frames around the supervised chunk. Results for BLSTM training are improved where tested. Caution: if you are tuning BLSTM things, you may need to re-run baselines after you merge this change.
1f990a4
to
dc26602
Compare
Added (B)LSTM scripts for ami/s5b and tedlium/s5_r2
72e3261
to
4611e41
Compare
…ontext. This prevents a crash that has been happening since pull request kaldi-asr#1197.
9a817a3
to
5026607
Compare
2. Added ability to provide strings with '=' as values in xconfig. 3. Added swbd recipes for TDNN, LSTM and BLSTM using xconfig.
5026607
to
22c45ba
Compare
…quired context values were incorrect.