Perform separate validation and test epochs per dataset when multiple files are specified (Fixes #1634 and #2043) #2038
Conversation
DeepSpeech.py
Outdated
| set_loss = run_set('dev', epoch, init_op, dataset=csv) | ||
| dev_loss += set_loss | ||
| log_progress('Finished validating epoch %d on %s - loss: %f' % (epoch, csv, set_loss)) | ||
| dev_loss = dev_loss / len(dev_csvs) |
There was a problem hiding this comment.
Does diving by len(dev_csvs) give the same result as before this PR?
There was a problem hiding this comment.
No, if the sizes of the dev sets are unbalanced then the mean of sample means will not match the mean of the whole population. We can get to the same result by doing a weighted average if we know the sample size, will have to return that value from the run_set function since the dataset size is only known after a full iteration has been done through it. I don't expect the difference to be large but I'll fix this.
|
Validation epochs are normally so fast that I don't think I've ever actually cached them in runs, so I can't give you numbers, but that goes to show that it probably does not make a huge difference :) As for test epochs, the acoustic model prediction is usually done in the same time as the validation sets, it's the decoding that takes the majority of the time. |
9076adf to
946828c
Compare
|
All green. |
|
@kdavis-mozilla review ping |
|
@reuben As I mentioned I dont' have time to do the review. If tilman gave it a r+, I'm fine with that. |
2351bf1 to
904ab1e
Compare
|
This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs. |
As a consequence this also removes support for caching to disk the processed validation and test sets. They're usually small enough that caching to disk makes very little difference, and the complexity of keeping track of multiple cache paths for each CSV is not worth it IMO.
Sorry for the continued churn on the progress bar/logging part of the code. This PR also includes some cleanup of that code using
progressbar.NullBarso that there's now a single check for the--show_progressbarflag. With this PR, the log with and without progress bars should be consistent in the information it provides.I recommend disabling whitespace changes for reviewing the last commit that touches
evaluate.py, otherwise the diff ends up looking like a mess.