Dmitriy Serdyuk
Dmitriy Serdyuk
@dwf , sure, I got some inspiration from your implementation of batch normalization. I'm trying to have a more general solution which makes use of the stuff we made before,...
I don't like that by default blocks overrides the model file with parameters corrupted with NaNs and I have to retrain the model from scratch. So throwing an exception seems...
Is it better to thow an exception in `Dump` class when trying to serialize NaN?
@dwf Yes, sure. It can be solved with a temp file.
Did you install requirements like in [documentation](https://blocks.readthedocs.io/en/latest/setup.html)? ``` pip install git+git://github.com/mila-udem/blocks.git \ -r https://raw.githubusercontent.com/mila-udem/blocks/master/requirements.txt ``` Blocks is not maintained anymore, so expect all kinds of rough edges.
@rizar , are you applying aggregations after replacements? Seems, that it should work.
I'm not sure that we should add a new callback. Blocks now have the same situation with `every_n_epochs`, for example, it turns into `after_epoch` inside the callback. I vote to...
Agree. It should be an option to drop the last batch.
@nouiz , it is a bit-wise or operator which can be overridden with `__or__`. I think about implementing a proxy object in the case if transformer was created without the...
Why not make it always optional? It shouldn't break others' code. When an outside user decides to use piping syntax, she just rewrites her custom transformers.