Dzmitry Bahdanau
Dzmitry Bahdanau
It appears to me that primary key information in [geography-schema.csv](https://github.com/jkkummerfeld/text2sql-data/blob/master/data/geography-schema.csv) is wrong. For example, how can `state_name` be the primary key in the `mountain` table?
Currently it doesn't, and because of that typos like `extra_ndims` instead of `extra_ndim` go unnoticed.
It says that this method takes unnormalized probabilities as input, whereas it actually takes unnormalized *log*-probabilities.
See the [test run](https://travis-ci.org/mila-udem/blocks/jobs/188971333). After the tests finish, python crashes: ``` *** glibc detected *** /home/travis/miniconda/bin/python: free(): invalid pointer: 0x00007f539273e020 *** ======= Backtrace: ========= /lib/x86_64-linux-gnu/libc.so.6(+0x7da26)[0x7f539164ba26] /home/travis/miniconda/bin/../lib/libpython3.4m.so.1.0(+0xc098e)[0x7f539236c98e] /home/travis/miniconda/bin/../lib/libpython3.4m.so.1.0(+0x7be0f)[0x7f5392327e0f] /home/travis/miniconda/bin/../lib/libpython3.4m.so.1.0(+0x91347)[0x7f539233d347] /home/travis/miniconda/bin/../lib/libpython3.4m.so.1.0(+0xa2f0f)[0x7f539234ef0f] /home/travis/miniconda/bin/../lib/libpython3.4m.so.1.0(+0xc4ac6)[0x7f5392370ac6]...
In Blocks code `warnings.warn` and `logger.warning` are used interchangeably: ``` ╰─$ grep --include='*.py' -R -n logger.warning . ./blocks/extensions/saveload.py:157: logger.warning("No dump found") ./blocks/main_loop.py:158: logger.warning("different parameters for model and algorithm") ./blocks/main_loop.py:287: logger.warning('Received...
E.g. ``` child.allocate() File "/u/bahdanau/Dist/blocks/blocks/bricks/base.py", line 612, in allocate '{}'.format(', '.join(missing_config))) ValueError: allocation config not set: input_dim ``` Here, we could mention the brick for which the configuration is missing.
I am working on the second version of the sequence generation functionality for Blocks. I would appreciate feedback and suggestions, especially from @janchorowski , @dmitriy-serdyuk @sotelo @orhanf , cause I...
I am thinking about adding the following feature to Blocks: an keyword argument for application methods, that when `True` would trigger automatic conversion of all non-Theano variable arguments into Theano...
In practice many people do not know about all nice features of this class: `set_conditions`, `add_conditions` and so know. A compact tutorial that presents these things would be great to...
Initialization schemes from `initialization.py` should have proper `__repr__` implementation, so that the user could use the parameters, such as for instance mean and std.