neuralmonkey
neuralmonkey copied to clipboard
ufal
An open-source tool for sequence learning in NLP built on TensorFlow.
Should the residual connections be really handled inside the SequenceLabeler? Shouldn't the underlying encoder be in charge of that? Right now, the labeler creates two matrices, enc_out_proj_M and enc_in_proj_M to...
The Objective.name in general should have more strict naming format, otherwise it cannot be used e.g. for naming graph parts (for example naming Delayed trainer accumulators with corresponding objective names...
D401 warns about the first line of the docstring not being in the imperative mood. This is bad especially for `@property` methods. Should we turn D401 off? At least until...
In case `len(tokens) == 0` (i.e. the input sentence is empty) ChrF crashes with `ZeroDivisionError` on line 89: https://github.com/ufal/neuralmonkey/blob/master/neuralmonkey/evaluators/chrf.py#L89 ``` File "/lnet/spec/work/people/kasner/neuralmonkey-ctc-decoder/neuralmonkey/bin/neuralmonkey/evaluators/chrf.py", line 89, in chr_p return precision / float(self.max_ord)...
As discussed on the meeting, main issues are: - [x] Adapt `@tensor` annotations (introduced in #191) - [x] Base class for object with `hidden_states` and `encoded` and rename these attributes...
https://pypi.org/project/pytypes/ `check_argument_types` is supported. Additionally, pytypes can do runtime type checks on demand to mimic `inisntance` or `issubclass` behavior.
We'd like to see e.g. this: `output_fw: ▁The ▁ ▁ ▁ ▁ ▁ s ` where the last ` `is the first of ``-only tokens (or ``)
We should increase pylint limits for `too-many-arguments` and `too-many-locals` (which is influenced by the arguments which count as well towards this number). For example, the current maximum of arguments is...
When chaining multiple dataset series preprocessor steps, e.g.: preprocessors=[("source", "source_wp", ), ("source_wp", "source_wp_other", )] The dataset.load can fail because there is no implicit order of processing the preprocessors list. See:...
When initializing Dataset with buffer_size=None, attribute self.buffer_size is not created even though it can be called later in the Dataset methods. Most of the time, it is checked (by checking...