Description

I trained a transformer model for English to French translation. It is working well when I give it a sentence to translate. However, when I give a whole document (or simply a paragraph with many sentences), it gives me back a very bad translation and sometimes it skips some sentences.

Did anyone encounter this kind of problem ?

We should generate a proper API documentation based on PyDoc strings. The question are:

1. How to make it look nice?
2. How to integrate it into the documentation?

Should finished #23 before doing this.

Why are tgt_vocab and normalization specified here as arguments when the constructor doesn't take them, but rather an unspecified argument criterion?

https://github.com/OpenNMT/OpenNMT-py/blob/5ea9cb2fa0d12e1f8b3b835ecf51a4b949fc0a14/onmt/utils/loss.py#L67-L89 # #

Is there a tutorial on how to train a new model with new custom dataset? I would like to train new ASR model for different languages but Lingvo does not seem to have a straightforward API for training a new model. Are you planning on writing a tutorial or can you point me to an existing one? The codelab example is not helpful in this regard.

Current documentation in README explains how to install the toolkit and how to run examples. However, I don't think this is enough for users who want to make some changes to the existing recipes or make their own new recipe. In that case, one needs to understand what run.sh does step by step, but I think docs for that are missing at the moment. It would be great if we provide documentation for:

1. Add CI test for building documentations (Do not ignore warnings and add spellcheck).
2. Fix docstrings with incorrect/inconsistent Sphinx format. Currently, such issues are treated as warnings in the docs building.

• We update papers about machine translation from top conferences, including ICLR, NeurIPS, ICML, ACL, EMNLP, NAACL, COLING, EACL and so on, as well as top journals including CL and TACL.
• Currently, we only add officially published papers as well as the archived papers which have triggered heated discussion (BERT, for example!).
• For those recently archived papers, or insightful but accidenta

fastText updated their embeddings. The link you have on the README no longer points to an existing page.
The alignment matrices you have were are binded to the embeddings of fastText that could still be obtained in version 0.2.0, 19 Dec 2018 .
Links to those embeddings can be found in the file [pretrain

Hi,
Is it possible to add benchmarks of some models into documentation for comparison purposes ?
Also run time would be helpful. For example 1M iteration takes a weekend with GTX 1080.

When there's no comment, extractlocstrings will generate entries like
/* No comment provided by engineer. */ all over the place. These comments should be removed by default on normalization. There might even be an option to remove comments alltogether.

• TransformerDecoder.forward: where does self.training come from?
  https://github.com/asyml/texar-pytorch/blob/d17d502b50da1d95cb70435ed21c6603370ce76d/texar/torch/modules/decoders/transformer_decoders.py#L448-L449

• All arguments should say their types explicitly in the docstring. E.g., what is the type of infer_mode? The [method signature](https://texar-pytorch.readthedocs.

Based on this line of code:
https://github.com/ufal/neuralmonkey/blob/master/neuralmonkey/decoders/output_projection.py#L125

Current implementation isn't flexible enough; if we train a "submodel" (e.g. decoder without attention - not containing any ctx_tensors) we cannot use the trained variables to initialize model with attention defined because the size of the dense layer matrix input become

The list is awesome. It'd be great if you can add Texar to the list :)

Texar is an open source toolkit aiming to support a broad set of machine learning especially text generation tasks, such as machine translation, dialog, summarization, content manipulation, language modeling, and more. Texar is based on TensorFlow, and is designed to be highly modularized, versatile, and extensible, to suppo

Documentation of MMT CLI is lacking a lot of small information. Things like:

1. For this command to work:
   ./mmt import -e ENGINE -p SOURCE_FILE TARGET_FILE
   SOURCE_FILE and TARGET_FILE needs to end to language IDs.
   I tried importing files with .txt extension but the number of units in
   curl -X GET http://localhost:8045/memories/imports/00000000-0000-0000-0000-000000000001
   were 1/10th