nicolay-r / arekit Goto Github PK

Such parameters were not passed:

The idea comes from an application of SynonymsCollection, which is (might be) useless during neural networks training process on already serialized data. However the latter is important on input data serialization stage.
Therefore, there is a need to split DataIO API of experiment subfolder onto:

1. SerializationData
2. TrainingData
   Considering renaming from DataIO -> BaseData.

Update the latter (c7da1a0 ) in 0.20.5, in order to provide and support an optional separation of a model source dir (which keeps input samples/opinions) from the related results. By default, it is considered that the result dir equals to model dir.

We provide embedding at nn compilation stage.

From WIMS-2020 paper.
This parameter might be significant in terms of affection on results.
Using \thetta results in reduced amount of contexts to be extracted.

0.19.3-lrec-rejected-rc:

Добавить в класс Entity такое своство, которое возвращает int либо None.
Для чего это необходимо:

1. Так можно уйти от испльзования коллекции синонимов, если синонимы применялись при составлении коллекции (уже были размечены сущности) => GroupIndex будет предоставлять такой индекс
2. Получение правильной связи синонимов, так как сейчас при обращении в коллекцию используется лемматизация, что искажает результат
3. Ускорение загрузки и обработки данных (за счет отключения коллекции синонимов)
4. Удаление чтения синонимов из RuAttittudes.

There is a need to provide another evaluation mode for 2-scale output classification format.

This issue proceeds the #37 .

As we now declare all the synonyms collection within the related experiment, there is no need to provide a custom SynonymsCollection iniitializer that supports the case of entries duplication.

As all words has unicode type, we can only demarcate paddings (using -1 placeholder) from the actual sample content (words).
Entities remain equal 0 at this point.
Here is a method:

Problem: Optimizer was declared and intialized as a static variable of the DefaultNetworkConfig class:

In some cases, we may consider to ModelIO and Callback as None, especially when there is no need to perform evaluation and utilize neural networks (ModelIO = None).
How to solve: merge changes from the related update in 0.20.4 version (57fa08d) to 0.20.5 version.

Provide evaluation of etalon results against etalon collection, in order to determine the maximal F1, P, R for a particular source (RuSentRel)

There is no need to store all the parsed news since sampling is an iterative process

From nn benchmark, there is a problem that such models are stuck at 1-st iteration.

There is a need to serialize such tags.

Possible modifiable parameters are:
dropout,
embedding_dropout

for opimizer:
learning rate

Now we keep all the documents from the related output, while later on we utilize only those one that supposed for comparison:
DocumentsOperations.iter_doc_ids_to_compare

Due to evaluation process assumes to perform mapping of model results towards etalon results, there is a need to utilize Synonyms in evaluation process.
It is also used for etalon collection initialization.

In general, it is important to have a read-only synonyms collection which could cover entries of input examples of a variety sources types, such as train, test, (dev) simultaneously.

It is necessary to clarify whether the source of synonym is trusted or not.
Here 'trusted' means that the synonyms collection has been obtained from the same corpora as opinions, i.e. we guarantee the absence of duplicated synonymous opinions within a document.
Otherwise, we are able to skip (ommit) a duplicated opinions during OpinionCollection initialization

Using specific experiment method that we refer to during experiment data serialization.
This allows us demarcate two type of SynonymCollection:

1. For results evaluation (OpinionOperations)
2. For additional search of synonymous entities during serialization stage. (described above)

Solution: Use merged collection only during serialization stage, for entity values only (not for opinions)!

We utilize vocab.npz and term_embedding.npz for so.
In case of k-fold CV, these files becomes rewritten k times.

Now we consider a news_id as a parameter, which is unique within a particular collection. Using multiple collections may result in news_ids overlapping. Hence, there is a need in ids mapping.

The latter causes Memory Limit Exceeded.
The problem arises from the fact that NetworkSingleTextProvider performs caching embeddings for every term.

Incorrect usage, we should avoid usage of StemmerBasedSynonymsCollection for RuAttitudes. Instead, it is assumed to obtain entity indices from the related labeling in RuAttitudes collection.

In https://github.com/nicolay-r/are-networks-experiments, 0.20.5-rc:

In prior versions:
0.19.4-lrec-rejected-rc

0.20.3-wims-rc: