Issue 150: Long Short-Term Memory (LSTM) Binary Classifier for texts

[Hector-hedb12]

Resolves #150

[csala]

batch_size is currently ignored by the keras adapter, so this line can be removed.

[Hector-hedb12]

928cfab

[csala]

A fixed hyperparameter called "verbose" with default false should be added. See #143 .

[Hector-hedb12]

@csala according to the doc, verbose should be an integer:

verbose: Integer. 0, 1, or 2. Verbosity mode. 0 = silent, 1 = progress bar, 2 = one line per epoch.

So, a bool might not work, right? 🤔

[Hector-hedb12]

1b2b7f4

[csala]

Well, it actually does work. I guess True/False is internally interpreted as 1/0. Feel free to change it to int type.

[csala]

Please set the default to 0

[Hector-hedb12]

6a366fa

[csala]

@Hector-hedb12 The primitive looks good and can be merged, but I'm afraid the pipeline does not.

According to the Keras examples, this architecture is suitable to be used with tokenized sequences as input.
However, the pipeline that you included has two problems:

1. It uses a StringVectorizer, which produces a token count matrix instead of token sequences.
2. It is uses a validation dataset (personae) that contains more information a part from the texts, but this primitive is intended to receive as input nothing but the sequence tokens.

I think that here we need to:

1. Add a new dataset for binary text classification
2. Use the keras.Tokenizer instead of the StringVectorizer in the preprocessing steps.

In order to avoid having this PR stuck I will merge it as it is right now, but I will create a new issue to improve the pipeline.