the purpose of B is to map statistics of vowel-appearance and map it to a ML-trainable format.

B iterates the characters in context by index+2 (the newspaper articles, preprocessed) and looks for vowels found in the vowel-list predefined.
If it finds a vowel in the list, several things happen.
to GT, the vowel-list-index is appended (h2)

following that (variable "r"), the code considers several positions around its current iteration.
r takes into consideration four characters by index: -2 of the vowel, -1 of the vowel, and +1 and +2 of the vowel of the current iteration (u[v+3/4])

r (using function G to create set-character-lengthed arrays) creates numpy-arrays based on the appearence of these previous and ahead-appearing characters. 

k = adjusted number of nodes in each of the hidden layers of the NN seen in Train.py's model.

r = adjusted number of epochs (training iteration loops/cycles for optimization)

m = input-text-file taken by A for preprocessing

h = position of torch.save (path to model / pickle)

Eval.py is a script taking three arguments (1 optional), input-data, model-pickle, and destination/writing file.

internally, Eval.py consists of a few functions:

	predict() takes the predictions made by train.py, turns it to a NP-array, iterates all the predicts of each vowel-case and returns the largest prediction via. argmax. 

	eval_accuracy takes pre-processed test-data as True Positives and cross-references the predictions in a simple mathematical equation saying (correct predictions)/all possible predictions

	overwriting is not so much an overwriting function as it actualy just produces a new list with the updated preditions.
	It iterates the preprocessed test data and when encountering a vowel it replaces it with the predicted one instead. 
			Had a few trouble getting the correct length to match. There are few discrepancies between test and predicted vowels throug some 
			anomalies and range-issues--- IF POSSIBLE, could you perhaps illustrate how to fix this? 

I have not done anything to optimize the pre-processing which will undoubtedly improve the performance. Here, more information about vowel-environment could be included, and low-frequency-vowels could be filtered, as low-frequency items tends to be more difficult to map/predict as less instances are known.

Evidently, the model performs better with a lower/medium number of epochs and a higher number of nodes.


Note: the performance of the model, when called through the terminal provides one accuracy / prediction, but when called in the notebook in the directory, the performance is slightly better. I cannot see how/why this is, as I have stared myself blind on the various vector-lists and vocabularies.

Perhaps you can briefly comment on where i connect things incorrectly.



The eval.py also takes a file-name for overwriting, although this has not been included for every instance. It is possible, though. 

attempts made to concatenate and calculate perplexity.

modeldropout.py contains a changed version, where dropout has been added as param to model.py.
the dropout has been set to a change-able var designable upon call/assignment.