import numpy as np
from neurox.data.writer import ActivationsWriter
def _balance_negative_class(words, activations, positive_class_size):
"""
Helper function to shuffle the negative class instances and select the number of instances equal to
the positive class. If negative class examples are fewer than positive class examples, it does not
perform any balancing and return the data as received
Parameters
----------
words: list
A list of words
activations: list
A list of word-wise activations
positive_class_size: number of words to select
Returns
-------
A list of words and activations equal or less than in size to the passed length
"""
if len(words) <= positive_class_size:
print ("No need of balancing the data. Negative class is equal or smaller in size to the positive class")
return words, activations
print ("Balancing Negative class ...")
indices = list(range((len(words))))
np.random.shuffle(indices)
swords = []
sactivations = []
for i in range(positive_class_size):
swords.append(words[indices[i]])
sactivations.append(activations[indices[i]])
return swords, sactivations
[docs]def save_files(words, labels, activations, output_prefix, output_type="hdf5", decompose_layers=False, filter_layers=None):
"""
Save word and label files in the text format and activations in the specified format (default hdf5 format)
Parameters
----------
words: list
A list of words
labels: list
A list of labels for every word
activations: list
A list of word-wise activations
output_prefix: string
Specify prefix of the output files
Returns
-------
Save word, label and activation files
"""
with open(output_prefix+".word", "w") as file:
print(*words, sep = "\n", file = file)
file.close()
with open(output_prefix+".label", "w") as file:
print(*labels, sep = "\n", file = file)
file.close()
writer = ActivationsWriter.get_writer(f"{output_prefix}.{output_type}", filetype=output_type, decompose_layers=decompose_layers, filter_layers=filter_layers)
for word_idx, word in enumerate(words):
writer.write_activations(word_idx, [word], activations[word_idx])
writer.close()
