MultiLayerPerceptron/letters.py

167 lines
6.2 KiB
Python
Raw Normal View History

import os
from pnmimage import PnmImage
class LettersData(object):
def __init__(self, data_folder="data/", filename_expected_data="list_expected_data.txt", car_filename_fmt="ext_ln{}_car{}.pgm"):
self.filename_expected_data = filename_expected_data
self.car_filename_fmt = car_filename_fmt
self.data_folder = data_folder
self.expected_data_list = []
self.expected_letters = []
self.letter_to_int = {}
self.letters_classes = {}
self.extract_vocab = []
self.vocab = {}
self._images = []
self._expected_values = []
self.nb_vocab = 0
self.input_image_size = 0
self.selection_vocab = {}
self.selection_batch = []
self.letter_to_vector = {}
self.vector_to_letter = {}
def _gen_ext_filenames_list(self):
for ln in range(100):
for car in range(100):
n = self.data_folder + self.car_filename_fmt.format(ln, car)
if os.path.exists(n) and n.replace(data_folder, '') in self.filenames_list:
yield(n)
def _get_expected_letters_list(self):
"""Get expected letters list from expected data file.
:return: list of tuples with the letter and the corresponding filename
"""
data_list = []
name = self.filename_expected_data
with open(name) as f:
lst = [(line.split(" ")[0], line.split(" ", 1)[1].replace('\n', '')) for line in f.readlines() if line.strip() != '']
# list of tuples with (letter, filename)
data_list = [(e[1], e[0]) for e in lst]
return data_list
def _extract_data(self):
"""Extract data.
Create list of expected letters/filenames, base letters with info, classes.
"""
self.expected_data_list = self._get_expected_letters_list()
"""list of tuples with letter and filename"""
self.extract_vocab = sorted(set([l[0] for l in self.expected_data_list]))
"""list of data vocab letters sorted"""
self.nb_vocab = len(self.extract_vocab)
self.vocab = {}
letters_list = [l[0] for l in self.expected_data_list]
for i, c in enumerate(self.extract_vocab):
vec = [0] * self.nb_vocab
vec[i] = 1
self.vocab[c] = {
'index': i,
'count': letters_list.count(c),
'vector': vec
}
self.letter_to_int = {c: i for i, c in enumerate(self.extract_vocab)}
"""dict of letters with index of each"""
self.letters_classes = {}
"""dict of letters with vector representation of each letter"""
2018-01-20 22:22:10 +01:00
for letter, idx in sorted(self.letter_to_int.items(), key=lambda x: x[0]):
cls = [0] * self.nb_vocab
cls[idx] = 1
self.letters_classes[letter] = cls
def get_vocab_with_min_count(self, min_count):
"""Get the vocab. A dictionary of letters with count, index and vector.
The index is re-computed and also the vector to match the number of subelements
"""
subvocab = {}
i = 0
for c, info in sorted(self.vocab.items(), key=lambda x: x[0]):
if info['count'] >= min_count:
subvocab[c] = {
'count': info['count'],
'index': i,
'vector': None
}
i += 1
nb_vocab = len(subvocab)
self.letter_to_vector = {}
self.vector_to_letter = {}
2018-01-20 22:22:10 +01:00
for c in sorted(subvocab):
vec = [0] * nb_vocab
vec[subvocab[c]['index']] = 1
subvocab[c]['vector'] = vec
self.letter_to_vector[c] = vec
return subvocab
def get_letter_of_vector(self, vector):
"""Get the letter corresponding to a given vector.
2018-01-20 22:22:10 +01:00
:param vector: a flatten list representing the vector representation of a letter
:return: the found letter else None
"""
ret = None
2018-01-20 22:22:10 +01:00
assert(type(vector) is list)
for letter, vec in self.letter_to_vector.items():
if vec == vector:
ret = letter
break
return ret
def get_batches(self, min_count=0, mini_batch_size=None):
"""Get the selection data based on min count of letters in the dataset
:param min_count: minimal count of same letters to be added (default 0 for the whole dataset)
:param mini_batch_size: size of a mini batch, if None the whole dataset size (default None)
if whole size is not factor of the mini batch size then the last mini batch
has a size < mini batch size
:return: a list of mini batches (at least list of one)
"""
if mini_batch_size is None:
mini_batch_size = len(self.expected_data_list)
self.selection_vocab = self.get_vocab_with_min_count(min_count)
self.selection_batch = []
X = []
Y = []
bsize = 0
for letter, name in self.expected_data_list:
path = '{}{}'.format(self.data_folder, name)
if letter in self.selection_vocab:
img = PnmImage()
if img.load(path) == False:
print("ERROR: failed to open '{}'".format(path))
image_size = img.get_size()
bsize += 1
X.append(img.get_data_bin())
Y.append(self.selection_vocab[letter]['vector'])
if bsize >= mini_batch_size:
self.selection_batch.append((X, Y))
X = []
Y = []
bsize = 0
if bsize > 0:
self.selection_batch.append((X, Y))
return self.selection_batch
def process(self):
self._extract_data()
def get_data_with_min_count(self, min_count):
vocab = self.get_vocab_with_min_count(min_count)
#todo
return images, expected_values
def get_vocab(self):
"""Get the vocab. A dictionary of letters with count, index and vector"""
return self.vocab
def get_classes(self):
return self.letters_classes
def get_class_element_size(self):
return self.nb_vocab
def get_input_image_size(self):
return self.input_image_size