def test_restoration_integrity(self):
char_table = CharacterTable()
char_original = 'c'
char_restored = char_table.decode(char_table.encode(char_original))
self.assertEqual(char_restored, char_original)
def test_consistency(self):
char_original = 'a'
first_encoding = CharacterTable().encode(char_original)
second_encoding = CharacterTable().encode(char_original)
self.assertEqual(first_encoding, second_encoding)
def predict(self, input_sequence):
# todo: refactor
labels = self._inference_model.predict(input_sequence)
char_table = CharacterTable()
s = ''.join([char_table.decode(label) for label in labels])
return s.strip()
def test_decode_without_blanks_and_repeatitions(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
labels = [char_table.encode(ch) for ch in self.original_text]
self.assertEqual(decoder.decode(labels), 'Helo, world!')
def test_decode_one_label_sequence(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
original = 'c'
labels = [char_table.encode(original)]
self.assertEqual(decoder.decode(labels), original)
def text_to_codes(self, text):
char_table = CharacterTable()
codes = [char_table.encode(ch) for ch in text]
blank = len(char_table)
seq = [blank]
for code in codes:
seq.append(code)
seq.append(blank)
return seq
def dummy_source():
sin = 'HHHH eee lll lll ooo ,,, www oooo rrr lll ddd'
sout = 'Hello, world'
char_table = CharacterTable()
codes = [char_table.encode(ch) for ch in sin]
x = to_categorical(codes, num_classes=len(char_table))
x = x.reshape(1, len(sin), -1)
return PreLoadedSource(x, [sout])
def test_decode_labels(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
blank = len(char_table)
num_repeated = 4
labels = [blank] * num_repeated
for ch in self.original_text:
label = char_table.encode(ch)
labels.extend([label] * num_repeated)
labels.extend([blank] * num_repeated)
self.assertEqual(decoder.decode(labels), self.original_text)
def test_decode_sequence_of_blanks(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
blanks = [len(char_table)] * 50
self.assertEqual(decoder.decode(blanks), '')
def setUp(self):
self.seqs_in = [
[[1, 1], [2, 2], [3, 3]],
[[4, 4]]
]
self.seqs_out = [
[34, 85, 23],
[28]
]
char_table = CharacterTable()
self.char_table = char_table
self.start = char_table.encode(char_table.start)
self.sentinel = char_table.encode(char_table.sentinel)
self.adapter = Seq2seqAdapter(self.start, self.sentinel,
num_classes=len(char_table))
def create_distribution(self, codes):
char_table = CharacterTable()
Tx = len(codes)
n = len(char_table) + 1
a = np.zeros((Tx, n))
for i, code in enumerate(codes):
a[i, code] = 1.0
return a
def test_mapping_is_one_to_one(self):
char_table = CharacterTable()
decoded_chars = []
for code in range(len(char_table)):
ch = char_table.decode(code)
decoded_chars.append(ch)
self.assertEqual(
len(decoded_chars), len(set(decoded_chars)),
'Got duplicate characters from different codes: {}'.format(
decoded_chars))
encoded_chars = []
for ch in decoded_chars:
encoded_chars.append(char_table.encode(ch))
self.assertEqual(
len(encoded_chars), len(set(encoded_chars)),
'2 or more characters got mapped to the same code:'.format(
encoded_chars))
def setUp(self):
transitions = {
("hello", "world"): 0.75,
("hello", "hello"): 0.25,
("world", "hello"): 0.3,
("world", "world"): 0.7,
}
self.char_table = CharacterTable()
self.hello_code = 0
self.world_code = 1
self.dictionary = WordDictionary(["hello", "world"], transitions)
def decode_next(self, prev_y, prev_state):
char_table = CharacterTable()
epsilon = 0.001
next_p = [epsilon] * len(char_table)
next_state = [epsilon] * len(char_table)
if self.counter >= len(self.result):
code = 27
else:
code = self.result[self.counter]
next_p[code] = 1.0
self.counter += 1
return next_p, next_state
def decode_next(self, prev_y, prev_state):
char_table = CharacterTable()
epsilon = 0.001
next_p = [epsilon] * len(char_table)
next_state = [epsilon] * len(char_table)
return next_p, next_state
def test_decode_sequence_with_unknown_characters(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
labels = [165, 123568, 123586]
self.assertEqual(decoder.decode(labels), '???')
def test_decode_empty_sequence(self):
char_table = CharacterTable()
decoder = CTCOutputDecoder(char_table)
labels = []
self.assertEqual(decoder.decode(labels), '')
def test_sentinel(self):
char_table = CharacterTable()
sentinel = char_table.sentinel
decoded = char_table.decode(char_table.encode(sentinel))
self.assertEqual(decoded, sentinel)
def get_initial_state(self):
char_table = CharacterTable()
epsilon = 0.001
return [epsilon] * len(char_table)
def test_decode_out_of_alphabet(self):
char_table = CharacterTable()
res = char_table.decode(len(char_table))
self.assertEqual(res, '?')
