def _GetCaptchaIdsFromImageFilename(image_filepath):
captcha_str = _ParseCaptchaFromImageFilename(image_filepath)
captcha_ids = numpy.zeros(len(captcha_str), dtype=numpy.int32)
for i, captcha_char in enumerate(captcha_str):
CHAR_VOCABULARY, CHARS = vocabulary.GetCharacterVocabulary(sys.argv[2])
captcha_ids[i] = CHAR_VOCABULARY[captcha_char]
return captcha_ids
def __init__(self,
learning_rate=0.01,
no_hidden_layers=2,
includeCapital=False,
num_rnn_steps=5,
saved_params_path=None,
multi_chars=True,
num_softmaxes=None,
use_mask_input=False,
lstm_layer_units=256,
cnn_dense_layer_sizes=[256],
bidirec=False,
lstm_grad_clipping=False):
if not learning_rate:
self.learning_rate = 0.01
else:
self.learning_rate = float(learning_rate)
if not no_hidden_layers:
self.no_hidden_layers = 2
else:
self.no_hidden_layers = int(no_hidden_layers)
self.num_rnn_steps = num_rnn_steps
self.includeCapital = includeCapital
self.CHAR_VOCABULARY, self.CHARS = vocabulary.GetCharacterVocabulary(
includeCapital)
if multi_chars:
if num_softmaxes:
self._network, self._train_fn, self._test_fn, self._inference_fn = (
self._InitializeModelThatPredictsCharsMultiSoftmax(
self.learning_rate, num_softmaxes=num_softmaxes))
else:
self._network, self._train_fn, self._test_fn, self._inference_fn = (
self._InitializeModelThatPredictsAllChars(
self.learning_rate,
use_mask_input=use_mask_input,
lstm_layer_units=lstm_layer_units,
cnn_dense_layer_sizes=cnn_dense_layer_sizes,
bidirectional_rnn=bidirec,
lstm_grad_clipping=lstm_grad_clipping))
else:
self._network, self._train_fn, self._test_fn, self._inference_fn = (
self._InitializeModelThatPredictsFirstChar(self.learning_rate))
self.prediction = lasagne.layers.get_output(self._network)
if saved_params_path:
'''
If saved params path is specified, then start from that parms value.
'''
f = numpy.load(saved_params_path)
param_values = [f['arr_%d' % i] for i in range(len(f.files))]
'''
Deleting a vector as it is not needed. It was taking memory. so needed to delete it explicitly.
'''
del f.f
f.close()
lasagne.layers.set_all_param_values(self._network, param_values)