def _aon(self, inputs, is_train):
"""
"""
assert inputs.get_shape()[1:] == (26, 26, 256)
with tf.variable_scope("aon"):
hfeatures = self._shared_cnn(inputs, is_train, reuse=False)
vfeatures = self._shared_cnn(tf.contrib.image.rotate(
inputs, math.pi / 2),
is_train,
reuse=True)
hfeatures = tf.transpose(hfeatures,
perm=[1, 0, 2],
name='h_time_major')
hfeatures = rnn_layer(hfeatures, None, self.rnn_size, 'hbdrnn')
vfeatures = tf.transpose(vfeatures,
perm=[1, 0, 2],
name='v_time_major')
vfeatures = rnn_layer(vfeatures, None, self.rnn_size, 'vbdrnn')
features = (hfeatures, tf.reverse(hfeatures, axis=[0])) + \
(vfeatures, tf.reverse(vfeatures, axis=[0]))
features = tf.stack(features, axis=1)
features = tf.transpose(features, [2, 1, 0, 3])
clue = self._clue_network(inputs, is_train)
return features, clue
def get_logits(self, image, is_train, **kwargs):
"""
"""
image = tf.reshape(image, [-1, 100, 100, 1])
# BCNN
features = self._bcnn(image, is_train)
assert features.get_shape()[1:] == (26, 26, 256)
# AON
features, clue = self._aon(features, is_train)
assert features.get_shape()[1:] == (4, 23, 512)
assert clue.get_shape()[1:] == (4, 23, 1)
# FG
features = tf.reduce_sum(features * clue, axis=1)
features = tf.nn.tanh(features)
assert features.get_shape()[1:] == (23, 512)
# LSTM
features = tf.transpose(features, [1, 0, 2], name='time_major')
features = rnn_layer(features, None, self.rnn_size, 'lstm')
logits, weights = attention_decoder(features, kwargs['label'],
len(self.out_charset),
self.rnn_size, is_train,
self.FLAGS.label_maxlen)
sequence_length = None
return logits, sequence_length
def get_logits(self, image, is_train, **kwargs):
"""
"""
widths = tf.ones(tf.shape(image)[0],
dtype=tf.int32) * tf.shape(image)[2]
features, sequence_length = self._convnet_layers(
image, widths, is_train)
features = tf.transpose(features, perm=[1, 0, 2], name='time_major')
attention_states = rnn_layer(features,
sequence_length,
self.rnn_size,
scope="rnn")
logits, weights = attention_decoder(attention_states, kwargs['label'],
len(self.out_charset),
self.rnn_size, is_train,
self.FLAGS.label_maxlen)
return logits, sequence_length
def get_logits(self, image, is_train, **kwargs):
"""
"""
# ResNet
widths = tf.ones(tf.shape(image)[0],
dtype=tf.int32) * tf.shape(image)[2]
features, sequence_length = self._convnet_layers(
image, widths, is_train)
# LSTM encoder
with tf.variable_scope("rnn"):
rnn_inputs = tf.nn.max_pool(features, (1, 8, 1, 1), (1, 1, 1, 1),
'VALID',
data_format='NHWC')
rnn_inputs = tf.squeeze(rnn_inputs, axis=[1])
rnn_inputs = tf.transpose(rnn_inputs,
perm=[1, 0, 2],
name='time_major')
holistic_features = rnn_layer(rnn_inputs,
sequence_length,
self.rnn_size,
scope='holistic')
holistic_feature = dense_layer(holistic_features[-1],
self.FLAGS.rnn_size,
name='holistic_projection')
# 2D LSTM decoder
logits, weights = self.twodim_attention_decoder(
holistic_feature, features, kwargs['label'], len(self.out_charset),
self.FLAGS.rnn_size, is_train, self.FLAGS.label_maxlen)
logits = tf.reshape(
logits, [-1, self.FLAGS.label_maxlen,
len(self.out_charset) + 1])
sequence_length = None
self.attention_weights = tf.expand_dims(weights, axis=1)
return logits, sequence_length
