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 = rnn_layers(features,
sequence_length,
self.rnn_size,
use_projection=True)
logits = dense_layer(features,
len(self.out_charset) + 1,
name='logits')
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)
attention_states = rnn_layers(features,
sequence_length,
self.rnn_size,
use_projection=True)
attention_states = dense_layer(attention_states,
self.rnn_size,
name='att_state_dense')
logits, weights = attention_decoder(attention_states,
kwargs['label'],
len(self.out_charset),
self.rnn_size,
is_train,
self.FLAGS.label_maxlen,
cell_type='gru')
return logits, sequence_length