def setup_cnn(self) -> None:
"""Создание свёрточных слоёв"""
cnn_in4d = tf.expand_dims(input=self.input_imgs, axis=3)
# Параметры свёрточного ядра, выходных нейронов, stride для каждого слоя
kernel_vals = [5, 5, 3, 3, 3]
feature_vals = [1, 32, 64, 128, 128, 256]
stride_vals = pool_vals = [(2, 2), (2, 2), (1, 2), (1, 2), (1, 2)]
num_layers = len(stride_vals)
# Cоздание слоёв
pool = cnn_in4d
for i in range(num_layers):
kernel = tf.Variable(
tf.random.truncated_normal([kernel_vals[i], kernel_vals[i], feature_vals[i], feature_vals[i + 1]],
stddev=0.1))
conv = tf.nn.conv2d(input=pool, filters=kernel, padding='SAME', strides=(1, 1, 1, 1))
conv_norm = tf.compat.v1.layers.batch_normalization(conv, training=self.is_train)
relu = tf.nn.relu(conv_norm)
pool = tf.nn.max_pool2d(input=relu, ksize=(1, pool_vals[i][0], pool_vals[i][1], 1),
strides=(1, stride_vals[i][0], stride_vals[i][1], 1), padding='VALID')
self.cnn_out_4d = pool
def setup_rnn(self) -> None:
"""Создание рекуррентных слоёв."""
rnn_in3d = tf.squeeze(self.cnn_out_4d, axis=[2])
num_hidden = 256
cells = [tf.compat.v1.nn.rnn_cell.LSTMCell(num_units=num_hidden, state_is_tuple=True) for _ in
range(2)] # 2 слоя LSTM
# базовые ячейки
stacked = tf.compat.v1.nn.rnn_cell.MultiRNNCell(cells, state_is_tuple=True)
# bidirectional RNN (двунаправленная)
(fw, bw), _ = tf.compat.v1.nn.bidirectional_dynamic_rnn(cell_fw=stacked, cell_bw=stacked, inputs=rnn_in3d,
dtype=rnn_in3d.dtype)
# BxTxH + BxTxH -> BxTx2H -> BxTx1X2H (конкатенация)
concat = tf.expand_dims(tf.concat([fw, bw], 2), 2)
# выход RNN блока
kernel = tf.Variable(tf.random.truncated_normal([1, 1, num_hidden * 2, len(self.char_list) + 1], stddev=0.1))
self.rnn_out_3d = tf.squeeze(tf.nn.atrous_conv2d(value=concat, filters=kernel, rate=1, padding='SAME'),
axis=[2])