neuro-lab8/model.py

import tensorflow as tf

from tensorflow.keras import layers as kl
from tensorflow.keras import models as km
from tensorflow.keras import losses as ks
from tensorflow.keras import optimizers as ko
from tensorflow.keras import callbacks as kc

from tensorflow import keras

from loss import CTCLoss

#g = tf.config.experimental.list_physical_devices('GPU')
#for i in g:
#  tf.config.experimental.set_memory_growth(i, True)


def model(input_dim, output_dim, rnn_layers = 5, rnn_units = 128):
    li = kl.Input((None, input_dim))
    l1 = kl.Reshape((-1, input_dim, 1))(li)

    l2 = kl.Conv2D(32,
                       kernel_size = [11, 41],
                       strides = [2, 2],
                       padding = 'same',
                       use_bias = False)(l1)
    l3 = kl.BatchNormalization()(l2)
    l4 = kl.ReLU()(l3)

    l5 = kl.Conv2D(32,
                       kernel_size = [11, 21],
                       strides = [1, 2],
                       padding = 'same',
                       use_bias = False)(l4)
    l6 = kl.BatchNormalization()(l5)
    l7 = kl.ReLU()(l6)

    lb = kl.Reshape((-1, l7.shape[-2] * l7.shape[-1]))(l7)

    for i in range(rnn_layers):
        i += 1

        r = kl.GRU(rnn_units,
                   activation = 'tanh',
                   recurrent_activation = 'sigmoid',
                   use_bias = True,
                   return_sequences = True,
                   reset_after = True)

        lb = kl.Bidirectional(r,
                              merge_mode = 'concat')(lb)

        if i < rnn_layers:
            lb = kl.Dropout(rate=0.5)(lb)

    lc1 = kl.Dense(rnn_units * 2, activation = 'relu')(lb)
    lc2 = kl.Dropout(0.5)(lc1)
    lo = kl.Dense(output_dim + 1, activation = 'softmax')(lc2)

    m = keras.Model(li, lo)
    m.compile(optimizer = ko.Adam(0.0001),
              loss = CTCLoss)

    return m