Tensorflow 1.4.0-rc1 CTC Loss Simple Test

CTCLossSimpleTest.py

import numpy as np
import tensorflow as tf

def CTCLossSimpleTest():
    """
    Alex Graves Paper example:
    #      a    -
    P = [[0.3, 0.7], # t = 0
         [0.4, 0.6]] # t = 1

    P(a) = P(a-) + P(-a) + P(aa) = 0.3*0.6 + 0.7*0.4 + 0.3*0.4 = 0.58
    negative log prob = -ln(0.58) = 0.544727175
    """
    # vocabulary list
    vocabularies = 'a'

    # input tensor
    #      a    -
    P = [[0.3, 0.7], # t = 0
         [0.4, 0.6]] # t = 1
    print(P)
    # might be something wrong with precision here?
    
    # convert to float32 array
    P = np.array(P, dtype=np.float32)
    
    max_time, num_classes = P.shape
    assert num_classes == len(vocabularies) + 1
    
    # then to convert tensor
    P = tf.convert_to_tensor(P)

    # expand dims for batch dimension
    P = tf.expand_dims(P, axis=1)

    assert P.shape == (max_time, 1, num_classes)  # shape = (Max_time, batch_dim, num_classes)
    
    # convert label string to list of indices
    label = 'a'
    label = [vocabularies.index(char) for char in label]
    
    # calculate sequence length
    sequence_length = len(label)
    
    # convert label to tf.SparseTensor
    value = label
    index = [[0, i] for i in range(sequence_length)]
    dense_shape = [1, sequence_length]
    labels = tf.SparseTensor(indices=index, values=value, dense_shape=dense_shape)
    
    # convert sequence length to tensor
    sequence_length = np.atleast_1d(sequence_length)
    sequence_length = tf.convert_to_tensor(sequence_length, dtype=tf.int32)
    
    # more assertions
    assert P.shape == (max_time, 1, num_classes)  # shape = (Max_time, batch_size, num_classes)
    assert sequence_length.shape == (1,)  #shape = (batch_size)

    # CTC loss
    loss = tf.nn.ctc_loss(labels, P, sequence_length)
    # run session
    sess = tf.Session()
    prob = sess.run([loss])
    
    # compare truth prob to calculated prob
    true_prob = 0.58
    neg_log_true_prob = -1. * np.log(true_prob)
    print('true negative log prob = %.4f' % neg_log_true_prob)
    print('Tensorflow CTC calculated negative log prob = %.4f' % prob[0])
    assert prob[0] == neg_log_true_prob

CTCLossSimpleTest()