0

我正在尝试实施以下论文:https ://arxiv.org/abs/1904.08779 ,以便在 Speech to Text 中取得更好的效果。
我正在尝试使用mozilla DeepSpeech repo 来实现它。它使用 tensorflow 数据集模型来加载数据。

dataset = (tf.data.Dataset.from_generator(generate_values,
                                              output_types=(tf.string, (tf.int64, tf.int32, tf.int64),tf.int64))
                              .map(entry_to_features, num_parallel_calls=tf.data.experimental.AUTOTUNE)
                              .cache(cache_path)
                              .map(augment_spec, num_parallel_calls=tf.data.experimental.AUTOTUNE)
                              .window(batch_size, drop_remainder=True).flat_map(batch_fn)
                              .prefetch(num_gpus))

音频转换为频谱图并计算 mfcc,因此当数据到达 augment_spec 函数时,它的形状为 (?, 26)。? 是对可变音频长度进行整形的结果。我试图掩盖图像的某些部分,为此我想乘以张量,一个是 1 和 0 的掩码,使用类似这样的代码

def augment_spec(features, features_len, transcript):
    # print("\n\n\n\n duration", duration.eval())
    sample_rate = 8000

    mask = np.ones_like(features)

    temp = tf.Variable(tf.ones_like(features))
    print(temp)

    time_len = features_len.shape[0]
    features_len = features_len

    n_time_masks = np.random.randint(0, 4)
    n_freq_masks = np.random.randint(0, 3)

    for _ in range(n_time_masks):
        time_delta = np.random.randint(int(sample_rate / 10), int(sample_rate / 2))
        time_start = np.random.randint(0, time_len - time_delta)
        print(time_start, time_delta)
        mask[time_start:time_start + time_delta] = 0

    for _ in range(n_freq_masks):
        freq_delta = np.random.randint(1, 4)
        freq_start = np.random.randint(0, features_len - freq_delta)
        print(freq_start, freq_delta)
        mask[:, freq_start:freq_start + freq_delta] = 0

    mask = tf.convert_to_tensor(mask, dtype=tf.float32)
    return tf.math.multiply(features, mask),  features_len, transcript

问题是这些说明:

    mask = np.ones_like(features)  

    time_len = features_len.shape[0]  

由于在构建图形时张量尚未定义形状,因此不起作用,所以我不知道如何实现这一点。你能帮我解决这个问题吗?非常感谢!!

更新:在@kempy 回答之后,我的代码现在看起来像这样:

def augment_spec(features, features_len, transcript):

    # print("\n\n\n\n duration", duration.eval())
    sample_rate = 8000

    mask = tf.Variable(tf.ones_like(features),validate_shape=False)

    time_len = tf.shape(features)[0]

    n_time_masks = np.random.randint(0, 4)
    n_freq_masks = np.random.randint(0, 3)
    # n_time_masks = tf.random.uniform(
    #         shape=(), minval=0, maxval=4, dtype=tf.int32)
    # n_freq_masks = tf.random.uniform(
    #         shape=(), minval=0, maxval=3, dtype=tf.int32)

    for _ in range(n_time_masks):

        time_delta = tf.random.uniform(
            shape=(), minval=int(sample_rate / 10), maxval=int(sample_rate / 2), dtype=tf.int32)
        time_start = tf.random.uniform(
            shape=(), minval=0, maxval=time_len-time_delta, dtype=tf.int32)

        # indexes = list(range(time_start,time_start+time_delta))
        indexes = tf.range(time_start, time_start+time_delta, delta=1, dtype=tf.int32, name='range')

        tf.scatter_update(mask, indexes, 0)

    mask = tf.transpose(mask,(1,0))
    for _ in range(n_freq_masks):
        # freq_delta = np.random.randint(1, 4)
        # freq_start = np.random.randint(0, features_len - freq_delta)

        freq_delta = tf.random.uniform(
            shape=(), minval=1, maxval=4, dtype=tf.int32)
        freq_start = tf.random.uniform(
            shape=(), minval=0, maxval=(features_len - freq_delta), dtype=tf.int32)


        # indexes = list(range(freq_start,freq_start+freq_delta))
        indexes = tf.range(freq_start, freq_start+freq_delta, delta=1, dtype=tf.int32, name='range')

        tf.scatter_update(mask, indexes, 0)


    mask = tf.transpose(mask,(1,0))
    mask = tf.convert_to_tensor(mask, dtype=tf.float32)
    masked = tf.multiply(features, mask)
    return masked,  features_len, transcript

但是现在我收到了这个错误:

ValueError: Tensor("Variable:0", dtype=float32_ref) must be from the same graph as Tensor("tower_0/Mean:0", shape=(), dtype=float32, device=/device:GPU:0).

我不知道如何解决这个问题,谢谢你的帮助

4

1 回答 1

0

简短的回答

使用tf版本而不是np函数。tf.ones_like输入形状应该可以正常工作(?, 26),您可以使用它tf.shape(features)[0]来动态获取特征的形状。再往下,你应该使用类似的东西tf.random.uniform

长答案

在图形模式下运行 TF 时(这是 TF 1.X 中的默认设置),你不能让 python 代码依赖于张量的输出,因为它还没有被执行,所以你应该使用 TF ops 而不是 python麻木代码。

我们可以构建一个具有动态第一维的图:

import numpy as np
import tensorflow as tf

# Feature dimensions
unknown_size = 3
feature_dim = 26

tf.reset_default_graph()

# features_input has dynamic first dimension
features_input = tf.placeholder(tf.int32, shape=(None, feature_dim))

# ones_like should work fine with argument of shape (?, 26)
batched_ones = tf.ones_like(features_input)

# dynamically get the shape of the features_input
time_len = tf.shape(features_input)[0]
time_start = tf.random.uniform(
    shape=(), minval=0, maxval=time_len, dtype=tf.int32)

并打印以下内容:

print('features_input.shape:')
print(features_input.shape)
print('batched_ones.shape:')
print(batched_ones.shape)
print('time_start.shape:')
print(time_start.shape)

我们看到的输出是:

features_input.shape:
(?, 26)
batched_ones.shape:
(?, 26)
time_start.shape:
()

如果我们然后尝试执行图表:

with tf.Session() as sess:
  # Create some input data
  features = np.arange(feature_dim)
  batched_features = np.tile(features, (unknown_size, 1))

  # Evaluate the tensors
  features_out, ones_out, time_start_out = sess.run(
      [features_input, batched_ones, time_start],
      feed_dict={features_input: batched_features})

并打印输出:

# Print out what the output looks like
print('\nOutput:')
print('\nFeatures:')

print(features_out)
print('shape:', features_out.shape)

print('\nOnes:')
print(ones_out)
print('shape:', ones_out.shape)

print('\nRandom between 0 and unknown_size:')
print(time_start_out)
print('shape:', time_start_out.shape)

我们可以看到它有效!

Output:

Features:
[[ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
  24 25]
 [ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
  24 25]
 [ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
  24 25]]
shape: (3, 26)

Ones:
[[1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
 [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]]
shape: (3, 26)

Random between 0 and unknown_size:
0
shape: ()
于 2019-04-25T21:42:02.933 回答