python - Tensorflow高效重叠相加

Question

我正在尝试实现 overlap-add在 Tensorflow 中，但我正在努力将 numpy output_seq[start:end] += chunk 转换为 Tensorflow。现在我是 output_seq = output_seq + tf.pad(chunk, [[start, length - end]]) 但这在长序列上真的很慢。

我也有一种预感，你可能可以通过聚集/分散来实现一些技巧，但我不太明白。以下是我的暴力尝试:

import tensorflow as tf

input = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]]

def overlap_add(overlap):
    with tf.Graph().as_default(), tf.Session() as sess:

        x = tf.constant(input)

        num_chunks = tf.shape(x)[0]
        chunk_size = tf.shape(x)[1]
        hop_length = chunk_size - overlap
        out_len = chunk_size + hop_length * (num_chunks - 1)

        y = tf.zeros((out_len,), dtype=tf.int32)

        def body(i, y):
            j = i * hop_length
            padding = [[j, out_len - (j + chunk_size)]]
            chunk = x[i]
            y = y + tf.pad(chunk, padding)
            return (i + 1, y)

        i = tf.constant(0)
        i, y = tf.while_loop(
            cond=lambda i, _: tf.less(i, num_chunks),
            body=body,
            loop_vars=[i, y])

        return sess.run(y)


for i in range(4):
    print 'overlap_add(%d): %s' % (i, overlap_add(i))

# overlap_add(0): [ 1  2  3  4  5  6  7  8  9 10 11 12]
# overlap_add(1): [ 1  2  3  9  6  7 17 10 11 12]
# overlap_add(2): [ 1  2  8 10 16 18 11 12]
# overlap_add(3): [ 1  7 18 21 19 12]

Answer 1

更新:现在有一个overlap_and_add Tensorflow 本身具有函数。

旧答案: 浏览文档并发现 unsorted_segment_sum:

import tensorflow as tf

input = [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]]

def tf_repeat(a, repeats):
    return tf.reshape(tf.tile(tf.reshape(a, [-1, 1]),
                              [1, repeats]), [-1])

def overlap_add(overlap):
    with tf.Graph().as_default(), tf.Session() as sess:

        x = tf.constant(input)
        x_flat = tf.reshape(x, [-1])

        num_chunks = tf.shape(x)[0]
        chunk_size = tf.shape(x)[1]
        hop_len = chunk_size - overlap
        flat_len = num_chunks * chunk_size
        out_len = chunk_size + hop_len * (num_chunks - 1)

        # e.g. [0,1,2,3, 2,3,4,5, 4,5,6,7] for overlap == 2
        indexes = tf.range(flat_len) - tf_repeat(tf.range(num_chunks), chunk_size) * overlap

        return sess.run(tf.unsorted_segment_sum(x_flat, indexes, out_len))


for i in range(4):
    print 'overlap_add(%d): %s' % (i, overlap_add(i))

# overlap_add(0): [ 1  2  3  4  5  6  7  8  9 10 11 12]
# overlap_add(1): [ 1  2  3  9  6  7 17 10 11 12]
# overlap_add(2): [ 1  2  8 10 16 18 11 12]
# overlap_add(3): [ 1  7 18 21 19 12]