Tensorflow 稀疏张量问题

Question

我正在尝试运行以下 TensorFlow 代码。它包括稀疏矩阵，但似乎不起作用。我修改了 tensorflow 文档中给出的示例 (Link)。我正在使用 tensorflow 版本 1.12.0。

代码：

import tensorflow as tf
import numpy as np

x = tf.sparse.placeholder(shape=[-1,8,8], dtype=np.float32)
x_reshaped = tf.sparse.reshape(x, shape=[-1,64],name='flow_sizes_reshaped')
layer = tf.Variable(initial_value=tf.random_normal(shape=[64, 32],stddev=.05), name='hidden_layer_0', dtype=np.float32)
x_final = tf.sparse.matmul(x_reshaped, layer)

with tf.Session() as sess:
  indices = np.array([[0, 2, 0], [0, 5, 1]], dtype=np.int64)
  values = np.array([1.0, 2.0], dtype=np.float32)
  shape = np.array([1, 8, 8], dtype=np.int64)

  print(sess.run(x_final, feed_dict={
    x: (indices, values, shape)}))

错误：

2018-12-11 13:24:39.039224: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Traceback (most recent call last):
  File "Test_Sparse.py", line 15, in <module>
    x: (indices, values, shape)}))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 929, in run
    run_metadata_ptr)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 1132, in _run
    raise ValueError('Tensor %s may not be fed.' % subfeed_t)
ValueError: Tensor Tensor("Const:0", shape=(3,), dtype=int64) may not be fed.

我认为这不是维度不匹配问题，因为如果我在代码中进行此更改：

indices = np.array([[2, 0], [5, 1]], dtype=np.int64)

我收到以下错误：

2018-12-11 13:30:01.538664: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
Traceback (most recent call last):
  File "Test_Sparse.py", line 15, in <module>
    x: (indices, values, shape)}))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 929, in run
    run_metadata_ptr)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 1128, in _run
    str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (2, 2) for Tensor u'Placeholder_1:0', which has shape '(?, 3)'

我尝试过很多事情，但每次都遇到这个问题。我现在正在考虑转到 TF 源代码并检查为什么会引发此错误。

Answer 1

这很奇怪，但它的发生是因为 tf.sparse.reshape。更准确地说，当 tensorflow 构建计算图时，它将所有常量张量添加为不可馈送，这是有道理的。因此，当您重塑并传递稀疏张量 x 时，它会将其 dense_shape 添加为常量张量（我认为在这个地方必须复制该张量）。最后，当您运行具有 sess.run 的图并通过 x 时，它由 3 个张量组成，包括 x.dense_shape，tensorflow 检查是否可以提供所有包含的张量并在 x.dense_shape。

另一种选择是为 x:

的形状制作一个占位符

x_shape = tf.placeholder(shape=[3], dtype=np.int64)
x = tf.sparse.placeholder(shape=x_shape, dtype=np.float32)
...
print(sess.run(x_final, feed_dict={
    x_shape: [-1, 8, 8],
    x: tf.SparseTensorValue(indices, values, shape)
}))

Tensorflow 稀疏张量问题

Tensorflow Sparse Tensors Issue

python

sparse-matrix

tensorflow