在TensorFlow中使用SciPy函数：ValueError（'invalid axis'）

Question

我正在使用TensorFlow训练自动编码器。因为我的输入通常很吵，所以我想使用SciPy的ndimage.gaussian_filter1d根据过滤后的数据评估网络的损失。当我将TensorFlow张量输入此函数时，我得到一个ValueError: invalid axis（见下文）。谁能建议我如何解决或解决此问题？

重现该错误的最小示例：

import tensorflow as tf
from tensorflow.contrib.layers import fully_connected
import numpy as np
import matplotlib.pyplot as plt
from scipy.ndimage import gaussian_filter1d


def run_trainer():

    tf.reset_default_graph()

    N_data = 1000

    noise = 0.2 * (np.random.rand(N_data) - 0.5)
    x = np.linspace(0, 2*np.pi, N_data)
    y = np.sin(x)
    y_noisy = (y + noise).reshape((1, -1))

    input = tf.placeholder(tf.float32, shape=[None, N_data])
    fc1 = fully_connected(inputs=input, num_outputs=500, activation_fn=tf.nn.relu)
    fc2 = fully_connected(inputs=fc1, num_outputs=250, activation_fn=tf.nn.relu)
    fc3 = fully_connected(inputs=fc2, num_outputs=500, activation_fn=tf.nn.relu)
    output = fully_connected(inputs=fc3, num_outputs=1000, activation_fn=tf.nn.tanh)

    smooth_input = gaussian_filter1d(input, sigma=10)
    smooth_output = gaussian_filter1d(output, sigma=10)
    loss = tf.reduce_mean(tf.square(smooth_input - smooth_output))
    # Comment out this guy for working version
    # loss = tf.reduce_mean(tf.square(input - output))

    optimiser = tf.train.AdamOptimizer(2e-5)
    train_step = tf.contrib.training.create_train_op(loss, optimiser)

    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True

    with tf.Session(config=config) as sess:
        sess.run(tf.global_variables_initializer())

        for i in range(int(1e3)):

            _, loss_train = sess.run(
                [train_step, loss],
                feed_dict={input: y_noisy}
            )

            if i % 100 == 0:
                print("step: %i \t loss: %.3e" % (i, loss_train))

        result = output.eval(feed_dict={input: y_noisy})

        plt.plot(x, y_noisy.flatten())
        plt.plot(x, result.flatten())
        plt.show()

if __name__ == "__main__":
    run_trainer()

错误消息：

Traceback (most recent call last):
  File "/***/stack_TF_axis.py", line 60, in <module>
    run_trainer()
  File "/***/stack_TF_axis.py", line 27, in run_trainer
    smooth_input = gaussian_filter1d(input, sigma=10)
  File "/***/anaconda3/envs/TensorFlow/lib/python3.6/site-packages/scipy/ndimage/filters.py", line 204, in gaussian_filter1d
    return correlate1d(input, weights, axis, output, mode, cval, 0)
  File "/***/anaconda3/envs/TensorFlow/lib/python3.6/site-packages/scipy/ndimage/filters.py", line 86, in correlate1d
    axis = _ni_support._check_axis(axis, input.ndim)
  File "/***/anaconda3/envs/TensorFlow/lib/python3.6/site-packages/scipy/ndimage/_ni_support.py", line 90, in _check_axis
    raise ValueError('invalid axis')
ValueError: invalid axis

Answer 1

基于@LukeDeLuccia的建议，我将this解决方案改编为用于一维张量的高斯滤波。作为参考，我在下面提供了一个工作示例：

import tensorflow as tf
from tensorflow.contrib.layers import fully_connected
import numpy as np
import matplotlib.pyplot as plt

# Based on: https://stackoverflow.com/a/52012658/1510542
# Credits to @zephyrus, @LukeDeLuccia, and @xdurch0


def gaussian_kernel(size, mean, std):
    d = tf.distributions.Normal(tf.cast(mean, tf.float32), tf.cast(std, tf.float32))
    vals = d.prob(tf.range(start=-size, limit=size+1, dtype=tf.float32))
    kernel = vals[:, tf.newaxis, tf.newaxis]
    return kernel / tf.reduce_sum(kernel)


def gaussian_filter(input, sigma):
    size = int(4*sigma + 0.5)
    x = input[:, :, tf.newaxis]
    kernel = gaussian_kernel(size=size, mean=0.0, std=sigma)
    conv = tf.nn.conv1d(x, kernel, stride=1, padding="SAME")
    return conv


def run_trainer():

    tf.reset_default_graph()

    # Define size of data, batch sizes
    N_data = 1000

    noise = 0.2 * (np.random.rand(N_data) - 0.5)
    x = np.linspace(0, 2*np.pi, N_data)
    y = np.sin(x)
    y_noisy = (y + noise).reshape((1, -1))

    input = tf.placeholder(tf.float32, shape=[None, N_data])
    fc1 = fully_connected(inputs=input, num_outputs=500, activation_fn=tf.nn.relu)
    fc2 = fully_connected(inputs=fc1, num_outputs=250, activation_fn=tf.nn.relu)
    fc3 = fully_connected(inputs=fc2, num_outputs=500, activation_fn=tf.nn.relu)
    output = fully_connected(inputs=fc3, num_outputs=1000, activation_fn=tf.nn.tanh)

    smooth_input = gaussian_filter(input, sigma=1)
    smooth_output = gaussian_filter(output, sigma=1)
    loss = tf.reduce_mean(tf.square(smooth_input - smooth_output))

    optimiser = tf.train.AdamOptimizer(2e-5)
    train_step = tf.contrib.training.create_train_op(loss, optimiser)

    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True

    with tf.Session(config=config) as sess:
        sess.run(tf.global_variables_initializer())

        for i in range(int(1e4)):

            _, loss_train = sess.run(
                [train_step, loss],
                feed_dict={input: y_noisy}
            )

            if i % 100 == 0:
                print("step: %i \t loss: %.3e" % (i, loss_train))

        result, smooth_result = sess.run(
            [output, smooth_output],
            feed_dict={input: y_noisy}
        )

        plt.plot(x, y_noisy.flatten())
        plt.plot(x, result.flatten())
        plt.plot(x, smooth_result.flatten())
        plt.show()


if __name__ == "__main__":
    run_trainer()