如何将tensorflow占位符变量转换为numpy数组？

Question

I would like to use scipy interpolation function in the tensorflow code.

Here is the example snippet similar to my situation.

import tensorflow as tf
from scipy import interpolate

def interpolate1D(Xval,Fval,inp): 
    Xval = np.array(Xval)
    Fval = np.array(Fval)
    f = interpolate.interp1d(Xval, Fval, fill_value="extrapolate")
    z = f(inp)
    return z

properties = {
    'xval': [200,400,600,800,1100],
    'fval': [100.0,121.6,136.2,155.3,171.0]
}

tensor = tf.placeholder("float")

interpolate = interpolate1D(properties['xval'],properties['fval'], tensor)

Once I get the interpolate I'll convert it into tensor using tf.convert_to_tensor(interpolate)

Here interpolate.interp1d is just an example. I'll be using other interpolation methods and output of those methods will be fed into another neuron.

I understand placeholder is empty variable so technically it's not possible to convert into numpy array. Also, I cannot use this interpolation function outside the tensorflow graph because in some situations I need to use output of a neural network as a input to interpolation function.

Overall, I would like to use scipy interpolation function with in the tensor graph.

Answer 1

You could use tf.py_func to use the SciPy function inside your graph, but a better option would be to implement the interpolation in TensorFlow. There is no function in the library that does this out of the box, but it is not difficult to implement it.

import tensorflow as tf

# Assumes Xval is sorted
def interpolate1D(Xval, Fval, inp):
    # Make sure input values are tensors
    Xval = tf.convert_to_tensor(Xval)
    Fval = tf.convert_to_tensor(Fval)
    inp = tf.convert_to_tensor(inp)
    # Find the interpolation indices
    c = tf.count_nonzero(tf.expand_dims(inp, axis=-1) >= Xval, axis=-1)
    idx0 = tf.maximum(c - 1, 0)
    idx1 = tf.minimum(c, tf.size(Xval, out_type=c.dtype) - 1)
    # Get interpolation X and Y values
    x0 = tf.gather(Xval, idx0)
    x1 = tf.gather(Xval, idx1)
    f0 = tf.gather(Fval, idx0)
    f1 = tf.gather(Fval, idx1)
    # Compute interpolation coefficient
    x_diff = x1 - x0
    alpha = (inp - x0) / tf.where(x_diff > 0, x_diff, tf.ones_like(x_diff))
    alpha = tf.clip_by_value(alpha, 0, 1)
    # Compute interpolation
    return f0 * (1 - alpha) + f1 * alpha

properties = {
    'xval': [200.0, 400.0, 600.0, 800.0, 1100.0],
    'fval': [100.0, 121.6, 136.2, 155.3, 171.0]
}

with tf.Graph().as_default(), tf.Session() as sess:
    tensor = tf.placeholder("float")
    interpolate = interpolate1D(properties['xval'], properties['fval'], tensor)
    print(sess.run(interpolate, feed_dict={tensor: [40.0, 530.0, 800.0, 1200.0]}))
    # [100.   131.09 155.3  171.  ]

Answer 2

The expected answer for this question is the use of tf.py_func and thanks to @jdehesa for mentioning that.

For users, who are expecting the solution for this question is mentioned below

import tensorflow as tf
import numpy as np
from scipy import interpolate

properties = {
    'xval': [200,400,600,800,1100],
    'fval': [100.0,121.6,136.2,155.3,171.0]
}

Xval = np.array(properties['xval'])
Fval = np.array(properties['fval'])
f = interpolate.interp1d(Xval, Fval, fill_value="extrapolate")

def comp(x):
    return np.float32(f(x))

tensor = tf.placeholder("float32")
interpolate = tf.py_func(comp, [tensor], tf.float32)

with tf.Session() as sess:
    print(sess.run(interpolate, feed_dict = {tensor: [230]}))