有没有办法计算张量流中{2}个图像之间的联合直方图

Question

我正在尝试计算tensorflow中的互信息，并将其与我的numpy代码进行比较：

import numpy as np
import tensorflow as tf
import nibabel as nib

def nmi(vol1,vol2,bin):
    m=vol1.shape[2]
    n=vol2.shape[2]

    if m>n:
       d=m-n
       slices=np.zeros([vol1.shape[0],vol2.shape[1],d+n])
       slices[:,:,0:n]=vol2
       vol2=slices
     else:
       d=n-m
       slices=np.zeros([vol1.shape[0],vol2.shape[1],d+m])
       slices[:,:,0:m]=vol1
       vol1=slices

  h,x_edges,y_edges=np.histogram2d(vol1.ravel(),vol2.ravel(),bins=bin)
  pxy=h/np.sum(h)
  px = np.sum(pxy, axis=1) 
  py = np.sum(pxy, axis=0) # marginal for y over x
  px_py = px[:, None] * py[None, :] # Broadcast to multiply marginals

＃现在我们可以使用pxy，px_py 2D数组进行计算       nzs =​​ pxy＆gt; 0＃只有非零pxy值对和

有贡献

  return np.sum(pxy[nzs] * np.log(pxy[nzs] / px_py[nzs]))

现在要与之比较的tensorflow代码如下：

def nmi_tf(x,y,bin):

    vol1=tf.placeholder(dtype=tf.float32,shape= 
    (x.shape),name='volume1')
    vol2=tf.placeholder(dtype=tf.float32,shape= 
    (x.shape),name='volume2')
    sess=tf.Session()
    tf.global_variables_initializer()
    value_range=[0,1]

    x1=tf.histogram_fixed_width(tf.cast(vol1,dtype=tf.float32),
    value_range,nbins=bin)

    x2=tf.histogram_fixed_width(tf.cast(vol2,dtype=tf.float32),
    value_range,nbins=bin)
    pxy=tf.concat([x1,x2],0)

    pxy=tf.cast(tf.histogram_fixed_width(pxy,value_range,nbins=bin),
    dtype=tf.float32)
    pxy=pxy/tf.reduce_sum(pxy)

    px=tf.reduce_sum(tf.cast(x1,dtype=tf.float32))
    py=tf.reduce_sum(tf.cast(x2,dtype=tf.float32))
    p_x_y=tf.multiply(px,py)
    nmi=tf.reduce_sum(pxy/p_x_y)
    print(sess.run(nmi,feed_dict={vol1:x,vol2:y}))