m的值始终是[1, 1, 1, 1]？为什么？然而，w和b的值却正确地发生了变化。感到困惑。

import tensorflow as tfimport numpyimport osimport syslearnRateT = 0.1batchMulti = 5000x = tf.placeholder(tf.float32)xData = numpy.array([[4, 5, 3, 7], [2, 1, 2, 6], [9, 8, 7, 6], [0, 1, 9, 3], [3, 3, 0, 3], [6, 2, 5, 8], [7, 4, 4, 7], [5.0, 1, 8, 0], [5.0, 1, 1, 0], [2, 1, 3, 1]], dtype=numpy.float32)rowSize = int(xData[0].size)rowCount = int(xData.size / rowSize)yTrain = tf.placeholder(tf.float32)yTrainData = numpy.array([1, 0, 1, 1, 0, 1, 0, 0, 1, 1], dtype=numpy.float32)m = tf.Variable(tf.ones([rowSize], dtype=tf.float32), dtype=tf.float32, trainable=True)m9 = tf.fill(m.shape, 9.0)mm = tf.mod(m * 10, m9) + 1y0 = tf.floor(tf.mod(x, mm))w = tf.Variable(tf.ones([rowSize]), dtype=tf.float32)b = tf.Variable(0, dtype=tf.float32)y = tf.reduce_mean(tf.nn.sigmoid(w * y0 + b))loss = tf.abs(y - tf.reduce_mean(yTrain))optimizer = tf.train.AdadeltaOptimizer(learnRateT)train = optimizer.minimize(loss)init = tf.global_variables_initializer()sess = tf.Session()sess.run(init)totalLossSum = 0.0for i in range(batchMulti):    lossSum = 0    for j in range(rowCount):        result = sess.run([loss, y, y0, yTrain, x, w, b, train, m, mm], feed_dict={x: xData[j], yTrain: yTrainData[j]})        lossSum = lossSum + float(result[0])        if i % 1000 == 0:            print("i: %d, j: %d, loss: %10.10f, y: %f, yTrain: %f, x: %s, y0: %s, m: %s, mm: %s" % (i, j, float(result[0]), float(result[1]), yTrainData[j], xData[j], result[2], result[8], result[9]))    if i % 1000 == 0:        print("avgLoss: %10.10f(%e)" % (lossSum / rowCount, lossSum / rowCount))print("Calculate result: ------")result = sess.run([y, w, b, loss], feed_dict={x: [5.0, 1, 6, 0], yTrain: 0})print(result)

输出结果类似于：

i: 4000, j: 0, loss: 0.1410247087, y: 0.858975, yTrain: 1.000000, x: [ 4.  5.  3.  7.], y0: [ 0.  1.  1.  1.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 1, loss: 0.7396742105, y: 0.739674, yTrain: 0.000000, x: [ 2.  1.  2.  6.], y0: [ 0.  1.  0.  0.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 2, loss: 0.2074543238, y: 0.792546, yTrain: 1.000000, x: [ 9.  8.  7.  6.], y0: [ 1.  0.  1.  0.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 3, loss: 0.1410146952, y: 0.858985, yTrain: 1.000000, x: [ 0.  1.  9.  3.], y0: [ 0.  1.  1.  1.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 4, loss: 0.8225950599, y: 0.822595, yTrain: 0.000000, x: [ 3.  3.  0.  3.], y0: [ 1.  1.  0.  1.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 5, loss: 0.2410957813, y: 0.758904, yTrain: 1.000000, x: [ 6.  2.  5.  8.], y0: [ 0.  0.  1.  0.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 6, loss: 0.7717698216, y: 0.771770, yTrain: 0.000000, x: [ 7.  4.  4.  7.], y0: [ 1.  0.  0.  1.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 7, loss: 0.7733184099, y: 0.773318, yTrain: 0.000000, x: [ 5.  1.  8.  0.], y0: [ 1.  1.  0.  0.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 8, loss: 0.1566338539, y: 0.843366, yTrain: 1.000000, x: [ 5.  1.  1.  0.], y0: [ 1.  1.  1.  0.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]i: 4000, j: 9, loss: 0.1410219669, y: 0.858978, yTrain: 1.000000, x: [ 2.  1.  3.  1.], y0: [ 0.  1.  1.  1.], m: [ 1.  1.  1.  1.], mm: [ 2.  2.  2.  2.]avgLoss: 0.4135602832(4.135603e-01)

回答：

tf.floor的梯度始终是0。请参见这个问题和这个GitHub问题。因此，没有更新传递到y0之前的节点：m和mm，这就是它们始终保持不变的原因。

顺便说一句，即使它能更新，tf.mod操作也不提供梯度（参见这个问题）。