python - Tensorflow 错误：没有为任何变量提供梯度，请检查您的图表中是否存在变量之间不支持梯度的操作
问题描述
我在使用 tensorFlow 时遇到了麻烦。以下代码是可以的。
来自未来的进口部门，print_function，absolute_import
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import tensorflow.contrib.slim as slim
from tensorflow.contrib.layers.python.layers import initializers
from tensorflow.python.ops import init_ops
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# Visualize decoder setting
# Parameters
training_epochs = 100
batch_size = 50
display_step = 40
LEARNING_RATE = 0.001
# LEARNING_RATE_BASE = 0.01
# LEARNING_RATE_DECAY = 0.99
min_after_dequeue = 1000
TRAINING_SAMPLE_SIZE = 3365
IMAGE_SIZE1 = 40
IMAGE_SIZE2 = 80
IMAGE_CHANNEL = 1
DEEP_SIZE = 8
def inference(x, reuse=False):
# Building the encoder
with tf.variable_scope('inference') as scope:
if reuse:
scope.reuse_variables()
with slim.arg_scope([slim.conv2d], padding='SAME',
weights_initializer=initializers.xavier_initializer(),
weights_regularizer=None,
biases_initializer=init_ops.zeros_initializer(),
biases_regularizer=None,
normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.relu):
net = slim.conv2d(x, DEEP_SIZE, [2, 2], stride=1, scope='conv1')
net = slim.conv2d(net, DEEP_SIZE*2, [2, 2], stride=1, scope='conv2')
net = slim.conv2d(net, DEEP_SIZE, [2, 2], stride=1, scope='conv3')
# net = slim.conv2d(net, DEEP_SIZE*2, [2, 2], stride=1, scope='conv4')
# net = slim.conv2d(net, DEEP_SIZE, [2, 2], stride=1, scope='conv5')
net = slim.conv2d(net, IMAGE_CHANNEL, [2, 2], stride=1, activation_fn=None, scope='conv6')
return net
# Construct model
def train():
# tf Graph input (only pictures)
x = tf.placeholder(tf.float32, [batch_size, IMAGE_SIZE1, IMAGE_SIZE2, IMAGE_CHANNEL])
y = tf.placeholder(tf.float32, [batch_size, IMAGE_SIZE1, IMAGE_SIZE2, IMAGE_CHANNEL])
y_pred = inference(x)
y_true = y
cost = tf.reduce_mean((y_pred - y_true) ** 2)
# cost = - tf.reduce_mean(y_true * tf.log(tf.clip_by_value(y_pred, 1e-10, 1.0)))
optimizer = tf.train.AdamOptimizer(LEARNING_RATE).minimize(cost)
saver = tf.train.Saver(tf.global_variables())
# Launch the graph
with tf.Session(config = tf.ConfigProto(allow_soft_placement = True, log_device_placement = True)) as sess:
sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
total_batch = TRAINING_SAMPLE_SIZE // batch_size
# Training cycle
loss_init = 100000000
for epoch in range(training_epochs):
# Loop over all batches
loss = 0
for i in range(total_batch):
batch_xs = sess.run(input_imgs)
batch_ys = sess.run(label_imgs)
batch_ys1 = (batch_ys<0.5)*0.
batch_ys2 = (batch_ys>0.5)*1.
batch_ys = batch_ys1 + batch_ys2
# Run optimization op (backprop) and cost op (to get loss value)
_, c = sess.run([optimizer, cost], feed_dict={x: batch_xs, y: batch_ys})
# Display logs per epoch step
if i % display_step == 0:
print(epoch+1,'\t', i+1, '\t', "{:.9f}".format(c))
# Compare original images with their reconstructions
a = 0
for j in range(1):
xs = sess.run(input_imgs)
ys = sess.run(label_imgs)
ys1 = (ys<0.5)*0.
ys2 = (ys>0.5)*1.
ys = ys1 + ys2
encode_decode = sess.run(y_pred, feed_dict={x: xs, y:ys})
for i in range(5):
# plt.imshow(xs[i].reshape(IMAGE_SIZE1, IMAGE_SIZE2), cmap='Greys_r')
# plt.axis('off') #不显示坐标轴
# fig = plt.gcf()
# fig.set_size_inches(IMAGE_SIZE2/300, IMAGE_SIZE1/300)
# plt.gca().xaxis.set_major_locator(plt.NullLocator())
# plt.gca().yaxis.set_major_locator(plt.NullLocator())
# plt.subplots_adjust(top = 1, bottom = 0, right = 1, left = 0, hspace = 0, wspace = 0)
# plt.margins(0,0)
# fig.savefig('autoencoder_AE/input/%d.jpg' % a, transparent=True, dpi=300, pad_inches = 0)
plt.imshow(ys[i].reshape(IMAGE_SIZE1, IMAGE_SIZE2), cmap='Greys_r')
plt.axis('off') #不显示坐标轴
fig = plt.gcf()
fig.set_size_inches(IMAGE_SIZE2/300, IMAGE_SIZE1/300)
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top = 1, bottom = 0, right = 1, left = 0, hspace = 0, wspace = 0)
plt.margins(0,0)
fig.savefig('autoencoder_AE/label/%d.jpg' % a, transparent=True, dpi=300, pad_inches = 0)
plt.imshow(encode_decode[i].reshape(IMAGE_SIZE1, IMAGE_SIZE2), cmap='Greys_r')
plt.axis('off') #不显示坐标轴
fig = plt.gcf()
fig.set_size_inches(IMAGE_SIZE2/300, IMAGE_SIZE1/300)
plt.gca().xaxis.set_major_locator(plt.NullLocator())
plt.gca().yaxis.set_major_locator(plt.NullLocator())
plt.subplots_adjust(top = 1, bottom = 0, right = 1, left = 0, hspace = 0, wspace = 0)
plt.margins(0,0)
fig.savefig('autoencoder_AE/predict/%d.jpg' % a, transparent=True, dpi=300, pad_inches = 0)
a = a + 1
if loss < loss_init:
saver.save(sess,'AE_simp_model/autoencoder.ckpt')
loss_init = loss
print("Optimization Finished!")
train()
但是，当 inference() 已更改为：
def inference(x, reuse=False):
# Building the encoder
with tf.variable_scope('inference') as scope:
if reuse:
scope.reuse_variables()
with slim.arg_scope([slim.conv2d], padding='SAME',
weights_initializer=initializers.xavier_initializer(),
weights_regularizer=None,
biases_initializer=init_ops.zeros_initializer(),
biases_regularizer=None,
normalizer_fn=slim.batch_norm,
activation_fn=tf.nn.relu):
net = slim.conv2d(x, DEEP_SIZE, [2, 2], stride=1, scope='conv1')
net = slim.conv2d(net, DEEP_SIZE*2, [2, 2], stride=1, scope='conv2')
net = slim.conv2d(net, DEEP_SIZE, [2, 2], stride=1, scope='conv3')
# net = slim.conv2d(net, DEEP_SIZE*2, [2, 2], stride=1, scope='conv4')
# net = slim.conv2d(net, DEEP_SIZE, [2, 2], stride=1, scope='conv5')
net = slim.conv2d(net, IMAGE_CHANNEL, [2, 2], stride=1, activation_fn=None, scope='conv6')
net = tf.reshape(net, [-1, 1])
a = tf.ones_like(net) * 0.5
net = tf.concat([net, a], 1)
net = tf.arg_max(net, 1)
net = tf.cast(net, tf.float32)
net = tf.reshape(net, [batch_size, IMAGE_SIZE1, IMAGE_SIZE2, IMAGE_CHANNEL])
return net
出现错误“没有为任何变量提供梯度，请检查您的图表中不支持梯度的操作，变量之间”。
解决方案
正如错误已经告诉您的那样，有一个操作 (op) 导致 tensorflow 无法计算梯度。在您的情况下，操作tf.arg_max是问题所在。该函数不可微分（即不存在导数，因此无法计算梯度）。因此，Tensorflow 不能为图中的任何变量创建梯度并抛出此错误。
解决方案是设计一个只使用可微函数的网络，并摆脱tf.arg_max. 由于您没有解释您的网络设计意图是什么，因此很难说出inference在这种情况下如何重新设计您的功能。我只能猜测您更希望softmax获得每个类的预测概率。然后可以将其用于训练的损失函数中。
即使您想要的目标图像只有值 0 和 1，强制网络只输出 0 和 1 也是不合理的。相反，您需要一个可以解释为（未缩放的）概率的连续输出。您需要连续值来计算定义明确的导数，以便您的损失函数可以通过梯度下降进行优化。
当然，在训练阶段之后，您可以argmax将预测值 <0.5 的所有像素设置为零，其他像素设置为 1。但是，在训练期间不应该这样做。
推荐阅读
- reactjs - 如何单击使用本机反应的功能呈现的按钮？
- git - 主分支有时在开发分支之前
- reactjs - 实现 css datepicker 样式在表单内无法正常工作
- spring-boot - 数组作为spring boot kotlin中蚂蚁匹配器的参数
- python - 调整 RGB 张量的大小
- c# - 检查一个类是否实现了一个精确的泛型类定义
- javascript - JS 表单事件和验证
- css - Chrome 中的 Flexbox 高度损坏
- java - 如何使用 Apache MINA SSHD 阻止 SFTP 删除操作
- bash - 计算 git 存储库给定分支中所有文件的所有行的出现总数