使用类先验优化输入图像

我正在尝试实现谷歌博客条目的第一部分起始论：深入探讨神经网络 https://research.googleblog.com/2015/06/inceptionism-going-deeper-into-neural.html在 TensorFlow 中。到目前为止，我发现了一些资源可以解释它自然语言 https://stackoverflow.com/questions/32856606/google-inceptionism-obtain-images-by-class或专注于其他部分或给出其他框架的代码片段 https://stackoverflow.com/questions/30994563/google-deep-dream-art-how-to-pick-a-layer-in-a-neural-network-and-enhance-it。我理解相对于先前的类优化随机输入图像的想法以及此中给出的其背后的数学论文，第 2 节 https://arxiv.org/pdf/1312.6034v2.pdf，但我无法自己使用 TensorFlow 实现它。

From 这个问题 https://stackoverflow.com/questions/39856291/how-to-visualize-dnns-dependent-of-the-output-class-in-tensorflow/39858096?noredirect=1#comment67705644_39858096以及有用的评论etarion，我现在知道您可以向优化器提供变量列表，而所有其他变量均保持不变。然而，当给优化器一个随机图像作为变量时会导致

File "mnist_test.py", line 101, in main
    optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(-cost, var_list=[rnd_img])
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 198, in minimize
  name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/optimizer.py", line 309, in apply_gradients
  (converted_grads_and_vars,))
ValueError: No gradients provided for any variable: ((None,<tensorflow.python.ops.variables.Variable object at 0x7feac1870410>),)

出于测试目的，我使用了一个精简的 MNIST 示例。我尝试使其尽可能短，同时仍具有可读性和可执行性：

def main():
    # parameters
    learning_rate = 0.001
    train_batches = 1000
    batch_size = 128
    display_step = 50

    # net parameters
    n_input = 784 #28x28
    n_classes = 10
    keep_prob = 0.75

    weights = {
        'wc1': tf.Variable(tf.truncated_normal([5, 5, 1, 32])),
        'wc2': tf.Variable(tf.truncated_normal([5, 5, 32, 64])),
        'wd1': tf.Variable(tf.truncated_normal([7*7*64, 1024])),
        'out': tf.Variable(tf.truncated_normal([1024, n_classes]))
    }

    biases = {
        'bc1': tf.Variable(tf.constant(0.1, shape=[32])),
        'bc2': tf.Variable(tf.constant(0.1, shape=[64])),
        'bd1': tf.Variable(tf.constant(0.1, shape=[1024])),
        'out': tf.Variable(tf.constant(0.1, shape=[n_classes]))
    }

    # tf inputs
    x = tf.placeholder(tf.float32, [None, n_input])
    y = tf.placeholder(tf.float32, [None, n_classes])
    dropout = tf.placeholder(tf.float32)

    # create net
    net = create_net(x, weights, biases, keep_prob)

    # define loss
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, y))

    # define optimizer
    optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)

    # evaluation
    pred_correct = tf.equal(tf.argmax(net, 1), tf.argmax(y, 1))
    accuracy = tf.reduce_mean(tf.cast(pred_correct, tf.float32))

    print "loading mnist data"
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

    sess = tf.Session()
    sess.run(tf.initialize_all_variables())

    for i in xrange(train_batches):
        batch_x, batch_y = mnist.train.next_batch(batch_size)

        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y, dropout: keep_prob})

        if i % display_step == 0:
            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x, y: batch_y, dropout: 1.0})
            print "batch: %i, loss: %.5f, accuracy: %.5f" % (i, loss, acc)

    acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels, dropout: 1.0})
    print "test accuracy: %.5f" % (acc)

    # ====== this is where the reconstruction begins =====

    rnd_img = tf.Variable(tf.random_normal([1, n_input]))
    one_hot = np.zeros(10)
    one_hot[4] = 1;

    # the next line causes the error
    optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(-cost, var_list=[rnd_img])

    for i in xrange(1000):
        session.run(optimizer2, feed_dict={x: rnd_img, y: one_hot, dropout: 1.0})

    sess.close()

if __name__ == "__main__":
    main()

我使用的辅助函数：

def create_net(x, weights, biases, dropout):
    x = tf.reshape(x, shape=[-1, 28, 28, 1])

    conv1 = conv2d_relu(x, weights['wc1'], biases['bc1'])
    conv1 = maxpool2d(conv1, 2)

    conv2 = conv2d_relu(conv1, weights['wc2'], biases['bc2'])
    conv2 = maxpool2d(conv2, 2)

    fc1 = fullyconnected_relu(conv2, weights['wd1'], biases['bd1'])
    fc1 = tf.nn.dropout(fc1, dropout)

    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])

    return out

def conv2d_relu(x, W, b, stride=1):
    conv = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding='SAME')
    conv = tf.nn.bias_add(conv, b)
    return tf.nn.relu(conv)

def maxpool2d(x, k=2, stride=2, padding='VALID'):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)

def fullyconnected_relu(x, W, b):
    fc = tf.reshape(x, [-1, W.get_shape().as_list()[0]])
    fc = tf.add(tf.matmul(fc, W), b)
    fc = tf.nn.relu(fc)

我发现一些消息来源说，当输出和要优化的变量之间的计算图中没有路径时，就会发生此错误，但我不明白为什么会出现这种情况。

我的问题是：

1. 为什么优化器无法应用任何梯度？
2. 这是实现类可视化的正确方法吗？

提前致谢。

Edit:

在合并已接受的答案之后，这里再次是完整的代码（对于任何感兴趣的人）。无论如何，结果仍然不符合预期，因为脚本在 100000 轮重建后基本上产生了随机图像。欢迎提出想法。

import tensorflow as tf
import numpy as np
import skimage.io

def conv2d_relu(x, W, b, stride=1):
    conv = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding='SAME')
    conv = tf.nn.bias_add(conv, b)
    return tf.nn.relu(conv)

def maxpool2d(x, k=2, stride=2, padding='VALID'):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)

def fullyconnected_relu(x, W, b):
    fc = tf.reshape(x, [-1, W.get_shape().as_list()[0]])
    fc = tf.add(tf.matmul(fc, W), b)
    fc = tf.nn.relu(fc)

    return fc;

def create_net(x, weights, biases, dropout):
    x = tf.reshape(x, shape=[-1, 28, 28, 1])

    conv1 = conv2d_relu(x, weights['wc1'], biases['bc1'])
    conv1 = maxpool2d(conv1, 2)

    conv2 = conv2d_relu(conv1, weights['wc2'], biases['bc2'])
    conv2 = maxpool2d(conv2, 2)

    fc1 = fullyconnected_relu(conv2, weights['wd1'], biases['bd1'])
    fc1 = tf.nn.dropout(fc1, dropout)

    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])

    return out

def save_image(img_data, name):
    img = img_data.reshape(28,28)

    mi = np.min(img)
    ma = np.max(img)

    img = (img-mi)/(ma-mi)
    skimage.io.imsave(name, img)

def main():
    # parameters
    learning_rate = 0.001
    train_batches = 1000
    batch_size = 100
    display_step = 50

    # net parameters
    n_input = 784 #28x28
    n_classes = 10
    keep_prob = 0.75

    weights = {
        'wc1': tf.Variable(tf.truncated_normal([5, 5, 1, 32])),
        'wc2': tf.Variable(tf.truncated_normal([5, 5, 32, 64])),
        'wd1': tf.Variable(tf.truncated_normal([7*7*64, 1024])),
        'out': tf.Variable(tf.truncated_normal([1024, n_classes]))
    }

    biases = {
        'bc1': tf.Variable(tf.constant(0.1, shape=[32])),
        'bc2': tf.Variable(tf.constant(0.1, shape=[64])),
        'bd1': tf.Variable(tf.constant(0.1, shape=[1024])),
        'out': tf.Variable(tf.constant(0.1, shape=[n_classes]))
    }

    # tf inputs
    x = tf.placeholder(tf.float32, [None, n_input])
    y = tf.placeholder(tf.float32, [None, n_classes])
    dropout = tf.placeholder(tf.float32)

    # create net
    net = create_net(x, weights, biases, dropout)

    # define loss
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net, y))

    # define optimizer
    optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost)

    # evaluation
    pred_correct = tf.equal(tf.argmax(net, 1), tf.argmax(y, 1))
    accuracy = tf.reduce_mean(tf.cast(pred_correct, tf.float32))

    print "loading mnist data"
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

    sess = tf.Session()

    sess.run(tf.initialize_all_variables())

    for i in xrange(train_batches):
        batch_x, batch_y = mnist.train.next_batch(batch_size)

        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y, dropout: keep_prob})

        if i % display_step == 0:
            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x, y: batch_y, dropout: 1.0})
            print "batch: %i, loss: %.5f, accuracy: %.5f" % (i, loss, acc)

    acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels, dropout: 1.0})
    print "test accuracy: %.5f" % (acc)

    # reconstruction part
    rnd_img = tf.Variable(tf.random_normal([1, n_input]))

    one_hot = np.zeros((1, 10))
    one_hot[0,1] = 1;

    net2 = create_net(rnd_img, weights, biases, dropout)
    cost2 = -tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(net2, y))
    optimizer2 = tf.train.AdamOptimizer(learning_rate).minimize(cost2, var_list=[rnd_img])

    init_var_list = []

    for var in tf.all_variables():
        if(not tf.is_variable_initialized(var).eval(session=sess)):
            init_var_list.append(var)

    sess.run(tf.initialize_variables(init_var_list))

    save_image(rnd_img.eval(sess), "bevor.tiff")

    for i in xrange(100000):
        _, loss = sess.run([optimizer2, cost2], feed_dict={y: one_hot, dropout: 1.0})
        if(i%10000 == 0):
            cur_img = rnd_img.eval(session=sess)
            print "loss:", loss, "mi:", np.min(cur_img), "ma:", np.max(cur_img)

    save_image(rnd_img.eval(sess), "after.tiff")

    sess.close()

if __name__ == "__main__":
    main()

一些解释：使用新的输入变量和优化器重建图形后，我必须初始化新变量，即 rnd_img 和 Adam 优化器使用的一些辅助变量，因此循环 all_variables() 并检查初始化状态。如果有人知道更优雅的方式，请告诉我。或者也许这就是我没有得到任何结果的原因？

rnd_img 需要是您优化的图形的一部分。在您的情况下，您只需创建一个变量并告诉优化器对其进行优化，但该变量与图中的损失无关。例如，您可以做的是使用 rnd_image 而不是 x 使用另一个对 create_net 的调用（但使用相同的权重！），为此创建成本，然后为该成本创建一个最小化操作。然后为了优化你只需要输入 y 。