GradientTape 在尝试计算梯度时返回一个空对象

Question

我想实现 Gatys 论文中讨论的方法一种艺术风格的神经算法。该方法需要一个自定义损失函数，虽然我没有在 TensorFlow 的文档中搜索它，但我还是想自己实现它。问题是 GradientTape 不计算损失函数的梯度。 GradientTape 似乎不监视需要监视的变量。我只是不能安静地把手指放在它上面。这是我的代码：

import numpy as np
import tensorflow as tf
from matplotlib import pyplot as plt


def gram(x):
    shape = x.shape
    x = np.reshape(x, (shape[1] * shape[2], shape[3]))
    return tf.matmul(x, x, True)


def load_image(path):
    image = tf.io.read_file(path)
    image = tf.image.decode_image(image, 3, tf.float32)
    assert image.shape[0] == image.shape[1]
    image = tf.image.resize(image, (224, 224))
    return image[tf.newaxis, :]


def content_loss(F, P):
    return tf.reduce_sum(((F - P) ** 2) / 2)


def style_loss(F, A):
    shape = F.shape
    G = gram(F)
    H = gram(A)
    return tf.reduce_sum(((G - H) ** 2) / (4 * shape[1] * shape[2] * shape[3] * 4))


def step(layer, activation):
    return layer(activation)


def feed(model, image, content_layers, style_layers, activations={}):
    activation = image
    for layer in model.keys():
        activation = step(model[layer], activation)
        if layer in content_layers or layer in style_layers:
            activations[layer] = activation
    return activations


def total_loss(layers, activations, content, style, content_layers, style_layers):
    loss = tf.Variable(0, trainable=True,  dtype=tf.float32)
    for layer in layers:
        if layer in content_layers:
            loss.assign_add(content_loss(activations[layer], content[layer]))
        if layer in style_layers:
            loss.assign_add(style_loss(activations[layer], style[layer]) * 1e03)
    return loss


content_layers = {'block5_pool'}
style_layers = {'block5_conv1', 'block5_conv2', 'block5_conv3', 'block5_conv4'}

VGG19 = tf.keras.applications.vgg19.VGG19(
    include_top=False, weights='imagenet', input_shape=(224, 224, 3))
vgg = {}
content = {}
style = {}
con_activation = load_image('./Data/pokion.png')
stl_activation = load_image('./Data/quadpon.jpg')
for layer in VGG19.layers:
    layer.trainable = False
    con_activation = step(layer, con_activation)
    stl_activation = step(layer, stl_activation)
    if layer.name in content_layers:
        content[layer.name] = con_activation
    if layer.name in style_layers:
        style[layer.name] = stl_activation
    vgg[layer.name] = layer
assert (len(content) != 0 and len(style) != 0)

del VGG19, con_activation, stl_activation

generated_image = tf.Variable(np.random.rand(1, 224, 224, 3), trainable=True, dtype=tf.float32)
optimizer = tf.optimizers.Adam()
for i in range(200):
    with tf.GradientTape(True) as tape:
        gen = feed(vgg, generated_image, content_layers, style_layers)
        print(generated_image.shape)
        loss = total_loss(vgg.keys(), gen, content, style, content_layers, style_layers)
        gradients = tape.gradient(loss, generated_image)
        optimizer.apply_gradients(zip([gradients], [generated_image]))
        fig, ax = plt.subplots()
        ax.imshow(generated_image[0])
        plt.show()

输出显示如下：

WARNING: Logging before flag parsing goes to stderr.
W0316 17:30:45.648816  6320 backprop.py:980] Calling GradientTape.gradient on a persistent tape inside its context is significantly less efficient than calling it outside the context (it causes the gradient ops to be recorded on the tape, leading to increased CPU and memory usage). Only call GradientTape.gradient inside the context if you actually want to trace the gradient in order to compute higher order derivatives.
Traceback (most recent call last):
  File "C:/Users/amran/Desktop/Projects/Roastilng Jabrils/tests/NST.py", line 85, in <module>
    gradients = tape.gradient(loss, generated_image)
  File "C:\ProgramData\Anaconda3\lib\site-packages\tensorflow_core\python\eager\backprop.py", line 990, in gradient
    with self:
  File "C:\ProgramData\Anaconda3\lib\site-packages\tensorflow_core\python\eager\backprop.py", line 803, in __enter__
    self._push_tape()
  File "C:\ProgramData\Anaconda3\lib\site-packages\tensorflow_core\python\eager\backprop.py", line 813, in _push_tape
    raise ValueError("Tape is already recording.")
ValueError: Tape is already recording.

有什么我错过的吗？

Answer 1

GradientTape 实例应该在“with”块结束后使用，这对于 Python 中上下文管理器的使用方式有点违反直觉。基本上你在“with”块中完成了所有需要观察的操作，然后使用磁带来获取相关的渐变。在块的内部尝试以下操作：

with tf.GradientTape(True) as tape:
    gen = feed(vgg, generated_image, content_layers, style_layers)
    print(generated_image.shape)
    loss = total_loss(vgg.keys(), gen, content, style, content_layers, style_layers)
gradients = tape.gradient(loss, generated_image)
optimizer.apply_gradients(zip([gradients], [generated_image]))
fig, ax = plt.subplots()
ax.imshow(generated_image[0])
plt.show()