当启用急切执行时，传递给 Optimizer.compute_gradients 的 `loss` 应该是一个函数

Question

我是 TensorFlow 的新手，我刚刚开始学习和理解它。我正在neural style transfer解决问题，我正在使用 tensorflow version 1.14。

我收到一个错误损失传递给 Optimizer.compute_gradients should be a function when Eage Execution is enabled。

我试图通过使用 TensorFlow 图而不是 来解决这个问题eager execution，但它不起作用。我想使用，eager execution因为它看起来更像是 Pythonic 的方式。

这是我的代码，很抱歉将整个代码放在这里，请在我的代码中提出更正建议。

import scipy
import tensorflow as tf
import tensorflow.contrib.eager as tfe
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
from scipy import misc
from skimage.transform import resize
from tensorflow.keras.applications.vgg19 import VGG19, preprocess_input
from tensorflow.keras import backend as K

tf.enable_eager_execution()
print('Eager execution {}'.format(tf.executing_eagerly()))

content_path = '800px-Green_Sea_Turtle_grazing_seagrass.jpg'
style_path = '800px-The_Great_Wave_off_Kanagawa.jpg'

content_img = plt.imread(content_path)
plt.imshow(content_img)
style_img = plt.imread(style_path)
plt.imshow(style_img)

MEANS = np.array([123.68, 116.779, 103.939]).reshape((1,1,1,3))
content_img = resize(content_img, (552,800,3)) #resized content img because style img has shape (552,800,3)

content_img = np.array(content_img)
content_img = np.reshape(content_img, ((1,)+content_img.shape))
style_img = np.array(style_img)
style_img = np.reshape(style_img, ((1,)+style_img.shape))

noise_img= np.random.uniform(-20,20,(1,552,800,3)).astype('float32')
generated_img = noise_img*0.6 + content_img*0.4
plt.imshow(generated_img[0])

content_img = content_img-MEANS
style_img = style_img-MEANS

model = VGG19(include_top=False, weights='imagenet')

def compute_content_cost(act_content_img, act_generated_img):
    return tf.reduce_mean(tf.square(act_content_img-act_generated_img))

def gram_matrix(A):
    gram = tf.matmul(A, tf.transpose(A))
    return gram

def style_loss_one_layer(act_style_img, act_generated_img):
    m,n_H,n_W,n_C = tf.shape(act_generated_img)               #act_generated_img.get_shape().as_list()
    gram_act_style_img = gram_matrix(act_style_img)
    gram_generated_img = gram_matrix(act_generated_img)
    return tf.reduce_mean(tf.square(gram_act_style_img-gram_generated_img))*(1/(4*n_C**2*(n_H*n_W)**2))

content_layer = ['block5_conv2']
style_layers = [('block1_conv1',0.2), 
                ('block2_conv1',0.2),
                ('block3_conv1',0.2),
                ('block4_conv1',0.2),
                ('block5_conv1',0.2)]

def compute_style_cost(model, style_layers):
    style_cost = total_style_cost = 0
    for layer, coeff in style_layers:
        act_style_img = model.get_layer(layer).output
        act_generated_img = model.get_layer(layer).output
        style_cost += style_loss_one_layer(act_style_img, act_generated_img)
        total_style_cost += coeff*style_cost 
    return total_style_cost

def compute_total_cost(J_content, J_style, alpha=10, beta=40):
    J = (alpha*tf.cast(J_content, tf.float64)) + (beta*J_style)
    return J

act_generated_img = model.get_layer('block5_conv2').output
act_content_img = model.get_layer('block5_conv2').output

J_content = compute_content_cost(act_content_img=act_content_img, act_generated_img=act_generated_img)
print(J_content)
J_style = compute_style_cost(model, style_layers=style_layers)
print(J_style)

J_total_cost = compute_total_cost(J_content, J_style, alpha=10, beta=40)
print(J_total_cost)

optimizer = tf.train.AdamOptimizer(2.0)

train_step = optimizer.minimize(J_total_cost)        #**getting error here**

Answer 1

上述错误主要是在您尝试使用 TensorFlow 1.x 但系统运行的是 tensor 2.0 时引起的。

使用以下代码初始化 TensorFlow，以确保您尝试使用 1.0 版

将 tensorflow.compat.v1 导入为 tf

您可以在初始化程序之后通过以下命令使系统禁用该行为。

tf.disable_v2_behavior()