TensorFlow中DeepDream代码报错：无法转换符号tf.Tensor为NumPy数组

错误信息 #

NotImplementedError: Cannot convert a symbolic tf.Tensor (Mean:0) to a numpy array.
This error may indicate that you're trying to pass a Tensor to a NumPy call, which is not supported.

运行代码 #

import tensorflow as tf 
import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt
import random

base_model = tf.keras.applications.InceptionV3(include_top=False, weights='imagenet')
base_model.summary()

names = ['mixed3','mixed5']
layers = [base_model.get_layer(name).output for name in names]

deepdream_model = tf.keras.Model(inputs=base_model.input, outputs=layers)

Sample_Image = tf.keras.preprocessing.image.load_img(r'apple.jpg', target_size =(225, 375))

np.shape(Sample_Image)


Sample_Image = np.array(Sample_Image)/255.0
Sample_Image.shape

plt.imshow(Sample_Image)

Sample_Image.max()
Sample_Image.min()

Sample_Image = tf.keras.preprocessing.image.img_to_array(Sample_Image) 
Sample_Image.shape

Sample_Image = tf.Variable(tf.keras.applications.inception_v3.preprocess_input(Sample_Image))

Sample_Image = tf.expand_dims(Sample_Image, axis = 0)
np.shape(Sample_Image)

activations = deepdream_model(Sample_Image)



def calc_loss(image, model):

    img_batch = tf.expand_dims(image, axis=0)
    layer_activations = model(img_batch)
    print('VALORES DE ACTIVACION (LAYER OUTPUT) =\n', layer_activations)

    losses = []
    for act in layer_activations:
        loss = tf.math.reduce_mean(act)
        losses.append(loss)

    print('PERDIDAS (DE MULTIPLES CAPAS DE ACTIVACION) = ', losses)
    print('FORMA DE PERDIDA (DE MULTIPLES CAPAS DE ACTIVACION) =', np.shape(losses))
    print('SUMA DE TODAS LAS CAPAS PERDIDAS (DE TODAS LAS CAPAS SELECCIONADAS) =', tf.reduce_sum(losses))

    return tf.reduce_sum(losses)


Sample_Image = tf.keras.preprocessing.image.load_img(r'apple.jpg', target_size =(225, 375))
Sample_Image = np.array(Sample_Image)/255.0
Sample_Image = tf.keras.preprocessing.image.img_to_array(Sample_Image)
Sample_Image = tf.Variable(tf.keras.applications.inception_v3.preprocess_input(Sample_Image))

loss = calc_loss(Sample_Image, deepdream_model)

loss

@tf.function
def deepdream(model, image, step_size):
    with tf.GradientTape() as tape:
        tape.watch(image)
        loss = calc_loss(image, model)

    gradients = tape.gradient(loss, image)

    print('GRADIENTES = \n', gradients)
    print('FORMA DE GRADIENTES =\n', np.shape(gradients))

    gradients /= tf.math.reduce_std(gradients)

    image = image + gradients * step_size
    image = tf.clip_by_value(image, -1, 1)

    return loss, image

def run_deep_dream_simple(model, image, steps=100, step_size=0.01):
    image = tf.keras.applications.inception_v3.preprocess_input(image)

    for step in range(steps):
        loss, image = deepdream(model, image, step_size)
    
        if step % 100 == 0:
            plt.figure(figsize=(12, 12))
            plt.imshow(deprocess(image))
            plt.show()
            print("Step {}, loss{}".format(step, loss))
    
    plt.figure(figsize=(12, 12))
    plt.imshow(deprocess(image))
    plt.show()

    return deprocess(image)

def deprocess(image):
    image = 255 * (image + 1.0) / 2.0
    return tf.cast(image, tf.uint8)

    Sample_Image= tf.keras.preprocessing.image.load_img(r'apple.jpg', target_size = (225, 375))
    Sample_Image = np.array(Sample_Image)
    dream_img = run_deep_dream_simple(model=deepdream_model, image=Sample_Image, steps=2000, step_size=0.001)

问题分析与修复方案 #

核心错误原因 #

报错根源是在@tf.function装饰的图模式函数中，调用了numpy的API处理符号张量：

• calc_loss里的np.shape(losses)试图将符号张量列表转换为numpy数组，图模式下不支持这种操作
• deepdream函数里的np.shape(gradients)存在同样问题
• 最后一段运行代码错误缩进在deprocess函数内部，导致逻辑无法正常执行

修复步骤 #

1. 替换numpy操作为TensorFlow原生方法

把所有np.shape()替换为TensorFlow的tf.shape()或张量自身的.shape属性，同时将普通print改为tf.print适配图模式：

# 修改calc_loss中的打印逻辑
def calc_loss(image, model):
    img_batch = tf.expand_dims(image, axis=0)
    layer_activations = model(img_batch)
    tf.print('VALORES DE ACTIVACION (LAYER OUTPUT) =\n', layer_activations)

    losses = [tf.math.reduce_mean(act) for act in layer_activations]

    tf.print('PERDIDAS (DE MULTIPLES CAPAS DE ACTIVACION) = ', losses)
    tf.print('FORMA DE PERDIDA (DE MULTIPLES CAPAS DE ACTIVACION) =', [tf.shape(l) for l in losses])
    total_loss = tf.reduce_sum(losses)
    tf.print('SUMA DE TODAS LAS CAPAS PERDIDAS (DE TODAS LAS CAPAS SELECCIONADAS) =', total_loss)

    return total_loss

# 修改deepdream中的打印逻辑
@tf.function
def deepdream(model, image, step_size):
    with tf.GradientTape() as tape:
        tape.watch(image)
        loss = calc_loss(image, model)

    gradients = tape.gradient(loss, image)

    tf.print('GRADIENTES = \n', gradients)
    tf.print('FORMA DE GRADIENTES =\n', tf.shape(gradients))

    gradients /= tf.math.reduce_std(gradients)

    image = image + gradients * step_size
    image = tf.clip_by_value(image, -1, 1)

    return loss, image

2. 修复代码缩进错误

将最后一段加载图片并运行DeepDream的代码移到deprocess函数外部：

def deprocess(image):
    image = 255 * (image + 1.0) / 2.0
    return tf.cast(image, tf.uint8)

# 移到函数外部，保证正常执行
Sample_Image= tf.keras.preprocessing.image.load_img(r'apple.jpg', target_size = (225, 375))
Sample_Image = np.array(Sample_Image)
dream_img = run_deep_dream_simple(model=deepdream_model, image=Sample_Image, steps=2000, step_size=0.001)

3. （可选）移除调试打印

如果不需要调试信息，可以直接删除所有打印语句，进一步避免图模式下的张量交互问题。

内容的提问来源于stack exchange，提问作者ROMERO VISOSO