ResNet50生成Grad-CAM热图遇sequential_2未调用错误求助

问题背景 #

使用Keras训练了皮肤病变分类的ResNet50模型，保存为ResNet50_skin_lesion_final_v2.keras，模型结构包含ResNet50基础层、全局平均池化层、全连接层、Dropout层及8分类输出层。尝试基于ResNet50的conv5_block3_out层生成Grad-CAM热图时，出现以下错误：

AttributeError: The layer sequential_2 has never been called and thus has no defined input.

相关代码 #

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow.keras.preprocessing import image
from tensorflow.keras.models import load_model

# Load trained model (ResNet50_skin_lesion_final_v2)
model = load_model('ResNet50_skin_lesion_final_v2.keras')

# Function to preprocess the input image
def preprocess_input_image(img_path, target_size=(224, 224)):
    img = image.load_img(img_path, target_size=target_size)
    img_array = image.img_to_array(img)
    img_array = np.expand_dims(img_array, axis=0)
    img_array /= 255.0  # Rescale the image like the training data
    return img_array

# Function to generate Grad-CAM heatmap
def generate_gradcam_heatmap(model, img_array, last_conv_layer_name, pred_index=None):
    # Forward pass to ensure the model input is called
    predictions = model.predict(img_array)
    if pred_index is None:
        pred_index = tf.argmax(predictions[0])

    # Get the ResNet50 base model
    base_model = model.get_layer('resnet50')

    # Build a model that outputs the activations of the last conv layer and the model output
    grad_model = tf.keras.models.Model(
        inputs=model.input, 
        outputs=[base_model.get_layer(last_conv_layer_name).output, model.output]
    )

    with tf.GradientTape() as tape:
        # Watch the gradients for the convolutional layer
        conv_outputs, predictions = grad_model(img_array)
        class_output = predictions[:, pred_index]

    # Compute the gradient of the class output with respect to the feature map
    grads = tape.gradient(class_output, conv_outputs)

    # Take the mean of the gradients across the channels
    pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))

    # Multiply the output feature map by the pooled gradients
    conv_outputs = conv_outputs[0]
    conv_outputs *= pooled_grads

    # Generate the heatmap
    heatmap = tf.reduce_mean(conv_outputs, axis=-1)
    heatmap = np.maximum(heatmap, 0) / np.max(heatmap)  # Normalize between 0 and 1
    return heatmap.numpy()

# Function to display the Grad-CAM heatmap on top of the original image
def display_gradcam(img_path, heatmap, alpha=0.4):
    # Load the original image
    img = image.load_img(img_path)
    img = image.img_to_array(img)

    # Resize the heatmap to match the original image size
    heatmap = np.uint8(255 * heatmap)
    heatmap = np.expand_dims(heatmap, axis=-1)
    heatmap = tf.image.resize(heatmap, (img.shape[0], img.shape[1])).numpy()

    # Create an RGB heatmap
    heatmap = np.uint8(plt.cm.jet(heatmap[..., 0]) * 255)

    # Superimpose the heatmap on the original image
    superimposed_img = heatmap * alpha + img

    # Display the image
    plt.figure(figsize=(8, 8))
    plt.imshow(superimposed_img.astype('uint8'))
    plt.axis('off')
    plt.show()

# Example usage
img_path = '/mnt/c/Users/arjay/Downloads/Ubuntu/ISIC_2019_Skin_Lesion_Data/ISIC_2019_Training_Input/manual_test/ISIC_0027665.jpg'
img_array = preprocess_input_image(img_path)

# Generate the Grad-CAM heatmap from the last conv layer of ResNet50
heatmap = generate_gradcam_heatmap(model, img_array, last_conv_layer_name='conv5_block3_out')

# Display the Grad-CAM heatmap on the original image
display_gradcam(img_path, heatmap)

错误栈 #

AttributeError                            Traceback (most recent call last)
Cell In[128], line 82
79 img_array = preprocess_input_image(img_path)
81 # Generate the Grad-CAM heatmap from the last conv layer of ResNet50
---> 82 heatmap = generate_gradcam_heatmap(model, img_array, last_conv_layer_name='conv5_block3_out')
84 # Display the Grad-CAM heatmap on the original image
85 display_gradcam(img_path, heatmap)

Cell In[128], line 30
26 base_model = model.get_layer('resnet50')
28 # Build a model that outputs the activations of the last conv layer and the model output
29 grad_model = tf.keras.models.Model(
---> 30     inputs=model.input,
31     outputs=[base_model.get_layer(last_conv_layer_name).output, model.output]
32 )
34 with tf.GradientTape() as tape:
35     # Watch the gradients for the convolutional layer
36     conv_outputs, predictions = grad_model(img_array)

File ~/projects/tf217/tf217/lib/python3.12/site-packages/keras/src/ops/operation.py:254, in Operation.input(self)
244 @property
245 def input(self):
246     """Retrieves the input tensor(s) of a symbolic operation.
247
...
292         f"{node_index}, but the operation has only "
293         f"{len(self._inbound_nodes)} inbound nodes."
294     )

AttributeError: The layer sequential_2 has never been called and thus has no defined input.

疑问解答 #

1. 为何错误提及sequential_2，而我加载的是ResNet50模型？ #

你的整体模型是一个Sequential容器（自动命名为sequential_2），其中封装了ResNet50基础层和后续的分类层（全局平均池化、全连接等）。加载模型后，直接访问model.input时，由于Sequential层的输入节点尚未通过实际输入数据完成初始化（即使调用了predict，部分情况下内部层的符号输入张量仍未被正确注册），导致报错。本质是模型加载后，容器层的输入张量没有被显式定义，构建grad_model时无法直接引用model.input。

2. 如何正确输入数据以生成Grad-CAM热图？ #

核心是确保构建grad_model前，模型的输入输出节点已完全初始化。以下是两种可靠的修正方案：

方案一：显式创建输入张量替代model.input

在构建grad_model时，用tf.keras.Input根据模型的输入形状创建输入张量，避免依赖未初始化的model.input。

方案二：加载模型后先执行一次前向传播

在调用generate_gradcam_heatmap前，先让模型处理一次输入数据（比如model(img_array)），确保所有层的输入节点被初始化。

修正后的完整代码 #

方案一：显式定义输入张量 #

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow.keras.preprocessing import image
from tensorflow.keras.models import load_model

# Load trained model
model = load_model('ResNet50_skin_lesion_final_v2.keras')

def preprocess_input_image(img_path, target_size=(224, 224)):
    img = image.load_img(img_path, target_size=target_size)
    img_array = image.img_to_array(img)
    img_array = np.expand_dims(img_array, axis=0)
    img_array /= 255.0
    return img_array

def generate_gradcam_heatmap(model, img_array, last_conv_layer_name, pred_index=None):
    # 先执行前向传播获取预测结果
    predictions = model.predict(img_array, verbose=0)
    if pred_index is None:
        pred_index = tf.argmax(predictions[0])

    base_model = model.get_layer('resnet50')
    
    # 显式创建输入张量，替代model.input
    input_tensor = tf.keras.Input(shape=model.input_shape[1:])
    x = input_tensor
    conv_output = None
    # 重建模型的前向传播路径，记录目标卷积层输出
    for layer in model.layers:
        x = layer(x)
        if layer.name == last_conv_layer_name:
            conv_output = x
    final_output = x

    # 构建grad_model
    grad_model = tf.keras.models.Model(
        inputs=input_tensor,
        outputs=[conv_output, final_output]
    )

    with tf.GradientTape() as tape:
        conv_outputs, predictions = grad_model(img_array)
        class_output = predictions[:, pred_index]

    grads = tape.gradient(class_output, conv_outputs)
    pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))

    conv_outputs = conv_outputs[0]
    conv_outputs *= pooled_grads

    heatmap = tf.reduce_mean(conv_outputs, axis=-1)
    heatmap = np.maximum(heatmap, 0) / np.max(heatmap)
    return heatmap.numpy()

def display_gradcam(img_path, heatmap, alpha=0.4):
    img = image.load_img(img_path)
    img = image.img_to_array(img)

    heatmap = np.uint8(255 * heatmap)
    heatmap = np.expand_dims(heatmap, axis=-1)
    heatmap = tf.image.resize(heatmap, (img.shape[0], img.shape[1])).numpy()

    heatmap = np.uint8(plt.cm.jet(heatmap[..., 0]) * 255)
    superimposed_img = heatmap * alpha + img

    plt.figure(figsize=(8, 8))
    plt.imshow(superimposed_img.astype('uint8'))
    plt.axis('off')
    plt.show()

# 示例调用
img_path = '/mnt/c/Users/arjay/Downloads/Ubuntu/ISIC_2019_Skin_Lesion_Data/ISIC_2019_Training_Input/manual_test/ISIC_0027665.jpg'
img_array = preprocess_input_image(img_path)

heatmap = generate_gradcam_heatmap(model, img_array, last_conv_layer_name='conv5_block3_out')
display_gradcam(img_path, heatmap)

方案二：提前初始化模型输入节点 #

修改generate_gradcam_heatmap函数，增加显式前向传播步骤：

def generate_gradcam_heatmap(model, img_array, last_conv_layer_name, pred_index=None):
    # 显式执行一次前向传播，初始化模型输入节点
    _ = model(img_array)
    predictions = model.predict(img_array, verbose=0)
    if pred_index is None:
        pred_index = tf.argmax(predictions[0])

    base_model = model.get_layer('resnet50')

    grad_model = tf.keras.models.Model(
        inputs=model.input, 
        outputs=[base_model.get_layer(last_conv_layer_name).output, model.output]
    )

    # 后续代码与原函数一致...

内容的提问来源于stack exchange，提问作者Arjay Alba