生成Grad-CAM热力图时遇'Tensor'无'output'属性错误求助

错误原因 #

• 代码中model.get_layer('vgg16').get_layer(last_conv_layer_name).output已返回Tensor对象（层的输出张量），而非层本身。后续调用last_conv_layer.output时，Tensor没有output属性，直接触发报错。
• 代码缺失get_img_array函数定义，会导致后续运行报错。
• 存在keras与tensorflow.keras混用的情况，可能引发版本冲突。

修复步骤及完整代码 #

1. 修正卷积层模型构建逻辑 #

获取最后卷积层的层对象而非输出张量，再通过层对象调用output属性。

2. 补充缺失的get_img_array函数 #

3. 统一使用tensorflow.keras导入，避免版本冲突 #

以下是修正后的完整代码：

import numpy as np
import tensorflow as tf
from tensorflow import keras
from IPython.display import Image, display
import matplotlib.pyplot as plt
import matplotlib.cm as cm
from tensorflow.keras.preprocessing.image import load_img, img_to_array
import os
from tensorflow.keras import models

# 统一tensorflow.keras导入，避免版本冲突
model_builder = keras.applications.xception.Xception
img_size = (299, 299)
preprocess_input = keras.applications.xception.preprocess_input
decode_predictions = keras.applications.xception.decode_predictions

img_path = "box.jpg"

def get_img_array(img_path, size):
    # 将图像转换为模型输入的数组格式
    img = load_img(img_path, target_size=size)
    array = img_to_array(img)
    # 添加batch维度
    array = np.expand_dims(array, axis=0)
    return array

def make_gradcam_heatmap(
    img_array, model, last_conv_layer_name, classifier_layer_names
):
    # 获取最后卷积层的层对象（而非输出张量）
    last_conv_layer = model.get_layer('vgg16').get_layer(last_conv_layer_name)
    # 构建输入到最后卷积层输出的模型
    last_conv_layer_model = keras.Model(model.inputs, last_conv_layer.output)

    # 构建分类器部分的模型（从最后卷积层输出到最终预测）
    classifier_input = keras.Input(shape=last_conv_layer.output.shape[1:])
    x = classifier_input
    for layer_name in classifier_layer_names:
        x = model.get_layer(layer_name)(x)
    classifier_model = keras.Model(classifier_input, x)

    # 计算梯度（Grad-CAM核心步骤）
    with tf.GradientTape() as tape:
        last_conv_layer_output = last_conv_layer_model(img_array)
        tape.watch(last_conv_layer_output)
        preds = classifier_model(last_conv_layer_output)
        # 获取预测概率最高的类别索引
        top_pred_index = tf.argmax(preds[0])
        top_class_channel = preds[:, top_pred_index]

    # 计算分类结果对最后卷积层输出的梯度
    grads = tape.gradient(top_class_channel, last_conv_layer_output)
    # 对每个通道的梯度取全局平均
    pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2))

    # 计算热力图：卷积层输出与池化后的梯度加权求和
    last_conv_layer_output = last_conv_layer_output[0]
    heatmap = last_conv_layer_output @ pooled_grads[..., tf.newaxis]
    heatmap = tf.squeeze(heatmap)
    # 归一化到0-1范围
    heatmap = tf.maximum(heatmap, 0) / tf.math.reduce_max(heatmap)
    return heatmap.numpy()

# 准备图像
img_array = preprocess_input(get_img_array(img_path, size=img_size))

# 加载模型
model = models.load_model("Box_Model_Augmented(15.11).h5")

last_conv_layer_name = "block5_conv3"
classifier_layer_names = ["global_max_pooling2d", "predictions"]

# 生成热力图
heatmap = make_gradcam_heatmap(
    img_array, model, last_conv_layer_name, classifier_layer_names
)

# 可选：将热力图叠加到原始图像上显示
def display_gradcam(img_path, heatmap, alpha=0.4):
    # 加载原始图像
    img = load_img(img_path)
    img = img_to_array(img)

    # 将热力图缩放到与原始图像相同尺寸
    heatmap = np.uint8(255 * heatmap)
    jet = cm.get_cmap("jet")
    jet_colors = jet(np.arange(256))[:, :3]
    jet_heatmap = jet_colors[heatmap]

    # 将热力图转换为图像格式
    jet_heatmap = keras.preprocessing.image.array_to_img(jet_heatmap)
    jet_heatmap = jet_heatmap.resize((img.shape[1], img.shape[0]))
    jet_heatmap = img_to_array(jet_heatmap)

    # 叠加热力图与原始图像
    superimposed_img = jet_heatmap * alpha + img
    superimposed_img = keras.preprocessing.image.array_to_img(superimposed_img)

    # 显示结果
    plt.figure(figsize=(10, 10))
    plt.imshow(superimposed_img)
    plt.axis('off')
    plt.show()

display_gradcam(img_path, heatmap)

额外说明 #

• 确保你的模型中确实包含名为vgg16的子模型，且该子模型存在block5_conv3层；如果是自定义模型，需核对层名称是否正确。
• 若使用Xception模型而非VGG16，需修改last_conv_layer_name为Xception的最后卷积层名称（如block14_sepconv2），同时移除model.get_layer('vgg16')调用。

内容的提问来源于stack exchange，提问作者SumSum