MoveNet姿态估计模型推理时出现维度不匹配错误求解决方案

我用TensorFlow官方姿态分类Notebook训练并通过model.save('model.keras')保存了MoveNet姿态估计模型，编写推理代码加载模型、预处理图像后执行预测时出现维度不匹配错误。

推理代码：

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

# Load the model
filepath = "C:\\Users\\golut\\OneDrive\\Documents\\PoseEstimation\\model.keras"
new_model = tf.keras.models.load_model(filepath)

# Read and preprocess the image
image_path = "C:\\Users\\golut\\OneDrive\\Documents\\PoseEstimation\\test_pos.jpg"
image = tf.io.read_file(image_path)
image = tf.image.decode_image(image, channels=3)
image = tf.image.convert_image_dtype(image, tf.float32)
image = tf.image.resize(image, (256, 256))  # Resize if necessary
image = np.expand_dims(image, axis=0)  # Add batch dimension

# Perform inference
predictions = new_model.predict(image)

# Process predictions as needed
# For example, print the predictions
print(predictions)

# Optionally, visualize the image and predictions
plt.imshow(image[0])  # Assuming only one image in the batch
plt.show()

错误提示：

Traceback (most recent call last):
  File "c:\Users\golut\OneDrive\Documents\PoseEstimation\app.py", line 18, in <module>
    predictions = new_model.predict(image)
  File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\utils\traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "C:\Users\golut\AppData\Local\Temp\__autograph_generated_filepea7q91y.py", line 15, in tf__predict_function
    retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), None, fscope)
ValueError: in user code:

    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\engine\training.py", line 2440, in predict_function  *
        return step_function(self, iterator)
    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\engine\training.py", line 2425, in step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\engine\training.py", line 2413, in run_step  **
        outputs = model.predict_step(data)
    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\engine\training.py", line 2381, in predict_step
        return self(x, training=False)
    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\utils\traceback_utils.py", line 70, in error_handler
        raise e.with_traceback(filtered_tb) from None
    File "C:\ProgramData\anaconda3\envs\pose_est\lib\site-packages\keras\src\engine\input_spec.py", line 298, in assert_input_compatibility
        raise ValueError(

    ValueError: Input 0 of layer "model" is incompatible with the layer: expected shape=(None, 51), found shape=(None, 256, 256, 3)

问题根源 #

你保存的不是MoveNet姿态估计模型，而是Notebook里最后训练的姿态分类模型。这个分类模型的输入不是原始图像，而是MoveNet输出的17个关键点的坐标+置信度（每个关键点包含x/y/置信度三个值，17×3=51维），所以模型期望输入形状是(None,51)，但你传入的是(None,256,256,3)的图像，直接导致维度不匹配。

解决办法 #

情况1：你需要的是姿态分类（判断动作类别，比如坐下/站立）

你需要先通过MoveNet模型提取图像的姿态关键点，再把关键点数据输入到你保存的分类模型中做预测：

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

# 加载你的姿态分类模型
classifier_model = tf.keras.models.load_model("C:\\Users\\golut\\OneDrive\\Documents\\PoseEstimation\\model.keras")

# 加载预训练的MoveNet姿态估计模型（以Lightning版本为例）
movenet = tf.keras.models.load_model("movenet_singlepose_lightning")

def preprocess_image(image_path, input_size=192):
    # MoveNet Lightning版本要求输入尺寸为192×192
    image = tf.io.read_file(image_path)
    image = tf.image.decode_jpeg(image, channels=3)
    image = tf.image.resize_with_pad(image, input_size, input_size)
    image = tf.cast(image, dtype=tf.int32)
    return tf.expand_dims(image, axis=0)

def extract_keypoints(movenet_model, image):
    outputs = movenet_model.predict(image)
    # 提取关键点坐标和置信度，展平为51维向量
    keypoints = outputs['output_0'].numpy().flatten()[:51]
    return np.expand_dims(keypoints, axis=0)

# 处理测试图像
image_path = "C:\\Users\\golut\\OneDrive\\Documents\\PoseEstimation\\test_pos.jpg"
input_image = preprocess_image(image_path)
keypoints = extract_keypoints(movenet, input_image)

# 用分类模型执行预测
predictions = classifier_model.predict(keypoints)
print("分类预测结果：", predictions)

# 可视化图像
plt.imshow(tf.squeeze(input_image).numpy().astype(np.uint8))
plt.show()

情况2：你需要的是姿态估计（提取人体关键点）

如果你想保存的是MoveNet姿态估计模型，说明你在Notebook里保存错了对象。回到训练Notebook，找到加载MoveNet的代码，直接保存该模型即可，示例代码：

# 加载MoveNet后保存模型
movenet = tf.keras.models.load_model("movenet_singlepose_lightning")
movenet.save("movenet_model.keras")

后续推理时，直接传入符合MoveNet输入尺寸（比如192×192）的预处理图像即可。

内容的提问来源于stack exchange，提问作者Maaz Shahid