ImageAI加载模型报错：变量形状与值形状不兼容

我使用imageai.Detection.Custom.CustomObjectDetection类加载通过imageai.Detection.Custom.DetectionModelTrainer训练的模型，对测试图像进行目标标注。已按官方文档配置好包含所有标签的数据集，但运行检测代码时出现维度不兼容的错误，想确认问题出在检测代码、训练代码还是数据集？

错误堆栈 #

Traceback (most recent call last):
  File "./model_predict.py", line 25, in <module>
    detector.loadModel()
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/imageai/Detection/Custom/__init__.py", line 666, in loadModel
    self.__model.load_weights(self.__model_path)
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/tensorflow/python/keras/engine/training.py", line 2234, in load_weights
    hdf5_format.load_weights_from_hdf5_group(f, self.layers)
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/tensorflow/python/keras/saving/hdf5_format.py", line 710, in load_weights_from_hdf5_group
    K.batch_set_value(weight_value_tuples)
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/tensorflow/python/util/dispatch.py", line 201, in wrapper
    return target(*args, **kwargs)
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/tensorflow/python/keras/backend.py", line 3706, in batch_set_value
    x.assign(np.asarray(value, dtype=dtype(x)))
  File "/home/thecoder3281f/Documents/cca/env/lib/python3.7/site-packages/tensorflow/python/ops/resource_variable_ops.py", line 891, in assign
    (tensor_name, self._shape, value_tensor.shape))
ValueError: Cannot assign to variable last1/kernel:0 due to variable shape (1, 1, 1024, 15) and value shape (45, 1024, 1, 1) are incompatible

环境信息 #

• OS: Ubuntu 22.04
• Python版本: 3.7.15
• GPU: NVIDIA Geforce RTX 2060
• CPU: AMD Ryzen 5 3600
• 文件夹结构符合官方文档规范

训练代码 #

import os
from imageai.Detection.Custom import DetectionModelTrainer

os.environ["SM_FRAMEWORK"] = "tf.keras"
trainer = DetectionModelTrainer()
trainer.setModelTypeAsYOLOv3()
trainer.setDataDirectory(data_directory="custom_image_db")
trainer.setTrainConfig(object_names_array=[], batch_size=4, num_experiments=20, train_from_pretrained_model="pretrained-yolov3.h5")
trainer.trainModel()

检测代码 #

import os

execution_path = os.getcwd()                                       
                                         
from imageai.Detection.Custom import CustomObjectDetection

detector = CustomObjectDetection()
detector.setModelTypeAsYOLOv3()
detector.setModelPath("custom_image_db/models (1)/detection_model-ex-020--loss-0030.816.h5")
detector.setJsonPath("custom_image_db/json (1)/detection_config.json")
detector.loadModel()
detections = detector.detectObjectsFromImage(input_image="testimg1.jpg", output_image_path="testimg1-detected.jpg", extract_detected_objects=True)
for detection in detections:
    print(detection["name"], " : ", detection["percentage_probability"], " : ", detection["box_points"])

样本XML标注文件 #

<?xml version="1.0" encoding="utf-8"?>
<annotation>
<folder>images</folder>
<filename>af278bd0-AdobeStock_15555240_Preview.jpeg</filename>
<source>
<database>MyDatabase</database>
<annotation>COCO2017</annotation>
<image>flickr</image>
<flickrid>NULL</flickrid>
<annotator>1</annotator>
</source>
<owner>
<flickrid>NULL</flickrid>
<name>Label Studio</name>
</owner>
<size>
<width>1000</width>
<height>667</height>
<depth>3</depth>
</size>
<segmented>0</segmented>
<object>
<name>Fall</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>35</xmin>
<ymin>100</ymin>
<xmax>921</xmax>
<ymax>603</ymax>
</bndbox>
</object>
<object>
<name>Pain</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>557</xmin>
<ymin>62</ymin>
<xmax>923</xmax>
<ymax>314</ymax>
</bndbox>
</object>
</annotation>

问题出在训练代码，核心错误是训练时未正确指定数据集的类别列表：

• 训练代码中trainer.setTrainConfig()的object_names_array参数设为空列表[]，但你的数据集明确包含Fall和Pain两个类别。
• YOLOv3的输出层维度由类别数量决定（计算公式：3*(类别数+5)，其中5对应坐标和置信度参数）。训练时未指定类别列表会导致模型按0类别构建输出层，与实际数据集的类别数不匹配，最终生成的模型权重维度混乱，检测加载时出现形状不兼容的错误。

解决步骤 #

1. 修改训练代码，将object_names_array设置为与数据集匹配的类别列表：

   trainer.setTrainConfig(object_names_array=["Fall", "Pain"], batch_size=4, num_experiments=20, train_from_pretrained_model="pretrained-yolov3.h5")
   

2. 删除之前训练生成的models和json文件夹，重新训练模型。
3. 使用新训练生成的模型文件和对应的detection_config.json运行检测代码。

额外检查点 #

• 确认所有标注文件中的类别名称与object_names_array里的内容完全一致，大小写、拼写不能出错。
• 检测代码加载的模型文件和json文件必须来自同一次训练，不能混用不同训练任务的文件。

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