导出Teachable Machine模型后本地加载报错：DepthwiseConv2D参数异常

报错信息 #

TypeError: Error when deserializing class 'DepthwiseConv2D' using config={'name': 'expanded_conv_depthwise', 'trainable': True, 'dtype': 'float32', 'kernel_size': [3, 3], 'strides': [1, 1], 'padding': 'same', 'data_format': 'channels_last', 'dilation_rate': [1, 1], 'groups': 1, 'activation': 'linear', 'use_bias': False, 'bias_initializer': {'class_name': 'Zeros', 'config': {}}, 'bias_regularizer': None, 'activity_regularizer': None, 'bias_constraint': None, 'depth_multiplier': 1, 'depthwise_initializer': {'class_name': 'VarianceScaling', 'config': {'scale': 1, 'mode': 'fan_avg', 'distribution': 'uniform', 'seed': None}}, 'depthwise_regularizer': None, 'depthwise_constraint': None}.

Exception encountered: Unrecognized keyword arguments passed to DepthwiseConv2D: {'groups': 1}

本地部署代码 #

from keras.models import load_model  # TensorFlow is required for Keras to work
from PIL import Image, ImageOps  # Install pillow instead of PIL
import numpy as np

# Disable scientific notation for clarity
np.set_printoptions(suppress=True)

# Load the model
model = load_model("keras_model.h5", compile=False)

# Load the labels
class_names = open("labels.txt", "r").readlines()

# Create the array of the right shape to feed into the keras model
# The 'length' or number of images you can put into the array is
# determined by the first position in the shape tuple, in this case 1
data = np.ndarray(shape=(1, 224, 224, 3), dtype=np.float32)

# Replace this with the path to your image
image = Image.open("jap span.jfif").convert("RGB")

# resizing the image to be at least 224x224 and then cropping from the center
size = (224, 224)
image = ImageOps.fit(image, size, Image.Resampling.LANCZOS)

# turn the image into a numpy array
image_array = np.asarray(image)

# Normalize the image
normalized_image_array = (image_array.astype(np.float32) / 127.5) - 1

# Load the image into the array
data[0] = normalized_image_array

# Predicts the model
prediction = model.predict(data)
index = np.argmax(prediction)
class_name = class_names[index]
confidence_score = prediction[0][index]

# Print prediction and confidence score
print("Class:", class_name[2:], end="")
print("Confidence Score:", confidence_score)

问题原因 #

Teachable Machine云端使用的是较新版本的TensorFlow，导出的模型配置中给DepthwiseConv2D层添加了groups参数，但本地环境安装的TensorFlow/Keras版本较低，该版本的DepthwiseConv2D不支持groups参数，导致反序列化模型时报错。

解决方案 #

方案1：升级TensorFlow版本 #

直接升级本地的TensorFlow到兼容版本（建议2.10及以上），执行以下命令：

pip install --upgrade tensorflow

升级完成后重新运行代码即可。

方案2：自定义兼容层（无需升级版本） #

如果无法升级环境，可以自定义一个忽略groups参数的DepthwiseConv2D层，加载模型时替换原层：
修改模型加载部分的代码为以下内容：

from keras.models import load_model
from keras.layers import DepthwiseConv2D
from keras.utils import get_custom_objects
import numpy as np
from PIL import Image, ImageOps

# 自定义兼容的DepthwiseConv2D层，移除不识别的groups参数
class CustomDepthwiseConv2D(DepthwiseConv2D):
    def __init__(self, *args, **kwargs):
        kwargs.pop('groups', None)
        super().__init__(*args, **kwargs)

# 注册自定义层
get_custom_objects()['DepthwiseConv2D'] = CustomDepthwiseConv2D

# Disable scientific notation for clarity
np.set_printoptions(suppress=True)

# Load the model
model = load_model("keras_model.h5", compile=False)

# 后续代码保持不变...

这样加载模型时会自动忽略groups参数，不影响模型的预测功能。

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