TensorFlow/Keras model.save失败疑因非英文字母路径求解决

问题描述 #

训练手写数字识别神经网络时，模型训练环节正常，但调用model.save()保存模型时报错，终端错误信息如下：

Traceback (most recent call last):
File "c:\Users\Käyttäjä\Desktop\Sampo\Neural\handwritten.py", line 33, in
model.save('handwritten.model.test1')
File "C:\Users\Käyttäjä\AppData\Local\Programs\Python\Python311\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\Käyttäjä\AppData\Local\Programs\Python\Python311\Lib\site-packages\tensorflow\python\lib\io\file_io.py", line 513, in recursive_create_dir_v2
_pywrap_file_io.RecursivelyCreateDir(compat.path_to_bytes(path))
tensorflow.python.framework.errors_impl.FailedPreconditionError: handwritten.model.test1\variables is not a directory

怀疑是用户名Käyttäjä中的特殊字符Ä导致路径编码问题，且无法修改现有目录，寻求可行解决办法。

使用的代码如下：

import os
import cv2
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf


# Load the dataset
mnist = tf.keras.datasets.mnist
(x_train, y_train), (x_test, y_test) = mnist.load_data()

# Use tf to normalize datasets between 0-1
x_train = tf.keras.utils.normalize(x_train, axis=1)
x_test = tf.keras.utils.normalize(x_test, axis=1)

# Create a basic sequential model. Also define the shape of input to be flattened 28 x 28 because of dataset beeing a 28 x 28 greyscale image.
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Flatten(input_shape=(28, 28)))
# Second, Dense, layer is a layer where every neuron is connected to every neuron. relu= rectified linear unit is the activation function (negative values turn to 0)
model.add(tf.keras.layers.Dense(128, activation='relu'))
# thrid layer, exactly same
model.add(tf.keras.layers.Dense(128, activation='relu'))
# final layer. 10 outputs for numbers 0-9. Softmax activation function is often used for the output. Gives a probability distribution as the output
model.add(tf.keras.layers.Dense(10, activation='softmax'))

# compile model using optimizer algorithm and loss function
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# finally train the model. Model is going to see each datapoint 3 times
model.fit(x_train, y_train, epochs=3)

# after the model has been trained, save the model as handwritten.model
model.save('handwritten.model.test1')

可行解决办法 #

1. 保存到不含特殊字符的绝对路径 #

选择路径中无特殊字符的目录（如系统临时目录、非用户目录的磁盘分区），使用绝对路径保存模型：

# 示例：保存到D盘指定目录
model.save(r'D:\saved_models\handwritten.model.test1')

# 或者保存到系统临时目录
import tempfile
temp_dir = tempfile.gettempdir()
model.save(os.path.join(temp_dir, 'handwritten.model.test1'))

2. 保存为HDF5格式（单个文件） #

HDF5格式将模型保存为单个文件，无需创建子目录结构，可直接绕过路径编码问题：

# 保存为.h5格式文件
model.save('handwritten_model.h5')

# 后续加载模型使用：
# model = tf.keras.models.load_model('handwritten_model.h5')

3. 强制TensorFlow使用UTF-8编码处理路径 #

在代码开头添加环境变量设置，强制TensorFlow用UTF-8解析文件路径：

import os
os.environ['TF_IO_ENCODING'] = 'UTF-8'
os.environ['PYTHONUTF8'] = '1'

# 之后再导入tensorflow
import tensorflow as tf
# 其余代码保持不变

验证建议 #

优先尝试方案2（HDF5格式），操作简单且直接解决问题；若需保留SavedModel格式，可选择方案1或方案3。

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