TensorFlow 2.10.0导出ONNX遇Conda包冲突问题求助

环境创建流程 #

我通过以下步骤创建Conda环境并完成MNIST模型训练：

conda create --name tf210 python=3.8
conda activate tf210
conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0
pip install tensorflow-gpu==2.10.0 keras==2.10.0 
pip install opencv-python==4.4.0.44 pillow==8.2 numpy==1.22 matplotlib scipy pandas scikit-learn tqdm imutils PyYAML tensorboard seaborn protobuf==3.20 chardet

pip install tensorflow-datasets==4.6.0
conda install ipykernel
python -m ipykernel install --user --name=tf210 --display-name="tf210"

conda install -c conda-forge widgetsnbextension ipywidgets

报错详情 #

安装tf2onnx和onnxruntime后，代码运行出现如下错误：

Traceback (most recent call last):
  File "F:\pythonProject\main.py", line 3, in <module>
    import tensorflow_datasets as tfds
  File "D:\Anaconda\envs\tf210\lib\site-packages\tensorflow_datasets\__init__.py", line 52, in <module>
    from tensorflow_datasets import image
  File "D:\Anaconda\envs\tf210\lib\site-packages\tensorflow_datasets\image\__init__.py", line 43, in <module>
    from tensorflow_datasets.image.dsprites import Dsprites
  File "D:\Anaconda\envs\tf210\lib\site-packages\tensorflow_datasets\image\dsprites.py", line 23, in <module>
    import h5py
  File "D:\Anaconda\envs\tf210\lib\site-packages\h5py\__init__.py", line 26, in <module>
    from . import _errors
ImportError: DLL load failed while importing _errors:The specified program cannot be found.

已知条件与需求 #

• protobuf 3.20.x与tensorflow-gpu 2.10.0、tensorboard 2.14.0适配正常
• 需保留tensorflow-gpu 2.10.0，解决当前报错并完成ONNX模型导出

解决方案 #

1. 修复h5py依赖问题 #

报错根源是pip安装的h5py与Conda环境系统库不兼容，导致DLL缺失。执行以下步骤替换h5py版本：

# 卸载当前pip安装的h5py
pip uninstall -y h5py

# 用conda安装适配Conda环境的h5py版本（选择与Python3.8、TF2.10兼容的3.7.0版本）
conda install -c conda-forge h5py=3.7.0

2. 安装适配TF2.10的tf2onnx与onnxruntime #

选择与TensorFlow 2.10.0兼容的版本，避免版本冲突：

# 安装适配TF2.10的tf2onnx
pip install tf2onnx==1.14.0

# 若需GPU加速，安装onnxruntime-gpu；否则安装onnxruntime
pip install onnxruntime-gpu==1.14.0
# CPU版本替代命令：pip install onnxruntime==1.14.0

3. 验证与测试 #

• 先验证tensorflow_datasets导入是否正常：

import tensorflow_datasets as tfds
print("tfds imported successfully")

• 再测试模型导出ONNX的流程，示例代码：

import tensorflow as tf
from tf2onnx import convert

# 加载已训练的MNIST模型
model = tf.keras.models.load_model("mnist_model.h5")

# 转换为ONNX格式
spec = (tf.TensorSpec((None, 28, 28, 1), tf.float32, name="input"),)
output_path = "mnist_model.onnx"
convert.from_keras(model, input_signature=spec, output_path=output_path)
print(f"Model exported to {output_path}")

内容的提问来源于stack exchange，提问作者Felix Ye