PyTorch初始化SummaryWriter遇AttributeError: tensorflow无io属性

问题描述 #

使用PyTorch的torch.utils.tensorboard.SummaryWriter训练DCGAN时，代码未直接导入或使用TensorFlow，但初始化SummaryWriter时触发如下报错：

Traceback (most recent call last):
  File "C:/Users/Gebruiker/PycharmProjects/BobRossGAN/train.py", line 47, in <module>
    writer_real = SummaryWriter(f"logs/real")
  File "C:\Users\Gebruiker\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\utils\tensorboard\writer.py", line 247, in __init__
    self._get_file_writer()
  File "C:\Users\Gebruiker\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\utils\tensorboard\writer.py", line 277, in _get_file_writer
    self.file_writer = FileWriter(
  File "C:\Users\Gebruiker\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\utils\tensorboard\writer.py", line 76, in __init__
    self.event_writer = EventFileWriter(
  File "C:\Users\Gebruiker\AppData\Local\Programs\Python\Python38\lib\site-packages\tensorboard\summary\writer\event_file_writer.py", line 73, in __init__
    if not tf.io.gfile.exists(logdir):
  File "C:\Users\Gebruiker\AppData\Local\Programs\Python\Python38\lib\site-packages\tensorboard\lazy.py", line 65, in __getattr__
    return getattr(load_once(self), attr_name)
AttributeError: module 'tensorflow' has no attribute 'io'

涉事代码片段 #

import torch
from torch.utils.tensorboard import SummaryWriter

# other imports and code

writer_real = SummaryWriter(f"logs/real")

当前环境版本：torch==2.2.0、tensorboard==2.13.0，已尝试降级相关库但问题未解决。

解决方案 #

1. 升级TensorFlow至兼容版本 #

报错核心是TensorBoard 2.13.0依赖的TensorFlow版本过低，缺少tf.io模块。安装与TensorBoard匹配的TensorFlow版本：

pip install tensorflow==2.13.0 --upgrade

2. 卸载TensorFlow，使用独立版TensorBoard #

若无需TensorFlow，可卸载现有TensorFlow，安装不依赖完整TensorFlow的独立版TensorBoard：

pip uninstall tensorflow -y
pip install tensorboard==2.13.0 --no-deps

3. 清理缓存并重装依赖 #

若环境存在损坏的安装缓存，执行以下命令彻底清理后重装：

pip cache purge
pip uninstall torch tensorboard tensorflow -y
pip install torch==2.2.0 tensorboard==2.13.0

GPU环境需确保安装对应CUDA版本的PyTorch，避免使用CPU版本。

4. 手动创建日志目录绕过TensorFlow API #

提前创建日志目录，避免TensorBoard调用TensorFlow的文件检查逻辑：

import os
from torch.utils.tensorboard import SummaryWriter

log_dir = "logs/real"
os.makedirs(log_dir, exist_ok=True)
writer_real = SummaryWriter(log_dir)

内容的提问来源于stack exchange，提问作者Djanger