如何解决Docker中pip与conda的opencv-python包冲突问题

我基于PyTorch镜像构建Docker镜像，Dockerfile中通过pip install -r requirements.txt安装包含opencv-python==4.7.0.68在内的所有Python依赖，但执行时出现以下错误：

我的Dockerfile片段 #

# PyTorch image
...

RUN pip install -r requirements.txt

报错信息 #

55.09 Installing collected packages: voluptuous, tensorboard-plugin-wit, sentencepiece, library, easygui, tensorboard-data-server, safetensors, regex, pyasn1-modules, protobuf, opencv-python, oauthlib, multidict, markdown, lightning-utilities, grpcio, ftfy, frozenlist, entrypoints, cachetools, async-timeout, absl-py, yarl, requests-oauthlib, huggingface-hub, google-auth, aiosignal, torchmetrics, tokenizers, google-auth-oauthlib, diffusers, aiohttp, accelerate, transformers, timm, tensorboard, altair, pytorch-lightning, open-clip-torch
56.38     Found existing installation: protobuf 3.20.3
56.39     Uninstalling protobuf-3.20.3:
58.77       Successfully uninstalled protobuf-3.20.3
58.87   Attempting uninstall: opencv-python
58.87     Found existing installation: opencv-python 4.6.0
58.87 ERROR: Cannot uninstall opencv-python 4.6.0, RECORD file not found. Hint: The package was installed by conda.

可行解决方案 #

• 方案1：用pip跳过已安装的opencv-python
  修改Dockerfile中的pip命令，添加--ignore-installed opencv-python参数，让pip跳过卸载已有的conda安装版本，直接安装指定版本：

  RUN pip install --ignore-installed opencv-python -r requirements.txt
  

  注意：仅针对opencv-python跳过，避免全局忽略引发其他依赖冲突。

• 方案2：先通过conda更新opencv-python
  如果基础镜像支持conda，先在Dockerfile中用conda更新到目标版本，再执行pip安装其他依赖：

  RUN conda install -y opencv-python=4.7.0.68
  RUN pip install -r requirements.txt
  

  这样pip不会再尝试更新opencv-python，彻底避免卸载冲突。

• 方案3：移除requirements.txt中的opencv-python，改用conda安装
  删除requirements.txt里的opencv-python==4.7.0.68条目，在Dockerfile中单独添加conda安装命令：

  RUN conda install -y opencv-python=4.7.0.68
  RUN pip install -r requirements.txt
  

  这种方式完全隔离pip和conda的包管理操作，杜绝冲突隐患。

内容的提问来源于stack exchange，提问作者ln9187