Fix the problem that Torch does not correspond to Torchvision version

[Lornatang]

fix #148 errors

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Adjusting PyTorch and torchvision to specific versions for consistency.

📊 Key Changes

• Changed PyTorch requirement from any version newer than 1.4 (>=1.4) to the specific version 1.4 (==1.4).
• Pinned torchvision to an exact version 0.5.0 (==0.5.0) instead of the latest available version.

🎯 Purpose & Impact

• Ensures compatibility and predictability by using specific versions of PyTorch and torchvision, reducing the likelihood of bugs or incompatibilities with new releases.
• Users can expect a more stable environment with less risk of encountering issues due to changing dependencies.
• May impact users who work with newer features of PyTorch and torchvision that are not available in the specified versions. 🔄

[glenn-jocher]

@Lornatang we can't force torch 1.4 install, as colab for example comes with 1.5 and the docker containers come with 1.5 also etc. 1.5, and now 1.5.1 have many bug fixes and improvements over 1.4.

[Lornatang]

@glenn-jocher This error seems to have found a good solution. If you use conda to install torch+torchvision you will get an error. No error when using pip.