loading...

What's your local dev setup for deep learning with PyTorch?

marcusturewicz profile image Marcus Turewicz ・1 min read

For development of deep learning models with PyTorch e.g. image classification, object detection and so forth, what works best for your local development machine:

  • Which operating system?
  • Which GPU?
  • Anaconda, pip or nothing?
  • Anything else to consider?

Would love to hear what you're using :)

Discussion

markdown guide
 

Not using pytorch but currently:

  • TF2
  • Windows 10 Professional with WSL 2
  • Dell XPS 13 with RTX 2070 Super connected via TB3
  • Vanilla python with venv

For quick and dirty solutions, I like to use google colab!

 

Thanks for letting me know!

I am keen to try out WSL 2 with deep learning. Do you see any major performance slow downs compared to native Linux?

 

I haven't tried any particular benchmarking, but I would expect a small performance dip due to the tb3 bandwidth limit

 

I really prefer pytorch over tensorflow

  • I have Arch, it makes it super easy to install packages ( especially nvidia)
  • NVIDIA gtx 970 , but it only has 4gb of ram, so depending on the task I use CPU...
  • Used conda before ( when only a specific python version was supported by TF), now I use pip, I like it more
  • Yes, buy and SSD, It makes loading data extremely fast. I have nvme, but sata should work a lot too!
  • Another thing, sometimes I like using AWS, but only when have tested the training locally
 

Thanks for sharing!

Arch has always seemed very mysterious to me, but I will definitely give it a go!

I have found conda to be quite bloated and cumbersome, especially on Windows. Pip is very straightforward.

 

Yes, pip is very simple compared to conda. I thought the same about Arch, but the installation was super easy thanks to the documentation it has. I used fedora before, but it always caused me problems when installing nvidia, it was a pain. Moreover I love the fact that Arch has a rolling release.