Comment by cweill | Hacker News Reader

Comment by cweill | Hacker News Reader

I worked in Google Research for over 5 years doing Machine Learning, and recently quit to build my own ML start-up. These days, I solve a mix of NLP, computer vision, and tabular problems, all with state of the art neural network techniques. I've tried many setups.

My advice is go with Colab Pro ($50/mo) and TensorFlow/Keras. You can go with Pytorch too if you prefer.

I made the mistake of buying a 2080Ti for my desktop thinking it would be better, but no. Consumer grade hardware is nowhere near as good/fast as the server grade hardware you get in Colab. Plus you have the option to use TPUs in Colab if you want to scale up quickly.

You really don't need to get fancy with this setup. The best part of using Colab is you can work on your laptop from anywhere, and never worry about your ML model hogging all your RAM (and swap) or compute and slowing your local machine down. Trust me, this sucks when it happens, and you have to restart!

As for your data, you can host it in a GCS bucket. For small data (<1TB) even better is Google drive (I know, crazy). Colab can mount your Google drive and loads from it extremely quickly. It's like having a remote filesystem, except with a handy UI and collaboration options, and an easy way to inspect and edit your data.

(OP, please don’t subject yourself to TensorFlow/Keras. The moment Jax became available on TPUs publicly, the moment I stopped using TF. And boy oh boy, “never looked back” is an understatement. I still cringe remembering all the time I spent trying to get tf.function to just please, please work, like a housewife alarmed that neither her partner nor herself are able to actually work.)

To support your point with data, here is a graph of usage of TF vs PyTorch in papers over time: https://horace.io/pytorch-vs-tensorflow/

I've seen this chart too. But research != Production. I bet that TensorFlow is still more commonly used in serving than Pytorch because of great tooling like TensorFlow Serving. I could be wrong though, as I'm not up to date with the latest in the Pytorch ecosystem.

FWIW PyTorch has TorchServe nowadays which does the same thing as TF Serving.

Great advice! BTW, your startup https://creatorml.com/ is very cool, what a creative idea.

Thank you! Feel free to reach out to me on Twitter or Discord (linked on the homepage) if you want to chat.

Note that Colab Pro is $10/month and Pro+ is $50/month.

The $10 is more than enough for learning Deep Learning.

100% agreed. Start with the $10 plan. I forgot how much it cost for the middle tier. The one benefit of the higher tier is you get access to better GPUs, and can run multiple colabs in parallel effectively getting multiple accelerators at once if you're doing distributed hyperparameter tuning.

The one benefit of the higher tier is you get access to better GPUs

I think it’s the same GPUs as Pro. I’m actually surprised you don’t recommend buying a 2080Ti. The best GPU you get with colab pro+ is P100 which is slower than 2080Ti. If you can afford it, having your own GPU workstation is a much better experience than dealing with colab.

+1 and you can even connect your Colab to a GCP Marketplace Colab runtime that has no time limit (but will cost us) if you e.g. need to run something for a few days (although then you don't get the awesome Google drive mounting - hope they fix this eventually)

Reply on news.ycombinator.com