160 points by ml_researcher 1 year ago flag hide 13 comments
hnuser1 1 year ago next
[Opening Comment] TensorFlow vs PyTorch: Performance Analysis on Large-Scale Datasets is an important topic in the Machine Learning community. I'm excited to see the results.
tfadvocate 1 year ago next
TensorFlow has a more mature ecosystem and offers features such as TensorBoard, which helps create visualizations. This leads to a better overall experience while monitoring.
tfadvocate 1 year ago next
Sure, PyTorch is easy to learn, but TensorFlow's performance is more consistent during inference, especially in large-scale production environments, making it a better choice.
pytorchfan 1 year ago prev next
While TensorFlow is strong with tooling and visualization, PyTorch provides an easier to learn and use API without compromising performance. PyTorch has great community support too.
newtoml 1 year ago next
Does anyone have concrete performance data that supports either TensorFlow or PyTorch on large-scale datasets like ImageNet?
performancedataguru 1 year ago next
Yes, I did an analysis with ImageNet recently, and results show that TensorFlow has a model loading time & inference edge over PyTorch, but PyTorch handles complex models better with comparable speed.
perfdatainterested 1 year ago next
Do you have a link to your analysis? I'd be curious to read what results you obtained.
performancedataguru 1 year ago next
Please find the analysis at [URL] . I tried to include the most relevant information I could find.
hnuser2 1 year ago prev next
What we really need is an industry standard for comparing performance between ML frameworks. It's difficult to get good benchmarks.
standardizedbenchmarker 1 year ago next
There are already several projects aiming to do just that. MLPerf (<https://mlperf.org/>) is one of the forerunners.
hnuser3 1 year ago prev next
TensorFlow was created by the research team behind DeepMind at Google, while PyTorch's creators are at Facebook Research. I wonder if this duality of origins implies the difference between choices.
originsagreed 1 year ago next
I agree with the difference in origin. Perhaps, TensorFlow focuses on optimizing for large-scale and heavily parallelized calculation whereas PyTorch targets ease of use and research.
divideandconquer 1 year ago prev next
Both have their advantages and disadvantages, but a researcher or developer should choose based on their particular needs.