"implement __dir__ for Tensors, so that editors can automatically auto-complete and query for the possible fields in Tensors"
This will probably be the most underrated part of this update. Used to bug me quite a bit.
From my experience, Pytorch has the best developper experience among deep learning frameworks. It seems this release will make it even better. Thanks and bravo to all contributors!
After tensorflow added support for eager evaluation, I don't understand why I might want to use pytorch. Can somebody give any ideas?
Because as with most things Tensorflow, support for eager mode is sparse and inconsistent. PyTorch "just works".
FWIW at this point Gluon/mxnet seems to be a nicer solution where you get a clean interactive API but also strong symbolic support (w/ better performance than TF).
Because tensorflow is pretty horrible to write.
for those wondering: Pytorch == Tensors and Dynamic neural networks in Python with strong GPU acceleration for Computer Vision, Text, Reinforcement Learning, etc.