Google seem to be particularly bad at this. They did the same with Tensorflow. It was kind of the de facto deep learning framework until Google decided to deprecate everything. Everyone responded by switching to Torch instead.
IMO tensorflow always had a worse API than PyTorch. There’s even the legendary issue “I fucking hate tensorflow”, now unfortunately censored as spam lol.
Tensorflow died because devs never bothered improving that. While PyTorch always had an increasing number of features and high level capabilities, TF has always felt like a lower level tool that only made sense choosing if you needed to run models for inference in other platforms. PyTorch Lightning on top of it was a great touch for researchers.
Google seem to be particularly bad at this. They did the same with Tensorflow. It was kind of the de facto deep learning framework until Google decided to deprecate everything. Everyone responded by switching to Torch instead.
IMO tensorflow always had a worse API than PyTorch. There’s even the legendary issue “I fucking hate tensorflow”, now unfortunately censored as spam lol.
Tensorflow died because devs never bothered improving that. While PyTorch always had an increasing number of features and high level capabilities, TF has always felt like a lower level tool that only made sense choosing if you needed to run models for inference in other platforms. PyTorch Lightning on top of it was a great touch for researchers.