Original article was published on Deep Learning on Medium
Was their any difference in training time? I’ve found PyTorch can be quite a bit faster usually. Tensorflow’s main advantages are first of all Keras, which makes building models significantly easier than native TF and easier than PyTorch, although there is a limit to the amount of custom stuff you’re able to do with Keras before it makes sense to use either PyTorch or base Tensorflow. The other advantage is the production capabilities and support provided by Tensorflow, although PyTorch is catching up in this regard.
Personally, I like PyTorch more as Tensorflow feels bloated and messy to me, and even with 2.x, I find that PyTorch’s approach to model building is still more intuitive.
What do you think?