One of the biggest complaints I heard about TF 1.x and Keras was that it made it hard to for researchers to implement custom layers/losses/additional functionality. I think the combo of TF 2.0 and Keras addresses those complaints and makes it easier for both researchers & devs.
Fully agree with Adrian. TF 2.0 isn't just a library that does GPU-accelerated computation, it's a ML platform and ecosystem spanning dozens of pieces, covering an incredible diversity of use cases. The scope and depth differ so much from PyTorch that you can't compare the two.https://twitter.com/PyImageSearch/status/1178999156909776896 …
-
-
This Tweet is unavailable.
-
This Tweet is unavailable.
- Show replies
-
-
-
-
Absolutely, the main source of friction in 1.x was having two API silos, high-level Keras on one side and low-level TF on the other side, that weren't well integrated. One of the key innovations in TF 2.0 is to bridge this gap completely. Thread:https://twitter.com/fchollet/status/1178738070381678595 …
End of conversation
New conversation -
-
-
Do you think tf will ever start abstracting away the batch axis? I'm still on cntk because of it.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
TF2.0+Keras is a big improvement, but what's absolutely unacceptable in my experience so far is that a model will happily build, compile, train, and save (either TF or HDF5) and then throw cryptic errors when loading it again. If my model is bad, don't let me save it at all!
-
This is very common in many probabilistic programming languages. It compiles, it runs inference... at a crawling speed, or just hangs. And you have little idea why.
End of conversation
New conversation -
-
-
Tensorflow is good machine learning library, I use in colab. Because I don't useful system for heavy computing
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
The C API appears to be still poorly documented.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
What sort of things does TF 2.0 do that you would you consider to be out of scope for PyTorch?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.