Google recently included in tensorflow's nightly builds its Eager mode, an imperative API to access tensorflow computation capabilities.
How do tensorflow eager compare to PyTorch?
Some aspects that could affect the comparison could be:
- Advantages and disadvantages of eager due to its static graph legacy (e.g. names in nodes).
- Intrinsic limitations of either of them that the other does not have.
- Areas in which one of them needs improvement (e.g. feature completeness, computational optimizations).
- Ecosystem differences (e.g. tensorboard?).
Note1: Yaroslav Bulatov wrote a review about eager's nice features.
Note2: In a previous question, I requested a comparison between PyTorch and Tensorflow Fold. At that time, it seemed to me that Fold could face PyTorch thanks to Google backing it. I was very very wrong: in the end, Google itself abandoned Fold in favour of Eager. I understand that this was due to intrinsic limitations in the normal tensorflow API that led Fold not to be very friendly, which constrained its adoption.