Both frameworks operate on tensors and view any model as a directed acyclic graph (DAG), but they differ drastically on how you can define them.
TensorFlow follows ‘data as code and code is data’ idiom. In TensorFlow you define graph statically before a model can run. All communication with outer world is performed via tf.Session object and tf.Placeholder which are tensors that will be substituted by external data at runtime.
In PyTorch things are way more imperative and dynamic: you can define, change and execute nodes as you go, no special session interfaces or placeholders. Overall, the framework is more tightly integrated with Python language and feels more native most of the times. When you write in TensorFlow sometimes you feel that your model is behind a brick wall with several tiny holes to communicate over. Anyways, this still sounds like a matter of taste more or less.
However, those approaches differ not only in a software engineering perspective: there are several dynamic neural network architectures that can benefit from the dynamic approach. Recall RNNs: with static graphs, the input sequence length will stay constant. This means that if you develop a sentiment analysis model for English sentences you must fix the sentence length to some maximum value and pad all smaller sequences with zeros. Not too convenient, huh. And you will get more problems in the domain of recursive RNNs and tree-RNNs. Currently Tensorflow has limited support for dynamic inputs via Tensorflow Fold. PyTorch has it by-default.
Reference:
https://medium.com/towards-data-science/pytorch-vs-tensorflow-spotting-the-difference-25c75777377b
[D] So… Pytorch vs Tensorflow: what’s the verdict on how they compare? What are their individual strong points?
byu/cjmcmurtrie inMachineLearning