PyTorch entered into the realm of DL framework with the promise of being “Numpy on GPU”. The obvious failures of static graph implementation for certain use cases is increasing industry wide adoption of PyTorch. Dynamic Computation Graph being the backbone of PyTorch, comes with some perks.
- Super intuitive - Shallow learning curve - Amazing community and discussion forum - Easy debugging
It puts you explicitly in control of your compute, there’s no compiler trying to be clever and “help you” or speed your code up, but in reality mostly leading to a massive amount of debugging headaches. PyTorch should be the go-to framework for the problem with “dynamicity” in the core, especially for RNNs and highly dynamic Reinforcement Learning algorithms.
My talk would be to convince the audience about the advantages of PyTorch. For what problems pytorch would be the best solution. I will have benchmarks with popular DL frameworks that review speed, performance, memory, and easiness. I’ll try to build on top of a comparison foundation with TF and NumPy so that the audience could correlate with their day to day research/development projects. The demo would be to portray the best cases where pytorch really shines over other DL frameworks.
Note: Code I developed for the presentation would be available under MIT licence
Slides and more info: anthillinside....