GTC Silicon Valley-2019: Deep Learning with Myia
GTC Silicon Valley-2019 ID:S9244:Deep Learning with Myia
Myia is an experimental deep learning framework that aims to help researchers implement differentiable models with complex control flow and run them efficiently on GPUs. These models may be implemented using a subset of Python that includes conditionals, loops, dataclasses, higher order functions, as well as recursion. Myia is then able to run powerful type and shape inference (without the need for type annotations) to guarantee correctness prior to running the model, differentiate the whole or parts of the model, and finally aggressively optimize the resulting code to execute it as fast as possible. Unlike other frameworks such as TensorFlow or PyTorch, which are more like domain-specific languages or libraries, our approach is grounded in programming language and compiler theory, which affords it greater generality. We demonstrate Myia on a variety of models requiring complex control flow and contrast program size and performance with other frameworks.