Page History

...

{'lhs': (1024, 1024), 'rhs': (1024, 1024), 'initializer': <function normal at 0x117b607b8>, 'run_backward': True, 'dtype': 'float64'}

Future Development and Ideas

The current proposed library is basic bare bone utility for easily running the operator benchmarks across different settings and generate the results in different formats. However, below are some ideas to evolve or build high level abstractions on top of this library:

1. Graphs and insights to understand implication of various settings, such as precision/storage_type/accelerator_type/tensor shapes/etc..., on a given operator.
2. Integration with tools such as MXBoard for debugging and visualization of performance bottlenecks, resource utilization and more fine grained visibility for users when training a model.
3. Integration with nnvm pass for optimization, dynamic accelerator choosing and other operator level decision making capabilities.

Development Plan / Milestones

...