Open sayakpaul opened 3 years ago
@sayakpaul , thanks for your great suggestions. we have an WiP pr here which implemented the space of Swish. We expect users to directly use this module in their PyTorch/TensorFlow model to automatically find out the best activation function specifically for their model.
Great. I will look into it for sure. Should I keep this issue open for other suggestions?
Great. I will look into it for sure. Should I keep this issue open for other suggestions?
yes, we can keep it open for now :)
Sure, thank you. You might want to add a relevant label to this issue for better/easier triage.
Added your vote for MobileNetV3 and MobileDet. Will consider them in future planning. :)
Sure. I personally think teaching the developers how to go about constructing informative search spaces itself is very rewarding.
What would you like to be added:
More case studies on the byproducts of Neural Architecture Search:
These case studies do not need to fully implemented. Minimal demonstrations should be sufficient enough for users to expand the case studies.
Also being able to specify hardware-aware constraints in the search space is an important use-case for NAS. An example showing how to do that would be useful.
Why is this needed:
Without establishing NNI's Retiarii capabilities transfer well to the existing line of works, it might be hard for researchers to adapt Retiarii for their own use.