custom MobileNetV2 on OFA #92
-
|
Hi, my group leader assigns the task to reduce the compute cost of mobileNetV2, initially did kernel-size refactoring. I thought to use OFA to check what are the possible modification that can be done. OFA doesn't come with mbv2 by default. can we run custom model on OFA, to get the model which is more efficient on specific dataset? Thank you. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Check out the |
Beta Was this translation helpful? Give feedback.
Check out the
mbv2-w0.35model file in the MCUNet model zoo, which is based on OFA. This model is a scaled-down version of MobileNetV2 designed for microcontrollers.Also, try using the
OFAProxylessNASNetsclass (mcunet/tinynas/elastic_nn/networks/ofa_proxyless.py) as a supernet to train your custom model.