How to implement dynamic inference on MutualNet about mutualnet HOT 5 CLOSED

@Liuyang829 Hi, in real time, you don't have to test all these configurations on every sample. We test all these configurations to get the best configuration for each constraint (e.g. FLOPs). Then this makes our query table where each FLOPs constraint has their corresponding configuration. In real time, you choose the configurations from the query table based on your FLOPs constraint.

from mutualnet.

I understand the usage of query table, but if you use FLOPs as the reference standard of constraint, maybe we can't know the FLOPs constraint of this device, so sometimes the query table may lacks actual meaning in real time.
In my opinion, If you can make it dynamic or adaptive in test, it will achieve more efficient usage.

from mutualnet.

You can also use the latency as the reference constraint, then you have the config-latency query table.

from mutualnet.

Thank you very much. When I try to train a model with ImageNet on MobileNetv1 backbone, I find that the Memory-Usage of GPU is always very high, but the volatile GPU-Util is very low, even 0%. I want to know why this situation happened and how to avoid this. Thanks.

from mutualnet.

This maybe caused by the data loader. You can improve the data loader speed by

1. increase the num_workers
2. put data on SSD

from mutualnet.