Do you have different checkpoints?

[tommiekerssies]

Hi, I am using your supernet for finding a good neural architecture for anomaly detection on MVTec. I want to find out whether the found optimal architecture is consistent. Did you train the supernet from different seeds for example, or can provide me with some other checkpoints somehow?

[NickChang97]

Sorry for that, we didn't test the consistentency for seached subnets. I think you may refer to the paper in supervised one-shot NAS methods, the conclusion maybe the same here. Another recommendation is to train super-net with 10% ImageNet, we verify results in such setting.

[NickChang97]

The reason for no more checkpoints is accually that a engineering problem which costs much time in training is encountered in training super-net. In each iteration, the network should forward three times and backward one times, while it would cause torch error. We use a ugly workround to solve it (refer to BigNasBasedRunner) forward then backward for three times in each iteration