Is SpatialDropout used in the code?

[liu09114]

In the code, spatialdropout is used like this:

But this can not drop channel to zeros? Is the code consistancy with the paper?

[liu09114]

gaussian dropout

[liu09114]

Results are not stable. The searched channels are very different for many times.

[danczs]

Thanks for your questions. Sorry for not answering in time.
Binary dropout is intuitive for data drop. Gaussian dropout is more like adding noise to data. Both binary dropout and Gaussian dropout work well for our method.

As for the stability.
One may get different architectures with different runs. For example, the channel number of a layer may increase from 64 to 120 in a run and only increase to 100 in another run, but the general trends are stable. Like the gradient-based optimization algorithms, more search iterations and small architerture learning rates can lead to more stable results. (e.g. –iters=20 –arch_learning_rate=0.1 )