e-lab / indoor-eye Goto Github PK

Need to disable dropout when p = 0, otherwise we include unnecessary complexity to the network.

No idea how and what to run. Please, add some documentation as it has been done for

After introducing new Cuda version of SpatialConvolutionMM, RAM usage computation must be verified.

Perhaps we want to look at this? Changing parameters on the fly is something quite useful, I think... Otherwise, if we change them manually, we will end up adding and pushing modifications without real necessity...

Needed more instructions about training.

• Do I need to re-run prepare_imagenet.lua when I change training size? What's the point of chaining it in run.lua if this does not change anything?
• What are the possible values for subsample_classes? The --help should contain all this kind of information, otherwise it's quite helpless.
• Prepare data: "See Data for details." There are no details to see there 👎