HDf5Data Layer transform params ??

[Ashwin Nair]

I've heard that the HD5 layer could be useful with multiple tops. But I ran into a few problems.

I've referred to the idea used here to create a data set with multiple tops.

The only change I've made to it is add a 1*4 vec 1,0,1,1 as a placeholder for the real data in a dataset named "data2"

I would like to pass the image data through a ready available network like alexnet.

This is how my model looks.

Model

Code for creation of dataset

However I was unable to perform transform actions like mean subtraction and coping or mirroring.

Does any one know how this could be done?

[Han]

It seems hdf5 layer doesn't support transform actions as the lmdb layer did. (correct me if I was wrong)

You can use a separate lmdb layer as the label if you want a vector as the label instead of a single number. 

But the problem is, if you have two separated lmdb files for data and label, you may need to transform them synchronously, which is very common in 'image to image' training. But I have no idea how to do them efficiently. Normally, I will write my own python layer to transform label and data simultaneously, but it is very slow. It's bother me a lot. Any suggestion is welcomed.

[Ashwin Nair]

I think your right HD5 does not allow any transformation.

I have the same problem of synchronization as well. That is Y I decided to use the HD5 layer as we can easily add data to it with mul tops.

But the lack to transformation can cause a few problems if I cant have a zero centered mean.(Possibly not sure. )

And would this lack for transformation also give the same global minimum for the same data ?