error with HDF5 output layer

[escorciav]

Hi guys,

I used HDF5_OUTPUT layer to save one blob on caffe test. I got an error when the number of iterations is greater than 1. I guest that it tried to overwrite the same file during each iteration.

Probably, I was target a functionality that this layer can't handle so I would like to know:
a) which is the scope of the HDF5 output layer?
b) Have you considered to save the same blob along many iterations?

An straightforward solution is to save n-datasets whose name are "data" plus "_ith-iteration". n:=num_iterations.
c) Could you give me an idea of how to get the iteration number?

[escorciav]

FYI

I read the code and I believe that it can't save many batches. The reason is that it tries to use the same name for the dataset.

I made a simple modification and it can save many datasets in the same hdf5. If you are interested on that, I put my code online.

As far as I understood, you can use my code to dump layers(blobs) on hdf5 files calling caffe binary.

I/you have to test that you will get the data in the same order. The reason is that I don't get the iteration number from Net. Instead, I use a new attribute on the hdf5_output_layer. It is possible that they are not synchronized.

Regards,