MemoryDataLayer from python

[JN]

Hello,

How do you correctly specify a MemoryData layer using the caffe.NetSpec() functionality in python?

I am trying to do a simple regression using pycaffe. I've defined a network as below. The middle bits are not so important at the moment, they may be right or wrong -  it is the MemoryData layers that cause problems.

My goal is to train the network prediction based on the data (from the first MemoryData layer) to match the target values (the second MemoryData layer).

n = caffe.NetSpec()

n.data = caffe.layers.MemoryData(batch_size=Nbatch, channels=n_statevar, height=1, width=1)

n.target = caffe.layers.MemoryData(batch_size=Nbatch, channels=n_actions, height=1, width=1)

n.data_activator = caffe.layers.TanH(n.data)

n.fc1 = caffe.layers.InnerProduct(n.data, num_output=64, weight_filler=dict(type='xavier'))

n.fc1_activator = caffe.layers.TanH(n.fc1)

n.fc2 = caffe.layers.InnerProduct(n.fc1_activator, num_output=n_actions, weight_filler=dict(type='xavier'))

n.fc2_activator = caffe.layers.TanH(n.fc2)

n.predvalues = caffe.layers.InnerProduct(n.fc2_activator)

n.loss = caffe.layers.EuclidianLoss(n.predvalues, n.target)

The above code writes a prototxt which looks OK but it must be offending somehow: When I try to load it into caffe.SGDSolver it complains about the number of tops from the MemoryData layer:

F0806 10:13:40.205143  9295 layer.hpp:358] Check failed: ExactNumTopBlobs() == top.size() (2 vs. 1) MemoryData Layer produces 2 top blob(s) as output.

I have seen in C++ examples that a "label" blob is associated with the MemoryDataLayer, which I suppose I could just leave hanging/not use in this case (?). But how to specify the prototxt correctly from Python? The resulting prototxt has the following layer def

layer {

  name: "data"

  type: "MemoryData"

  top: "data"

  memory_data_param {

    batch_size: 100

    channels: 2

    height: 1

    width: 1

  }

}

Suggestions would be greatly appreciated!

As the goal is to have a convenient way to load data into the solver and iteratively train it, any alternative route avoiding MemoryDataLayer is also of interest.

Thanks

[Philip H]

try this:

n.data, n.target = caffe.layers.MemoryData(batch_size=Nbatch, channels=n_statevar, height=1, width=1, ntop=2)

and remove the line n.target = ...

cheers,

phil

[JN]

Thanks for the advice, it will surely come handy at some point!

Although I think in this case it doesn't fit my situation; I think the second top blob of a MemoryData layer is assumed to have the same dimensions as the other blob that goes into a loss layer together with labels - correct me if I am wrong on this one. It seems like it is adapted to a classifier situation.

That's why I was trying to go for two separate MemoryLayers so I can control the size of the "target" array. What I am trying to do is more of an arbitrary function approximation given some input variables.

What I ended up doing is using one MemoryData layer, but concatenate the "data" and "target" arrays before feeding to caffe, then use a Slice layer to separate them. The separated data blob goes through the network, the target goes straight to the final loss layer.

MemoryData seems to assume and require a label vector so to keep it happy I feed it in, but never use it.

This seems to be doing the trick in my situation! I can train and run the net forward within a python script now.