How to define a layer with parameter include

[王驰航]

In python we can do something like this:

    n = caffe.NetSpec()
    n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
                             transform_param=dict(scale=1./255), ntop=2)
    n.conv1 = L.Convolution(n.data, kernel_size=5, num_output=20, weight_filler=dict(type='xavier'))

But how do we define a layer which has flag include:{phase: TEST}?

I tried something like:

n.accuracy = L.Accuracy(n.ip1, n.label, incluude=dict(phase=caffe.TEST))

But apparently doesn't work

[I-Chao Shen]

I have the same issue.

My problem is that I want to indicate different data layers for train and test, respectively.

However, when I trace the code in caffe_pb2, I found I can use 'include', but I can not successfully use it.

[KaitBristowe]

I don't think you can include TEST/TRAIN phases.  I would just make a test net and a train net and then specify both of these in the solver prototxt:

net: "train net"

test_net: "test_net"

test_interval: X

test_iter: X

[徐珍琦]

I'm facing the same problem. How do you deal with this issue?

在 2015年7月7日星期二 UTC+8下午11:43:24，Chihang Wang写道：

[Georg Waltner]

This should work:

n.accuracy = L.Accuracy(n.ip1, n.label, include=[dict(phase=1)])

HTH, Georg.

[Shai Bagon]

How can I do this for an input layer? That is having a data layer that has the SAME TOP name for TRAIN and TEST phases?

[Jeremy Rutman]

i bump this question, have the same issue.

obviously there will be a conflict if you do 

n.data, n.label = L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=test_lmdb,
                             transform_param=dict(scale=1./255), ntop=2,include=[dict(phase=1)])

and then try to define same variables using a different db with include=[dict(phase=0)]

The mechanism by which pycaffe allows dicts with multiple uses of the same key is with lists - for instance to get

  transform_param {

    scale: 0.00392156862745

    mean_value: 112

    mean_value: 123

    mean_value: 136

  }

we do transform_param=dict(scale=1./255,mean_value=[meanB,meanG,meanR])

so conceivably to get multiple layer defs with the same name one could analogously try a list  

n.data, n.label = [L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=train_lmdb,
                             transform_param=dict(scale=1./255), ntop=2,include=[dict(phase=0) ,

                   L.Data(batch_size=batch_size, backend=P.Data.LMDB, source=lmdb,
                             transform_param=dict(scale=1./255), ntop=2,include=[dict(phase=1)]) ]

i will try it as soon as i obtain two donuts

[Jeremy Rutman]

on second thought that will not work as a,b = f(x)  can't be replaced by a,b=[f(x1),f(x2)]
furthermore there are no donuts

[Marina V]

Hey, have you figured this one out?

[Jeremy Rutman]

I think I wound up using Lisa's answer namely separate definitions of training and test nets