Give the python layer parameter/weight blobs.

[philkr]

This PR allows the python layer to have its own parameter/weight blobs.
As any python layer inherits form caffe.Layer it already has a self.blobs variable that gets saved and loaded properly, however it wasn't possible to add a new blob to that BlobVec. This PR adds a function add_blob to the BlobVec which allows a python layer to extend the vector with a new blob. The arguments of the add_blob function are the dimensions of the blob to be added.

Here is a simple layer that demonstrates the use.

import caffe

class Param(caffe.Layer):
    def setup(self, bottom, top):
        self.blobs.add_blob(1,2,3)
        self.blobs[0].data[...] = 0

    def reshape(self, bottom, top):
        top[0].reshape(10)

    def forward(self, bottom, top):
        print(self.blobs[0].data)
        self.blobs[0].data[...] += 1

    def backward(self, top, propagate_down, bottom):
        pass


def main():
    from caffe import NetSpec, layers as L, params as P, to_proto
    from os import path
    ns = NetSpec()

    A = L.Python(module='test_phase', layer='Param', ntop=1, name='foo')
    open('/tmp/test.prototxt', 'w').write(str(to_proto(A)))
    net = caffe.Net('/tmp/test.prototxt', caffe.TRAIN )
    if path.exists('/tmp/test.caffemodel'):
        net.copy_from('/tmp/test.caffemodel')
    print( "Loaded" )
    net.forward()
    net.forward()
    net.save('/tmp/test.caffemodel')

if __name__ == "__main__":
    main()

[longjon]

Generally for this purpose I am using #2079, which ought to eventually in some way in accord with #1474 provide a more natural way to deal with parameters, esp. when using net spec.

But this seems like a reasonable exposure of the current state affairs. It would be nice if Layer.blobs was treated as a more Pythonic list-like thing, but that's understandably difficult with vector_indexing_suite, so this interface seems pretty reasonable and the code is unobtrusive.

For merge:

• Could we provide a basic test of this functionality?
• Please do follow the existing style (even where not checked by lint); in particular, one space between functions, and use explicit braces around blocks.

[philkr]

Fixed the style and added some tests.

[longjon]

Looks good, thanks @philkr!

[longjon]

Ah, I just merged this, so this comment is too late, but one more thing for the future: please squash style changes (here they appear in the test commit, which is confusing).

[peerajak]

Thank you very much. Exactly what I have been looking for

It terns out that this is perfectly what I want! the solver do update the weights!. I tested it by setting initial weights to zeros and backward with ones. the result diff is with 0-base_lr*1.
the way to access this new blobvec is as follow
solver.net.params['loss'][0].data