Python net specification

[longjon]

Protobuf is powerful, but is not a general-purpose language, and can be cumbersome in its current use for writing down complicated chains of functions, i.e., nets. See previous discussion at (nonexhaustively) #1290 and #1169.

Python is a general-purpose language with clean syntax and an intimate relationship with Caffe. This PR allows net specification as (domain-specific) Python code (and consequently, effortless programmatic net construction).

Currently this is just a proof-of-concept sketch, which lightly sprinkles Python magic over the caffe_pb2 protobuf interface.

LeNet can be written as:

from caffe import layers as L
from caffe import params as P

data = L.data(batch_size=64, backend=P.data.LMDB, source='examples/mnist/mnist_train_lmdb')
conv1 = L.convolution(data, kernel_size=5, num_output=20)
pool1 = L.pooling(conv1, kernel_size=2, stride=2, pool=P.pooling.MAX)
conv2 = L.convolution(pool1, kernel_size=5, num_output=50)
pool2 = L.pooling(conv2, kernel_size=2, stride=2, pool=P.pooling.MAX)
ip1 = L.relu(L.inner_product(pool2, num_output=500))
prob = L.softmax(L.inner_product(ip1, num_output=10))

And can be instantiated via this hackiness, to be cleaned up later:

with open('tmp.prototxt', 'w') as f:
    print >>f, prob.to_proto()
net = caffe.Net('tmp.prototxt')

Things to think about:

• Perhaps there should be a way to construct Nets directly from Layers, without going through protobuf.
• The magic in use here means syntax doesn't really gets checked until the Net is created.
• This will need to be rebased on Layer type is a string #1694 and more significantly Treat bottoms and params more uniformly? #1474.
• It would be nicer if the functions constructed here simply were Nets rather than having to make Nets out of them.
• There is no way to control blob names right now, but there are ways to make that happen...

[netheril96]

Perhaps there should be a way to construct Nets directly from Layers, without going through protobuf.

This is because the constructor of net unnecessarily involves I/O. It should instead take a NetParameter as argument, instead of (or in addition to) a filename as string.

[jeffdonahue]

This looks nifty. What happens with multiple tops? (e.g. in your example, can you do data, label = L.data(...)?)

the constructor of net unnecessarily involves I/O. It should instead take a NetParameter as argument, instead of (or in addition to) a filename as string.

I agree, there should definitely be a way to construct a pycaffe Net by passing a caffe_pb2.NetParameter thing (if you have one) rather than writing it to a file -- PRs welcome (or I'll probably try to do this at some point if nobody else does).

[longjon]

Re: multiple tops, I had in mind your syntax @jeffdonahue, but I don't think I implemented it here yet.

Yes, there should be a way to construct Nets from NetParameters, although that was not what I had in mind with the comment above, but rather a way to construct Nets without involving protobuf at all. In this sketch, one constructs some special Python class which specifies a net and then turns it into a Net. Why doesn't one just construct a Net to begin with?

[longjon]

Closing to reopen against master.

[kevin-li]

Is there a way to modify existing network from within python? Or is there a way to edit prototxt by coding instead of manully changing them?

My experiment requires freezing/training different layers, how can I do this through python code?

Thank you.