Python net specification

[longjon]

master edition of #1733. Still rough, but should be usable by the intrepid.

• Now including an AlexNet (CaffeNet variant) generation example.
• Fixed an error in uncamel which broke acronym names (e.g., LRN) (thanks @sontran).
• Now supports fillers (thanks @sontran).

[longjon]

This should now support repeated Messages as lists of dicts (like param or dummy_data_param's shape; @erictzeng, you asked about this earlier).

I think that means you can now specify any NetParameter as Python code. Once layer naming has been cleaned up a bit (for which I have an idea in mind), I think this will have reached mergeability as a thin wrapper around prototxt.

[muupan]

uncamel('HDF5Data') wrongly returns 'hd_f5_data'.

[muupan]

And uncamel('PReLU') wrongly returns 'p_relu'.

I think 'HDF5Data' -> 'hdf5_data' is normal uncamelling but 'PReLU' -> 'prelu' is exceptional. Maybe we have to find some other way.

[seanbell]

@muupan The layers that break the rule could just be hardcoded in the "uncamel" function, e.g. with a dictionary. Something like this:

_UNCAMEL_EXCEPTIONS = {
    'HDF5Data': 'hdf5_data',
    'PReLU': 'prelu',
}

def uncamel(s):
    """Convert CamelCase to underscore_case."""
    return _UNCAMEL_EXCEPTIONS.get(s, 
        re.sub('(?!^)([A-Z])(?=[^A-Z])', r'_\1', s).lower())

[Shaunakde]

@longjon I am having an issue using this PR. Running the example: http://nbviewer.ipython.org/github/BVLC/caffe/blob/tutorial/examples/01-learning-lenet.ipynb gives me the following error:

 File "/home/shaunak/caffe-pr2086/examples/wine/classify.py", line 18, in lenet
    n = caffe.NetSpec()

  File "../../python/caffe/layers.py", line 84, in __init__
    super(NetSpec, self).__setattr__('tops', OrderedDict())

TypeError: must be type, not None

Update

I tried adding an import statement for caffe as well and the following happened:

import numpy as np
import matplotlib.pyplot as plt

# Make sure that caffe is on the python path:
caffe_root = '../../'  # this file is expected to be in {caffe_root}/examples/wine
import sys
sys.path.insert(0, caffe_root + 'python')

from pylab import *

import caffe 

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

def logreg(hdf5, batch_size):
    # logistic regression: data, matrix multiplication, and 2-class softmax loss
    n = caffe.NetSpec()
    n.data, n.label = L.HDF5Data(batch_size=batch_size, source=hdf5, ntop=2)
    n.ip1 = L.InnerProduct(n.data, num_output=2, weight_filler=dict(type='xavier'))
    n.accuracy = L.Accuracy(n.ip1, n.label)
    n.loss = L.SoftmaxWithLoss(n.ip1, n.label)
    return n.to_proto()

with open('../../examples/hdf5_classification/logreg_auto_train.prototxt', 'w') as f:
    f.write(str(logreg('examples/hdf5_classification/data/train.txt', 10)))

with open('../../examples/hdf5_classification/logreg_auto_test.prototxt', 'w') as f:
    f.write(str(logreg('examples/hdf5_classification/data/test.txt', 10)))

causes this error:

runfile('/home/shaunak/caffe-pr2086/examples/wine/classify.py', wdir='/home/shaunak/caffe-pr2086/examples/wine')
Reloaded modules: caffe, caffe.proto, caffe._caffe, caffe.pycaffe, caffe.detector, caffe.proto.caffe_pb2, caffe.io, caffe.classifier, caffe.layers
Traceback (most recent call last):

  File "<ipython-input-9-694741de221d>", line 1, in <module>
    runfile('/home/shaunak/caffe-pr2086/examples/wine/classify.py', wdir='/home/shaunak/caffe-pr2086/examples/wine')

  File "/home/shaunak/anaconda/lib/python2.7/site-packages/spyderlib/widgets/externalshell/sitecustomize.py", line 682, in runfile
    execfile(filename, namespace)

  File "/home/shaunak/anaconda/lib/python2.7/site-packages/spyderlib/widgets/externalshell/sitecustomize.py", line 78, in execfile
    builtins.execfile(filename, *where)

  File "/home/shaunak/caffe-pr2086/examples/wine/classify.py", line 26, in <module>
    f.write(str(logreg('examples/hdf5_classification/data/train.txt', 10)))

  File "/home/shaunak/caffe-pr2086/examples/wine/classify.py", line 23, in logreg
    return n.to_proto()

  File "../../python/caffe/layers.py", line 97, in to_proto
    top.fn._to_proto(layers, names, autonames)

  File "../../python/caffe/layers.py", line 78, in _to_proto
    assign_proto(layer, k, v)

  File "../../python/caffe/layers.py", line 25, in assign_proto
    setattr(proto, name, val)

AttributeError: 'LayerParameter' object has no attribute 'source'

Discussion: http://stackoverflow.com/questions/29774793/typeerror-python-class

[escorciav]

Hi @Shaunakde , you safe me hours of reading code so I felt that I should help you (sorry if you already noticed that). You should use the comment of seanbell for your example. I guess that the reason is that HDF5Data is an CamelCase tricky layer.

Thank you longjon for this tool. I have spent hours debugging prototxt without noticing minor difference such as layers instead of layer.

[BlGene]

I was looking over this and from what I can tell the to_proto function from PR #1733 has moved
into NetSpec here.

In this case NetSpec should be imported and all layer variables should be defined to it. This would then also print the layers in their correct order.

The alexnet function would look as follows:

def alexnet(lmdb, batch_size=256, include_acc=False):
    n = NetSpec()

    n.data, n.label = L.Data(source=lmdb, backend=P.Data.LMDB, batch_size=batch_size, ntop=2,
        transform_param=dict(crop_size=227, mean_value=[104, 117, 123], mirror=True))

    # the net itself
    n.conv1, n.relu1 = conv_relu(n.data, 11, 96, stride=4)
    n.pool1 = max_pool(n.relu1, 3, stride=2)
    n.norm1 = L.LRN(n.pool1, local_size=5, alpha=1e-4, beta=0.75)
    n.conv2, n.relu2 = conv_relu(n.norm1, 5, 256, pad=2, group=2)
    n.pool2 = max_pool(n.relu2, 3, stride=2)
    n.norm2 = L.LRN(n.pool2, local_size=5, alpha=1e-4, beta=0.75)
    n.conv3, n.relu3 = conv_relu(n.norm2, 3, 384, pad=1)
    n.conv4, n.relu4 = conv_relu(n.relu3, 3, 384, pad=1, group=2)
    n.conv5, n.relu5 = conv_relu(n.relu4, 3, 256, pad=1, group=2)
    n.pool5 = max_pool(n.relu5, 3, stride=2)
    n.fc6, n.relu6 = fc_relu(n.pool5, 4096)
    n.drop6 = L.Dropout(n.relu6, in_place=True)
    n.fc7, n.relu7 = fc_relu(n.drop6, 4096)
    n.drop7 = L.Dropout(n.relu7, in_place=True)
    n.fc8 = L.InnerProduct(n.drop7, num_output=1000)
    n.loss = L.SoftmaxWithLoss(n.fc8, n.label)

    if include_acc:
        n.acc = L.Accuracy(n.fc8, n.label)

    return n.to_proto()

BR, Max

[BlGene]

Hi,

I've been using this PR for a few days and I have been able to write all the models I wanted to using it. I am very pleased with it and Ty for writing it. 👍 @longjon

The PR mentions that it fillers, but I don't see how these can be accessed from python, was this omitted from the PR?

I was wondering if there is an easy way to specify parameters for python layers, here a few ideas came to mind:

1. Just generating the python layer code with the variables in place
2. Extending PythonParameter in order to smuggle a dict through to the python layer:

message PythonParameter {
  optional string module = 1;
  optional string layer = 2;
  repeated string param_keys = 3;
  repeated string param_values = 3;
}

1. Extending PythonParameter with param_string, then people can put pickled parameters in it/ do whatever they want.

So for this I was wondering if there is an easier way.

Also in case it was overlooked, its worthwhile looking at the Theano version of this, called Mariana.

[bhack]

See also Lasagne on Theano.

[longjon]

Getting ready to update this; here's list of TODOs:

Cleanups and fixes before first mergeable version, which should be considered basically a "protobuf wrapper/generator":

• Try a better solution to the implicit layer <-> parameter correspondence (currently the messy uncamel function); the fact that the current protobuf system leaves this unspecified means there is no clean answer, but I have an idea for a hack that's a little more robust.
• Finish the transition to the NetSpec class (basically a namespace to avoid the previous atrocious hack of abusing locals (this was left as a WIP commit).
• Minimal tests.
• Minimal docs/examples (as begun).

Desiderata and future work to come after merge (PRs welcome!):

• Operator overloading for arithmetic.
• Update in accord with the solution to Treat bottoms and params more uniformly? #1474.
• Make Python layer integration more natural (possibly also by updating the way Python layers are specified in protobuf, in accord with Switch LayerType to string for extensibility #1685).
• Provide a direct path from spec -> Net without serializing through a file.
• Check things that can be checked at specification time (e.g., layer name validity, unused tops).
• Support silence layers.

[longjon]

TODOs all done for now; marking this ready for review!

@muupan @seanbell and others, uncamel is gone; the layer-parameter correspondence is now determined through inspection of the caffe_pb2 module. The only assumption is that the parameter type of a layer named X is XParameter, which is true for all existing layers and should remain true.

@Shaunakde, note that we generally aren't able to provide support for PRs, especially in-progress ones. You're welcome to contribute to the development discussion, but otherwise please use caffe-users or other venues.

@BlGene, see the tests for an example of specifying fillers. Parameters for Python layers are a different (though related) issue; see, e.g., #2001.

[shelhamer]

Trivial, but this should be called caffenet since it has our usual inversion.

[shelhamer]

@longjon amended names and the (AttributeError, KeyError) exception handling and pushed. Merging.

[kjmonaghan]

Very helpful, thank you!

How do I go about changing the decay_mult parameter?

[jeffdonahue]

e.g., L.Convolution(data, kernel_size=5, num_output=20, param=[dict(decay_mult=0.5)]). Any proto-generated object (in this case a ParamSpec) can be specified using a dict. The brackets [] are needed to make a list of them, since param is a repeated field (to support multiple parameters). You could also explicitly use a ParamSpec object (rather than a dict that gets converted into one) by doing:

filter_spec = caffe_pb2.ParamSpec()
filter_spec.decay_mult = 0.5
L.Convolution(data, kernel_size=5, num_output=20, param=[filter_spec])

(untested, could be slightly wrong)

[jmerkow]

Is it possible to specify dummy data with this?
I've tried about 1000 things to get it to work and can't seem to get it...

Based on caffenet.py, I would expect that it would be something like this:

import caffe
from caffe import layers as L, params as P, to_proto
from caffe.proto import caffe_pb2
from __future__ import print_function
def gen_inputs():
    data = L.DummyData(name="data",ntop=1,shape=dict(dim=[1,2,3]))
    label = L.DummyData(name="label",ntop=1,shape=dict(dim=[1,2,3]))
    return data,label

def caffenet(include_acc=False):
    data,label=gen_inputs()

    loss = L.SoftmaxWithLoss(data, label)
    return to_proto(loss)

def make_net(output_dir='./',net_name='train'):
    fname = os.path.join(output_dir,net_name+'.prototxt')
    with open(fname, 'w') as f:
        print(caffenet(), file=f)
make_net()

this produces:

layer {
  name: "data"
  type: "DummyData"
  top: "DummyData1"
}
layer {
  name: "label"
  type: "DummyData"
  top: "DummyData2"
}
layer {
  name: "SoftmaxWithLoss1"
  type: "SoftmaxWithLoss"
  bottom: "DummyData1"
  bottom: "DummyData2"
  top: "SoftmaxWithLoss1"
}

I've tried various other things using caffe_pb2 and caffe.params.
And I tried assigning dummy_data_param explicitly and with dictionaries:

def gen_inputs():
    data = L.DummyData(name="data",ntop=1,dummy_data_param=dict(shape=dict(dim=[1,2,3])))
    label = L.DummyData(name="label",ntop=1,dummy_data_param=dict(shape=dict(dim=[1,2,3])))
    return data,label

def gen_inputs():
    data_shape = caffe_pb2.BlobShape()
    data_shape.dim = [1,2,3]
    data_param = caffe_pb2.DummyDataParameter()
    data_param.shape = data_shape
    data = L.DummyData(name="data",ntop=1,dummy_data_param=data_param)
    label = L.DummyData(name="label",ntop=1,dummy_data_param=data_param)
    return data,label

Any thoughts?
--Jameson

[BlGene]

Maybe this?:
(Because shape is a repeated field in DummyDataParameter it should be given as a list.)

from __future__ import print_function
import caffe
from caffe import layers as L, params as P, to_proto
from caffe.proto import caffe_pb2
import os

def gen_inputs():
    data  = L.DummyData(name="data", ntop=1,dummy_data_param=dict(shape=[dict(dim=[1,2,3])]))
    label = L.DummyData(name="label",ntop=1,dummy_data_param=dict(shape=[dict(dim=[1,2,3])]))
    return data,label

def caffenet(include_acc=False):
    data,label=gen_inputs()

    loss = L.SoftmaxWithLoss(data, label)
    return to_proto(loss)

def make_net(output_dir='./',net_name='train'):
    fname = os.path.join(output_dir,net_name+'.prototxt')
    with open(fname, 'w') as f:
        print(caffenet(), file=f)

make_net()

BR, Max

[longjon]

@BlGene is right. If bad parameters are being silently ignored, however (in current master), that's a bug and you're welcome to open an issue.

[jmerkow]

I can open it with this as a test case. It may be the name 'shape' which has meaning in some contexts?

[dfagnan]

@longjon Does this caffenet example get all the weight_filler and bias_fillers correct? I'm not able to easily see that the default weight_filler is somehow gaussian with std = 0.01 here. Is this true?