DeepCL_sl.py

import cPickle as pickle
import numpy
import PyDeepCL
import re


# Warning: this module is just an approximate seriallize to the net
# It does not cover each property of NeuralNet 
# Use it carefully 
# Working in Python 2.7, PyDeepCL 10.3.0a3 / DeepCL 10.1.0

def NNSave(net):
    # get a PyDeepCL.NeuralNet Object and return a serialized result with pickle protocol
    DB = [None, ]*net.getNumLayers()
    for idx in range(net.getNumLayers()):
        try:
            M = net.getLayer(idx)
            DB[idx] = {
                "NetdefString": M.getNetdefString(),
                "OutPutSize": M.getOutputSize(),
                "Planes" : M.getOutputPlanes(),
                "Name": M.getClassName(),
                "Weight": M.getWeights()
            }
            NetString = net.asString()
            Q = net.getLayer(idx).getWeights()
            # print(numpy.array(Q).shape, M.getClassName())
        except:
            import traceback
            print "Caught error when dealing layer%s "%idx, net.getLayer(idx).getClassName()
            traceback.print_exc()
            

    return pickle.dumps((DB, net.getNetdef(), net.asString()))

def NNLoad(cl, Seriallized_net):

    # I don't check if is illegal string of input, take care about it'
    assert isinstance(cl, PyDeepCL.DeepCL)
    H, NetString, NetasString = pickle.loads(Seriallized_net)
    # Input Layer
    net = PyDeepCL.NeuralNet(cl, H[0]["Planes"], H[0]['OutPutSize'])
    # Uh......
    for i in range(1, len(H)):
        Layer_Now = H[i]
        Layer_Name = re.sub("Layer$", "", Layer_Now["Name"])

        Dict = {
            "RandomTranslations": ("rt(\d+)",         (lambda p1:         PyDeepCL.RandomTranslationsMaker().translateSize(int(p1)))),
            "Convolutional"     : ("(\d+)c(\d+)(z?)", (lambda p1, p2, p3: PyDeepCL.ConvolutionalMaker().numFilters(int(p1)).filterSize(int(p2)).padZeros().biased())),
            "Activation"        : ("\S+",             (lambda p1:         getattr(PyDeepCL.ActivationMaker(), p1)())),
            "Pooling"           : ("mp(\d+)(z?)",     (lambda p1, p2:     PyDeepCL.PoolingMaker().poolingSize(int(p1)))),
            "FullyConnected"    : ("(\d+)n",          (lambda p1:         PyDeepCL.FullyConnectedMaker().numPlanes(Layer_Now["Planes"]).imageSize(1).biased())),  # ImageSize is a default value
            "SoftMax"           : ("",                (PyDeepCL.SoftMaxMaker)),
            "Input"             : ("",                (lambda : (_ for _ in ()).throw(Exception("InputLayer should be created in initialize, There must be something wrong")))),
            "Normalization"     : ("",                (lambda : (_ for _ in ()).throw(Exception("I don't know how to rebuild this layer, could you help me?"))))
        }
        print "Adding %s%s......."%(Layer_Name, " - %s"%Layer_Now["NetdefString"] if Layer_Now["NetdefString"] else "")
        if Layer_Name in Dict.keys():
            Tmp = Dict[Layer_Name]
            G = re.findall(Tmp[0], Layer_Now["NetdefString"])[0]
        
            if isinstance(G, tuple):
                net.addLayer(Tmp[1](*G))
            else:
                net.addLayer(Tmp[1](G))
        else:
            raise BaseException("Layer_Name %s cannot be added beacause of NOT FOUND"%Layer_Name)
    print "rebuild the net are as follows:"
    print net.asString()
    print "the net storage in file are as follows, please check it:"
    print NetasString

    print "Loading Weights......"
    for i in range(0, len(H)):
        if not isinstance(H[i]["Weight"],type(None)):
            net.getLayer(i).setWeights(H[i]["Weight"])
    return net