def __init__(self,
inputDim,
convFilterNum,
convFilterDim,
convImageSize,
poolFilterDim,
poolImageSize,
hiddenDim,
outputDim,
seed=1):
#畳み込み+プーリング付き３層ニューラルネット
self.inputLayer = layer.InputLayer(inputDim)
self.convLayer = layer.ConvolutionalLayer(convFilterNum, convFilterDim,
convImageSize,
self.inputLayer, seed)
self.poolLayer = layer.PoolingLayer(poolFilterDim, poolImageSize,
self.convLayer)
self.hiddenLayer = layer.HiddenLayer(hiddenDim, self.poolLayer, seed)
self.outputLayer = layer.OutputLayer(outputDim, self.hiddenLayer, seed)
self.hiddenLayer.setActivator(util.relu)
self.hiddenLayer.setBackPropagator(learning.backPropReLU)
# 正解データ
self.answer = np.zeros(outputDim)
def __init__(self, sizes, cost=CrossEntropyCost, act=SigmoidActivation):
"""The list ``sizes`` contains the number of neurons in the
respective layers of the network. For example, if the list
was [2, 3, 1] then it would be a three-layer network, with the
first layer containing 2 neurons, the second layer 3 neurons,
and the third layer 1 neuron. The biases and weights for the
network are initialized randomly, using a Gaussian
distribution with mean 0, and variance 1. Note that the first
layer is assumed to be an input layer, and by convention we
won't set any biases for those neurons, since biases are only
ever used in computing the outputs from later layers."""
self.num_layers = len(sizes)
self.sizes = sizes
self.default_weight_initializer()
self.cost = cost
self.activ_fn = act
# line zip is to generate a full connected network weights
self.layers = [
layer.BNLayer(w, b, act)
for w, b in zip(self.weights[:-1], self.biases[:-1])
]
' output can be configured by output config string '
cost_act = None if cost.combined() else act
self.output_layer = layer.OutputLayer(self.weights[-1],
self.biases[-1], cost_act, cost)
def __init__(self, inputDim, hiddenDim, outputDim, seed=1):
# 三層ニューラルネットワーク
self.inputLayer = layer.InputLayer(inputDim)
self.hiddenLayer = layer.HiddenLayer(hiddenDim, self.inputLayer, seed)
self.outputLayer = layer.OutputLayer(outputDim, self.hiddenLayer, seed)
# 正解データ
self.answer = np.zeros(outputDim)
def calcLossTest():
driver = DriverLayer()
outputLayer = layer.OutputLayer(9, driver)
out = outputLayer.calculate()
label = np.array([[0], [0], [0], [1], [0], [0], [0], [0], [0]])
print(out)
print(label)
print(util.crossEntropy(out, label))
def __init__(self, inputDim, hiddenDim, outputDim, seed=1):
# 三層ニューラルネットワーク
self.inputLayer = layer.InputLayer(inputDim)
self.hiddenLayer = layer.HiddenLayer(hiddenDim, self.inputLayer, seed)
self.outputLayer = layer.OutputLayer(outputDim, self.hiddenLayer, seed)
self.hiddenLayer.setActivator(util.relu)
self.hiddenLayer.setBackPropagator(learning.backPropReLU)
# 正解データ
self.answer = np.zeros(outputDim)
def outputLayerTest():
driver = DriverLayer()
outputLayer = layer.OutputLayer(9, driver)
print(outputLayer.calculate())
outputLayer.setActivator(util.correctedSoftmax)
print(outputLayer.calculate())
outputLayer.setWeight(
outputLayer.maker.getIdentityComplex(outputLayer.weightSize))
outputLayer.setShift(np.zeros(outputLayer.shiftSize))
print(outputLayer.calculate())
#print(outputLayer.getOutput())
return outputLayer.getOutput()
def __init__(self, layers):
# 输入一个list，存放每一个hiddenLayer的节点数
self.network = []
self.network.append(layer.InputLayer()) # 输入层
i = 0
while i < len(layers):
self.network.append(
layer.HiddenLayer(layers[i], self.network[i], None))
i = i + 1
self.network.append(layer.OutputLayer(self.network[i]))
i = 0
while i < len(self.network) - 1:
self.network[i].next_layer = self.network[i + 1]
if i != 0:
self.network[i].initialize() # 对中间层重新进行初始化
i = i + 1
self.network[-1].initialize()
def inputConvTest():
inputM = ioex.InputManager()
outputM = ioex.OutputManager()
trainingData = inputM.getMnistTrainingData()
sample = trainingData.getImageBatch(5)
outputM.showPictureFromBatch(sample, (28, 28))
inputLayer = layer.InputLayer(28 * 28)
conv = layer.ConvolutionalLayer(4, 3, inputLayer)
outputLayer = layer.OutputLayer(10, conv)
#conv.setWeight(np.array([[1,1,1,1,-8,1,1,1,1],[1,0,1,0,-4,0,1,0,1],[0,1,0,1,-4,1,0,1,0],[0,0,0,0,1,0,0,0,0]]))
inputLayer.setInputBatch(sample)
for i in range(0, 10000):
outputLayer.calculate()
if i % 1000 == 0:
outputM.showPictureFromBatch(conv.getOutput(), (28 * 4, 28))
print(conv.getWeight())
outputLayer.update(trainingData.getAnswerVecotrBatch(5))
import ioex
import layer
import util
print("### task1 ###")
# 入出力準備
inputM = ioex.InputManager()
outputM = ioex.OutputManager()
testingData = inputM.getMnistTestingData()
# ３層ニューラルネットワーク
inputLayer = layer.InputLayer(28 * 28)
hiddenLayer = layer.HiddenLayer(50, inputLayer, 1, 1)
outputLayer = layer.OutputLayer(10, hiddenLayer, 1, 1)
# ニューラルネットワークの設定
#outputLayer.setActivator(util.correctedSoftmax)
#入出力
loop = True
while (loop):
# 入力
targetNum = inputM.selectNumber()
sample = testingData.getSingleData(targetNum)
inputLayer.setInput(sample.getImage())
# 出力
result = outputLayer.calculate()
outputM.printMaxLikelihood(result)
print(result)
def _set_layers(self):
last_layer = self.depth - 1
self.layers.append(layer.InputLayer(self.form[0]))
for i in range(1, last_layer):
self.layers.append(layer.HiddenLayer(self.form[i], [], []))
self.layers.append(layer.OutputLayer(self.form[last_layer], [], []))
import sys import time print "Initialising the training data (golden standard)" netInputVector = [[0.05, 0.1], [0.1, 0.2]] netOutputVector = [[0.01, 0.99], [0.04, 0.99]] print "Setting up the net" net = net.Net() inputLayer = layer.InputLayer(2, net, neuron.SigmoidNeuron) net.setInputLayer(inputLayer) hiddenLayer1 = layer.HiddenLayer(3, net, neuron.SigmoidNeuron) net.addHiddenLayer(hiddenLayer1) hiddenLayer2 = layer.HiddenLayer(5, net, neuron.SigmoidNeuron) net.addHiddenLayer(hiddenLayer2) outputLayer = layer.OutputLayer(2, net, neuron.SigmoidNeuron) net.setOutputLayer(outputLayer) print "Training the net" #for index in range(0, len(netInputVector)): # net.inputLayer.setActivationVector(netInputVector[index]) # net.perform() # error = net.outputLayer.error(netOutputVector[index]) # epoch = 1 # while error > 0.01: # net.backpropagate(netOutputVector[index]) # net.perform() # error = net.outputLayer.error(netOutputVector[index]) # epoch += 1 # print net # print "Training epochs: " + str(epoch)