def test_optimizer_override(self):
ls = MS.GradientDescent(lr=0.5)
cost = MC.NegativeLogLikelihood()
inp = ML.Input(1, 'inp')
h = ML.Hidden(5,
activation=MA.Tanh(),
learningScenari=[MS.Fixed("b")],
name="h")
o = ML.SoftmaxClassifier(
2,
learningScenari=[MS.GradientDescent(lr=0.5),
MS.Fixed("W")],
cost=cost,
name="out")
net = inp > h > o
net.init()
ow = o.getP('W').getValue()
ob = o.getP('b').getValue()
hw = h.getP('W').getValue()
hb = h.getP('b').getValue()
for x in xrange(1, 10):
net["out"].train({
"inp.inputs": [[1]],
"out.targets": [1]
})["out.drive.train"]
self.assertTrue(sum(ow[0]) == sum(o.getP('W').getValue()[0]))
self.assertTrue(sum(ob) != sum(o.getP('b').getValue()))
self.assertTrue(sum(hb) == sum(h.getP('b').getValue()))
self.assertTrue(sum(hw[0]) != sum(h.getP('W').getValue()[0]))
def testRecurrences(self):
import Mariana.recurrence as MREC
import Mariana.reshaping as MRES
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
# for clas in [MREC.RecurrentDense, MREC.LSTM, MREC.GRU] :
for clas in [MREC.RecurrentDense]:
inp = ML.Input((None, 3), 'inp')
r = clas(2, onlyReturnFinal=True, name="rec")
reshape = MRES.Reshape((-1, 2), name="reshape")
o = ML.SoftmaxClassifier(
2,
cost=cost,
learningScenari=[MS.GradientDescent(lr=0.5)],
name="out")
net = inp > r > reshape > o
net.init()
inputs = [[[1, 1], [1, 0], [1, 1]], [[1, 0], [0, 1], [1, 0]]]
oldWih = r.getP("W_in_to_hid").getValue()
oldWhh = r.getP("W_hid_to_hid").getValue()
for x in xrange(1, 100):
net["out"].train({
"inp.inputs": inputs,
"out.targets": [1, 1, 1]
})["out.drive.train"]
self.assertTrue(
oldWih.mean() != r.getP("W_in_to_hid").getValue().mean())
self.assertTrue(
oldWhh.mean() != r.getP("W_hid_to_hid").getValue().mean())
def test_multiinputs(self):
ls = MS.GradientDescent(lr=0.1)
inpA = ML.Embedding(2, 2, 2, name="IA")
inpB = ML.Input(2, name="IB")
inpNexus = ML.Composite(name="InputNexus")
h1 = ML.Hidden(32,
activation=MA.ReLU(),
decorators=[],
regularizations=[],
name="Fully-connected1")
o = ML.Regression(2,
decorators=[],
activation=MA.ReLU(),
learningScenario=ls,
costObject=MC.CrossEntropy(),
name="Out",
regularizations=[])
inpA > inpNexus
inpB > inpNexus
m = inpNexus > h1 > o
m.init()
def test_save_load_pickle(self) :
import os
import Mariana.network as MN
ls = MS.GradientDescent(lr = 0.1)
cost = MC.NegativeLogLikelihood()
i = ML.Input(2, 'inp')
h = Hidden_layerRef(i, 10, activation = MA.ReLU(), name = "Hidden_0.500705866892")
o = ML.SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
mlp = i > h > o
self.xor_ins = numpy.array(self.xor_ins)
self.xor_outs = numpy.array(self.xor_outs)
for i in xrange(1000) :
mlp.train(o, inp = self.xor_ins, targets = self.xor_outs )
mlp.save("test_save")
mlp2 = MN.loadModel("test_save.mar.mdl.pkl")
o = mlp.outputs.values()[0]
v1 = mlp.propagate( o.name, inp = self.xor_ins )["outputs"]
v2 = mlp2.propagate( o.name, inp = self.xor_ins )["outputs"]
self.assertEqual(numpy.sum(v1), numpy.sum(v2))
self.assertEqual(mlp["Hidden_0.500705866892"].otherLayer.name, mlp2["Hidden_0.500705866892"].otherLayer.name)
os.remove('test_save.mar.mdl.pkl')
def test_embedding(self):
"""the first 3 and the last 3 should be diametrically opposed"""
data = [[0], [1], [2], [3], [4], [5]]
targets = [0, 0, 0, 1, 1, 1]
ls = MS.GradientDescent(lr=0.5)
cost = MC.NegativeLogLikelihood()
emb = ML.Embedding(1, 2, len(data), learningScenario=ls, name="emb")
o = ML.SoftmaxClassifier(2,
learningScenario=MS.Fixed(),
costObject=cost,
name="out")
net = emb > o
miniBatchSize = 2
for i in xrange(2000):
for i in xrange(0, len(data), miniBatchSize):
net.train(o,
emb=data[i:i + miniBatchSize],
targets=targets[i:i + miniBatchSize])
embeddings = emb.getEmbeddings()
for i in xrange(0, len(data) / 2):
v = numpy.dot(embeddings[i], embeddings[i + len(data) / 2])
self.assertTrue(v < -1)
def getModel(inpSize, filterWidth) :
ls = MS.GradientDescent(lr = 0.5)
cost = MC.NegativeLogLikelihood()
pooler = MCONV.MaxPooling2D(1, 2)
i = ML.Input(inpSize, name = 'inp')
ichan = MCONV.InputChanneler(1, inpSize, name = 'inpChan')
c1 = MCONV.Convolution2D(
nbFilters = 5,
filterHeight = 1,
filterWidth = filterWidth,
activation = MA.ReLU(),
pooler = pooler,
name = "conv1"
)
c2 = MCONV.Convolution2D(
nbFilters = 10,
filterHeight = 1,
filterWidth = filterWidth,
activation = MA.ReLU(),
pooler = pooler,
name = "conv2"
)
f = MCONV.Flatten(name = "flat")
h = ML.Hidden(5, activation = MA.ReLU(), decorators = [], regularizations = [ ], name = "hid" )
o = ML.SoftmaxClassifier(2, decorators = [], learningScenario = ls, costObject = cost, name = "out", regularizations = [ ] )
model = i > ichan > c1 > c2 > f > h > o
return model
def test_ae(self):
data = []
for i in xrange(8):
zeros = numpy.zeros(8)
zeros[i] = 1
data.append(zeros)
ls = MS.GradientDescent(lr=0.1)
cost = MC.MeanSquaredError()
i = ML.Input(8, name='inp')
h = ML.Hidden(3, activation=MA.ReLU(), name="hid")
o = ML.Regression(8,
activation=MA.ReLU(),
learningScenario=ls,
costObject=cost,
name="out")
ae = i > h > o
miniBatchSize = 2
for e in xrange(2000):
for i in xrange(0, len(data), miniBatchSize):
ae.train(o,
inp=data[i:i + miniBatchSize],
targets=data[i:i + miniBatchSize])
res = ae.propagate(o, inp=data)["outputs"]
for i in xrange(len(res)):
self.assertEqual(numpy.argmax(data[i]), numpy.argmax(res[i]))
def test_save_load_64h(self):
import os
import Mariana.network as MN
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
i = ML.Input(2, 'inp')
o = ML.SoftmaxClassifier(nbClasses=2,
cost=cost,
learningScenari=[ls],
name="out")
prev = i
for i in xrange(64):
h = ML.Hidden(size=10, activation=MA.ReLU(), name="Hidden_%s" % i)
prev > h
prev = h
mlp = prev > o
mlp.init()
mlp.save("test_save")
mlp2 = MN.loadModel("test_save.mar")
mlp2.init()
v1 = mlp["out"].propagate["test"]({
"inp.inputs": self.xor_ins
})["out.propagate.test"]
v2 = mlp2["out"].propagate["test"]({
"inp.inputs": self.xor_ins
})["out.propagate.test"]
self.assertTrue((v1 == v2).all())
os.remove('test_save.mar')
def getModel(inpSize, filterWidth):
ls = MS.GradientDescent(lr=0.5)
cost = MC.NegativeLogLikelihood()
i = ML.Input((1, 1, inpSize), name='inp')
c1 = MCONV.Convolution2D(numFilters=5,
filterHeight=1,
filterWidth=filterWidth,
activation=MA.ReLU(),
name="conv1")
pool1 = MSAMP.MaxPooling2D(poolHeight=1, poolWidth=2, name="pool1")
c2 = MCONV.Convolution2D(numFilters=10,
filterHeight=1,
filterWidth=filterWidth,
activation=MA.ReLU(),
name="conv2")
pool2 = MSAMP.MaxPooling2D(poolHeight=1, poolWidth=2, name="pool2")
h = ML.Hidden(5, activation=MA.ReLU(), name="hid")
o = ML.SoftmaxClassifier(nbClasses=2,
cost=cost,
learningScenari=[ls],
name="out")
model = i > c1 > pool1 > c2 > pool2 > h > o
model.init()
return model
def test_embedding(self):
"""the first 3 and the last 3 should be diametrically opposed"""
data = [[0], [1], [2], [3], [4], [5]]
targets = [0, 0, 0, 1, 1, 1]
ls = MS.GradientDescent(lr=0.5)
cost = MC.NegativeLogLikelihood()
inp = ML.Input(1, 'inp')
emb = ML.Embedding(nbDimensions=2,
dictSize=len(data),
learningScenari=[ls],
name="emb")
o = ML.SoftmaxClassifier(2,
learningScenari=[MS.Fixed()],
cost=cost,
name="out")
net = inp > emb > o
net.init()
miniBatchSize = 2
for i in xrange(2000):
for i in xrange(0, len(data), miniBatchSize):
net["out"].train({
"inp.inputs": data[i:i + miniBatchSize],
"out.targets": targets[i:i + miniBatchSize]
})["out.drive.train"]
embeddings = emb.getP("embeddings").getValue()
for i in xrange(0, len(data) / 2):
v = numpy.dot(embeddings[i], embeddings[i + len(data) / 2])
self.assertTrue(v < -1)
def ae1(data):
'''Using a regression layer. This layer needs an explicit target'''
miniBatchSize = 2
ls = MS.GradientDescent(lr=0.1)
cost = MC.MeanSquaredError()
i = ML.Input(8, name='inp')
h = ML.Hidden(3,
activation=MA.ReLU(),
initializations=[MI.SmallUniformWeights(),
MI.ZeroBias()],
name="hid")
o = ML.Regression(
8,
activation=MA.ReLU(),
initializations=[MI.SmallUniformWeights(),
MI.ZeroBias()],
learningScenario=ls,
costObject=cost,
name="out")
ae = i > h > o
for e in xrange(1000):
for i in xrange(0, len(data), miniBatchSize):
ae.train(o,
inp=data[i:i + miniBatchSize],
targets=data[i:i + miniBatchSize])
return ae, o
def test_concatenation(self):
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
inp = ML.Input(2, 'inp')
h1 = ML.Hidden(5, activation=MA.Tanh(), name="h1")
h2 = ML.Hidden(5, activation=MA.Tanh(), name="h2")
o = ML.SoftmaxClassifier(nbClasses=2,
cost=cost,
learningScenari=[ls],
name="out")
inp > h1
inp > h2
c = ML.C([h1, h2], name="concat")
mlp = c > o
mlp.init()
self.assertEqual(c.getIntrinsicShape()[0],
h1.getIntrinsicShape()[0] + h2.getIntrinsicShape()[0])
for i in xrange(10000):
ii = i % len(self.xor_ins)
miniBatch = [self.xor_ins[ii]]
targets = [self.xor_outs[ii]]
mlp["out"].train({
"inp.inputs": miniBatch,
"out.targets": targets
})["out.drive.train"]
for i in xrange(len(self.xor_ins)):
self.assertEqual(
mlp["out"].predict["test"]({
"inp.inputs": [self.xor_ins[i]]
})["out.predict.test"], self.xor_outs[i])
def ae2(data):
"""This one uses an Autoencode layer. This layer is a part of the graph and does not need a specific traget"""
miniBatchSize = 1
ls = MS.GradientDescent(lr=0.1)
cost = MC.MeanSquaredError()
i = ML.Input(8, name='inp')
h = ML.Hidden(3,
activation=MA.ReLU(),
initializations=[MI.SmallUniformWeights(),
MI.ZeroBias()],
name="hid")
o = ML.Autoencode(
i.name,
activation=MA.ReLU(),
initializations=[MI.SmallUniformWeights(),
MI.ZeroBias()],
learningScenario=ls,
costObject=cost,
name="out")
ae = i > h > o
# ae.init()
# o.train.printGraph()
for e in xrange(2000):
for i in xrange(0, len(data), miniBatchSize):
ae.train(o, inp=data[i:i + miniBatchSize])
return ae, o
def test_composite(self):
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
inp = ML.Input(2, 'inp')
h1 = ML.Hidden(5, activation=MA.Tanh(), name="h1")
h2 = ML.Hidden(5, activation=MA.Tanh(), name="h2")
o = ML.SoftmaxClassifier(2,
learningScenario=ls,
costObject=cost,
name="out")
c = ML.Composite(name="Comp")
inp > h1 > c
inp > h2 > c
mlp = c > o
for i in xrange(10000):
ii = i % len(self.xor_ins)
mlp.train(o, inp=[self.xor_ins[ii]], targets=[self.xor_outs[ii]])
self.assertEqual(mlp.predict(o, inp=[self.xor_ins[0]])["class"], 0)
self.assertEqual(mlp.predict(o, inp=[self.xor_ins[1]])["class"], 1)
self.assertEqual(mlp.predict(o, inp=[self.xor_ins[2]])["class"], 1)
self.assertEqual(mlp.predict(o, inp=[self.xor_ins[3]])["class"], 0)
def test_ae(self) :
data = []
for i in xrange(8) :
zeros = numpy.zeros(8)
zeros[i] = 1
data.append(zeros)
ls = MS.GradientDescent(lr = 0.1)
cost = MC.MeanSquaredError()
i = ML.Input(8, name = 'inp')
h = ML.Hidden(3, activation = MA.ReLU(), initializations=[MI.SmallUniformWeights(), MI.ZeroBias()], name = "hid")
o = ML.Autoencode(targetLayerName = "inp", activation = MA.ReLU(), initializations=[MI.SmallUniformWeights(), MI.ZeroBias()], learningScenario = ls, costObject = cost, name = "out" )
ae = i > h > o
miniBatchSize = 1
for e in xrange(2000) :
for i in xrange(0, len(data), miniBatchSize) :
ae.train(o, inp = data[i:i+miniBatchSize])
res = ae.propagate(o, inp = data)["outputs"]
for i in xrange(len(res)) :
self.assertEqual( numpy.argmax(data[i]), numpy.argmax(res[i]))
def getMLP(self, nbInputs=2, nbClasses=2):
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
i = ML.Input(nbInputs, 'inp')
h = ML.Hidden(size=6, activation=MA.ReLU(), name="Hidden_0.500705866892")
o = ML.SoftmaxClassifier(nbClasses=nbClasses,
cost=cost,
learningScenari=[ls],
name="out")
mlp = i > h > o
mlp.init()
return mlp
def test_ae_reg(self):
powerOf2 = 3
nbUnits = 2**powerOf2
data = []
for i in xrange(nbUnits):
zeros = numpy.zeros(nbUnits)
zeros[i] = 1
data.append(zeros)
ls = MS.GradientDescent(lr=0.1)
cost = MC.MeanSquaredError()
i = ML.Input(nbUnits, name='inp')
h = ML.Hidden(powerOf2,
activation=MA.ReLU(),
initializations=[
MI.Uniform('W', small=True),
MI.SingleValue('b', 0)
],
name="hid")
o = ML.Regression(nbUnits,
activation=MA.ReLU(),
initializations=[
MI.Uniform('W', small=True),
MI.SingleValue('b', 0)
],
learningScenari=[ls],
cost=cost,
name="out")
ae = i > h > o
ae.init()
miniBatchSize = 1
for e in xrange(2000):
for i in xrange(0, len(data), miniBatchSize):
miniBatch = data[i:i + miniBatchSize]
ae["out"].train({
"inp.inputs": miniBatch,
"out.targets": miniBatch
})["out.drive.train"]
res = ae["out"].propagate["test"]({
"inp.inputs": data
})["out.propagate.test"]
for i in xrange(len(res)):
self.assertEqual(numpy.argmax(data[i]), numpy.argmax(res[i]))
def trainMLP_xor(self) :
ls = MS.GradientDescent(lr = 0.1)
cost = MC.NegativeLogLikelihood()
i = ML.Input(2, 'inp')
h = ML.Hidden(10, activation = MA.ReLU(), name = "Hidden_0.500705866892")
o = ML.SoftmaxClassifier(2, learningScenario = ls, costObject = cost, name = "out")
mlp = i > h > o
self.xor_ins = numpy.array(self.xor_ins)
self.xor_outs = numpy.array(self.xor_outs)
for i in xrange(1000) :
mlp.train(o, inp = self.xor_ins, targets = self.xor_outs )
return mlp
def test_merge(self):
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
inp1 = ML.Input(1, 'inp1')
inp2 = ML.Input(1, 'inp2')
merge = ML.M((inp1 + inp2) / 3 * 10 - 1, name="merge")
inp1 > merge
mdl = inp2 > merge
mdl.init()
self.assertEqual(merge.getIntrinsicShape(), inp1.getIntrinsicShape())
v = mdl["merge"].propagate["test"]({
"inp1.inputs": [[1]],
"inp2.inputs": [[8]]
})["merge.propagate.test"]
self.assertEqual(v, 29)
def trainMLP_xor(self):
ls = MS.GradientDescent(lr=0.1)
cost = MC.NegativeLogLikelihood()
i = ML.Input(2, 'inp')
h = ML.Hidden(4,
activation=MA.Tanh(),
decorators=[dec.GlorotTanhInit()],
regularizations=[MR.L1(0), MR.L2(0)])
o = ML.SoftmaxClassifier(2,
learningScenario=ls,
costObject=cost,
name="out")
mlp = i > h > o
self.xor_ins = numpy.array(self.xor_ins)
self.xor_outs = numpy.array(self.xor_outs)
for i in xrange(1000):
ii = i % len(self.xor_ins)
mlp.train(o, inp=[self.xor_ins[ii]], targets=[self.xor_outs[ii]])
return mlp
But Mariana style.
This version uses a trainer/dataset mapper setup:
* automatically saves the best model for each set (train, test, validation)
* automatically saves the model if the training halts because of an error or if the process is killed
* saves a log if the process dies unexpectedly
* training results and hyper parameters values are recorded to a file
* allows you to define custom stop criteria
* training info is printed at each epoch, including best scores and at which epoch they were achieved
"""
if __name__ == "__main__":
# Let's define the network
ls = MS.GradientDescent(lr=0.01)
cost = MC.NegativeLogLikelihood()
i = ML.Input(28 * 28, name='inp')
h = ML.Hidden(500, activation=MA.Tanh(), decorators=[MD.GlorotTanhInit()], regularizations=[MR.L1(0), MR.L2(0.0001)], name="hid")
o = ML.SoftmaxClassifier(10, learningScenario=ls, costObject=cost, name="out", regularizations=[MR.L1(0), MR.L2(0.0001)])
mlp = i > h > o
mlp.saveDOT("mnist_mlp")
mlp.saveHTML("mnist_mlp")
# And then map sets to the inputs and outputs of our network
train_set, validation_set, test_set = load_mnist()
trainData = MDM.Series(images=train_set[0], numbers=train_set[1])
trainMaps = MDM.DatasetMapper()
def test_multiout_fctmixin(self):
i = ML.Input(1, name='inp')
o1 = ML.Autoencode(targetLayer=i,
activation=MA.Tanh(),
learningScenari=[MS.GradientDescent(lr=0.1)],
cost=MC.MeanSquaredError(),
name="out1")
o2 = ML.Regression(1,
activation=MA.Tanh(),
learningScenari=[MS.GradientDescent(lr=0.2)],
cost=MC.MeanSquaredError(),
name="out2")
i > o1
ae = i > o2
ae.init()
preOut1 = ae["out1"].test({"inp.inputs": [[1]]})["out1.drive.test"]
preOut2 = ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"]
ae["out1"].train({"inp.inputs": [[1]]})["out1.drive.train"]
self.assertTrue(
preOut1 > ae["out1"].test({"inp.inputs": [[1]]})["out1.drive.test"]
)
self.assertTrue(preOut2 == ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"])
preOut1 = ae["out1"].test({"inp.inputs": [[1]]})["out1.drive.test"]
preOut2 = ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"]
ae["out2"].train({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.train"]
self.assertTrue(preOut1 == ae["out1"].test({"inp.inputs": [[1]]})
["out1.drive.test"])
self.assertTrue(preOut2 > ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"])
preOut1 = ae["out1"].test({"inp.inputs": [[1]]})["out1.drive.test"]
preOut2 = ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"]
fct = ae["out1"].train + ae["out2"].train + ae["inp"].propagate["train"]
ret = fct({"inp.inputs": [[1]], "out2.targets": [[1]]})
self.assertEqual(len(ret), 3)
self.assertTrue(
preOut1 > ae["out1"].test({"inp.inputs": [[1]]})["out1.drive.test"]
)
self.assertTrue(preOut2 > ae["out2"].test({
"inp.inputs": [[1]],
"out2.targets": [[1]]
})["out2.drive.test"])
#return model
def train(self, inputs, targets):
#because of the channeler there is no need to reshape the data besfore passing them to the conv layer
return self.model.train("out", inp=inputs, targets=targets)
def test(self, inputs, targets):
return self.model.test("out", inp=inputs, targets=targets)
def saveHTML(self, name):
return self.model.saveHTML(name)
if __name__ == "__main__":
ls = MS.GradientDescent(lr=1e-3)
cost = MC.NegativeLogLikelihood()
maxEpochs = 200
miniBatchSize = 10
model = ConvWithChanneler(ls, cost)
model.saveHTML('cnn_catdog')
datafile = "small_set"
train_set = load_data(datafile)
#shuffling the dataset
indices = [i for i in xrange(len(train_set[0]))]
numpy.random.shuffle(indices)
train_set = (train_set[0][indices], train_set[1][indices])