def plot_host_usage(self, sample_size=100):
counter = {}
for h, durations in self.__hosts.items():
for start, end in durations:
start, end = floor(start, sample_size), ceil(end, sample_size)
cur_end = min(start + sample_size, end)
while cur_end < end:
timerange = (cur_end - sample_size, cur_end)
if h not in counter.get(timerange, set()):
counter.setdefault(timerange, set()).add(h)
cur_end += sample_size
x = list(sorted(counter.keys()))
y = [len(counter[k]) for k in x]
return x, y
def test_ceil():
assert_equal(ceil(2.5), 3)
def rbm(data, numHid, modelType = "BB", **kwargs) :
"""
rbm defination
data : when type is BB, should be binary, or in [0,1] to be interpreted as probabilities
when type is GB, should be continuous real value. data should have a format of *.npy
numHid : number nodes of and hidden layer
type : rbm type, can be set as BB or GB
additional inputs (specified as name value pairs or in struct)
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
penalty weight decay factor
batchsize The number of training instances per batch
verbose For printing progress
anneal Flag. If set true, the penalty is annealed linearly
through epochs to 10% of its original value
OUTPUTS:
model.type Type of RBM (i.e. type of its visible and hidden units)
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.top The activity of the top layer, to be used when training
DBN's
errors The errors in reconstruction at every epoch
"""
arg = util.processOptions(kwargs, \
method = "CD", \
eta = 0.1, \
momentum = 0.9,\
maxEpoch = 50, \
avgLast = 0, \
penalty = 0, \
batchSize = 50, \
verbose = True, \
anneal = False)
[method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose, anneal] = [\
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"],\
arg["anneal"]
]
# from which step, we start to compute the average
avgStart = maxEpoch - avgLast
# for weight decay use
oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(data.shape)
if verbose :
print "processing data"
numVis = numDims
numBatch = util.ceil(numCases,batchSize)
# shuffle the data
np.random.shuffle(data)
# init CUDA
# cm.cuda_set_device()
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
# init weights
weight = cm.CUDAMatrix(0.1*np.random.randn(numVis,numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis,numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1,numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1,numHid)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidActP2 = cm.empty((batchSize, numHid))
visState = cm.empty((batchSize,numVis))
hidState = cm.empty((batchSize, numHid))
t = 1
for epoch in range(maxEpoch) :
error = []
if anneal :
# apply linear weight decay
penalty = oldPenalty - 0.9 *epoch/maxEpoch*oldPenalty
for batch in range(numBatch) :
# train each data batch
if batchSize*(batch+1) > numCases :
visTrue = deviceData.get_row_slice(batchSize*batch, numCases)
batchSize = visTrue.shape[0]
else :
visTrue = deviceData.get_row_slice(batchSize*batch, batchSize*(batch+1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
# positive phase
cm.dot(visActP, weight, target = hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidState)
if cmp(method, "SML") == 0 :
if np.logical_and(np.equal(epoch,1), np.equal(batch,1)) :
pass # here does not need in practical use
elif cmp(method, "CD") == 0 :
pass
# negetive phase
if cmp(modelType, "BB") == 0 :
cm.dot(hidState, weight.transpose(), target = visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
visState.fill_with_rand()
visState.less_than(visActP, target = visState)
elif cmp(modelType, "GB") == 0 :
cm.dot(hidState, weight.transpose(), target = visActP)
visActP.add_row_vec(biasV)
visActP.add(np.random.randn(batchSize, numVis),target=visState)
# another positive phase
cm.dot(visState, weight, target = hidActP2)
hidActP2.add_row_vec(biasH)
hidActP2.apply_sigmoid()
hidState.fill_with_rand()
hidState.less_than(hidActP2, target=hidState)
#update weight and bias
dWeight = cm.dot(visTrue.transpose(), hidActP)
dWeight.subtract_dot(visState.transpose(), hidActP2)
dBiasV = visTrue.sum(axis = 0).subtract(visState.sum(axis = 0))
dBiasH = hidActP.sum(axis=0).subtract(hidActP2.sum(axis = 0))
dWeight.divide(batchSize).subtract(weight.mult(penalty))
dBiasV.divide(batchSize)
dBiasH.divide(batchSize)
weightInc.mult(momentum).add_mult(dWeight, eta)
biasVInc.mult(momentum).add_mult(dBiasV, eta)
biasHInc.mult(momentum).add_mult(dBiasH, eta)
weight.add(weightInc)
biasV.add(biasVInc)
biasH.add(biasHInc)
if epoch > avgStart :
# apply average
weightAgv.subtract(weightAgv.subtract(weight).mult(1.0/t))
biasVAgv.subtract(biasVAgv.subtract(biasV).mult(1.0/t))
biasHAgv.subtract(biasHAgv.subtract(biasH).mult(1.0/t))
t = t+1
else :
weightAgv = weight
biasVAgv = biasV
biasHAgv = biasH
# reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm() ** 2)
if verbose :
print "epoch %d/%d. Reconstruction error is %f " % (epoch+1, maxEpoch, sum(error))
# save rbm model
top = cm.CUDAMatrix(np.zeros((numCases, numHid)))
cm.dot(deviceData, weightAgv, target = top)
top.add_row_vec(biasHAgv)
top.apply_sigmoid()
model_ = m.rbmModel(weightAgv,biasVAgv,biasHAgv,type = modelType,top = top)
cm.shutdown()
return model_
def rbmFit(X, numHid, y, isSaveModel=False, name=None, **kwargs) :
"""
X ... data. should be binary, or in [0,1] interpreted as
... probabilities
numhid ... number of hidden units
y ... List of discrete labels
nClass number of classes
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
batchsize The number of training instances per batch
verbose For printing progress
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.weightlabel ... The weights on labels layer
model.biasLabel ... The biases on labels layer
errors The errors in reconstruction at each epoch
"""
arg = util.processOptions(kwargs, \
nClass = np.unique(y).size, \
method = "CD", \
eta = 0.1, \
momentum = 0.5,\
maxEpoch = 500, \
avgLast = 0, \
penalty = 0, \
batchSize = 100, \
verbose = True)
[nClass, method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose] = [\
arg["nClass"],\
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"]
]
if verbose :
print "Processing data ..."
# from which step, we start to compute the average
# avgStart = maxEpoch - avgLast
# for weight decay use
# oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(X.shape)
numVis = numDims
uniqueLabel = np.unique(y)
numBatch = util.ceil(numCases, batchSize)
y = util.matrixLabel(y)
# shuffle data and label
data = copy.deepcopy(X)
[data, label] = util.shuffle(data, y)
# init CUDA
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
deviceLabel = cm.CUDAMatrix(cm.reformat(label))
# init weights
weight = cm.CUDAMatrix(0.1*np.random.randn(numVis,numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
weightLabel = cm.CUDAMatrix(0.1*np.random.randn(nClass, numHid))
biasLabel = cm.CUDAMatrix(np.zeros((1,nClass)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis,numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1,numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1,numHid)))
weightLabelInc = cm.CUDAMatrix(np.zeros((nClass, numHid)))
biasLabelInc = cm.CUDAMatrix(np.zeros((1,nClass)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
for epoch in range(maxEpoch) :
error = []
for batch in range(numBatch) :
# train each data batch
if batchSize*(batch+1) > numCases :
visTrue = deviceData.get_row_slice(batchSize*batch, numCases)
labelTrue = deviceLabel.get_row_slice(batchSize*batch, numCases)
batchSize = visTrue.shape[0]
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
else :
visTrue = deviceData.get_row_slice(batchSize*batch, batchSize*(batch+1))
labelTrue = deviceLabel.get_row_slice(batchSize*batch, batchSize*(batch+1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
#apply momentum
weightInc.mult(momentum)
biasVInc.mult(momentum)
biasHInc.mult(momentum)
weightLabel.mult(momentum)
biasLabel.mult(momentum)
# positive phase
cm.dot(visActP, weight, target = hidActP)
hidActP.add_dot(labelTrue, weightLabel)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.add_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0)
biasHInc.add_sums(hidActP, axis=0)
weightLabelInc.add_dot(labelTrue.T, hidActP)
biasLabelInc.add_sums(labelTrue, axis=0)
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidActP)
if cmp(method, "SML") == 0 :
if np.logical_and(np.equal(epoch,1), np.equal(batch,1)) :
pass # here does not need in practical use
elif cmp(method, "CD") == 0 :
pass
# negative phase
cm.dot(hidActP, weight.T, target = visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
cm.dot(hidActP, weightLabel.T, target = labelTrue)
labelTrue.add_row_vec(biasLabel)
labelTrue = util.softmax(labelTrue)
# another positive phase
cm.dot(visActP, weight, target = hidActP)
hidActP.add_dot(labelTrue, weightLabel)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.subtract_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0, mult=-1)
biasHInc.add_sums(hidActP, axis=0, mult=-1)
weightLabelInc.subtract_dot(labelTrue.T, hidActP)
biasLabelInc.add_sums(labelTrue, axis=0, mult=-1)
# update weights and bias
weight.add_mult(weightInc, eta/batchSize)
biasV.add_mult(biasVInc, eta/batchSize)
biasH.add_mult(biasHInc, eta/batchSize)
weightLabel.add_mult(weightLabelInc, eta/batchSize)
biasLabel.add_mult(biasLabelInc, eta/batchSize)
# calculate reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm()**2)
# free memory
visTrue.free_device_memory()
labelTrue.free_device_memory()
if verbose :
print "Epoch %d/%d, reconstruction error is %f " % (epoch+1, maxEpoch, sum(error))
# save rbm model
weight.copy_to_host()
biasV.copy_to_host()
biasH.copy_to_host()
weightLabel.copy_to_host()
biasLabel.copy_to_host()
model_ = m.rbmModel(weight.numpy_array, biasV.numpy_array, biasH.numpy_array, \
weightLabel = weightLabel.numpy_array,\
biasLabel = biasLabel.numpy_array, labels = uniqueLabel)
# free device memory
deviceData.free_device_memory()
deviceLabel.free_device_memory()
weight.free_device_memory()
biasV.free_device_memory()
biasH.free_device_memory()
weightLabel.free_device_memory()
biasLabel.free_device_memory()
weightInc.free_device_memory()
biasVInc.free_device_memory()
biasHInc.free_device_memory()
weightLabelInc.free_device_memory()
biasLabelInc.free_device_memory()
hidActP.free_device_memory()
visActP.free_device_memory()
hidState.free_device_memory()
cm.shutdown()
if isSaveModel :
modelList = []
modelList.append(model_)
model = np.array(modelList)
np.save(name,model)
return model_
def rbmFit(X, numHid, y, isSaveModel=False, name=None, **kwargs):
"""
X ... data. should be binary, or in [0,1] interpreted as
... probabilities
numhid ... number of hidden units
y ... List of discrete labels
nClass number of classes
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
batchsize The number of training instances per batch
verbose For printing progress
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.weightlabel ... The weights on labels layer
model.biasLabel ... The biases on labels layer
errors The errors in reconstruction at each epoch
"""
arg = util.processOptions(kwargs, \
nClass = np.unique(y).size, \
method = "CD", \
eta = 0.1, \
momentum = 0.5,\
maxEpoch = 500, \
avgLast = 0, \
penalty = 0, \
batchSize = 100, \
verbose = True)
[nClass, method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose] = [\
arg["nClass"],\
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"]
]
if verbose:
print "Processing data ..."
# from which step, we start to compute the average
# avgStart = maxEpoch - avgLast
# for weight decay use
# oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(X.shape)
numVis = numDims
uniqueLabel = np.unique(y)
numBatch = util.ceil(numCases, batchSize)
y = util.matrixLabel(y)
# shuffle data and label
data = copy.deepcopy(X)
[data, label] = util.shuffle(data, y)
# init CUDA
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
deviceLabel = cm.CUDAMatrix(cm.reformat(label))
# init weights
weight = cm.CUDAMatrix(0.1 * np.random.randn(numVis, numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
weightLabel = cm.CUDAMatrix(0.1 * np.random.randn(nClass, numHid))
biasLabel = cm.CUDAMatrix(np.zeros((1, nClass)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis, numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1, numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1, numHid)))
weightLabelInc = cm.CUDAMatrix(np.zeros((nClass, numHid)))
biasLabelInc = cm.CUDAMatrix(np.zeros((1, nClass)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
for epoch in range(maxEpoch):
error = []
for batch in range(numBatch):
# train each data batch
if batchSize * (batch + 1) > numCases:
visTrue = deviceData.get_row_slice(batchSize * batch, numCases)
labelTrue = deviceLabel.get_row_slice(batchSize * batch,
numCases)
batchSize = visTrue.shape[0]
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
else:
visTrue = deviceData.get_row_slice(batchSize * batch,
batchSize * (batch + 1))
labelTrue = deviceLabel.get_row_slice(batchSize * batch,
batchSize * (batch + 1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
#apply momentum
weightInc.mult(momentum)
biasVInc.mult(momentum)
biasHInc.mult(momentum)
weightLabel.mult(momentum)
biasLabel.mult(momentum)
# positive phase
cm.dot(visActP, weight, target=hidActP)
hidActP.add_dot(labelTrue, weightLabel)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.add_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0)
biasHInc.add_sums(hidActP, axis=0)
weightLabelInc.add_dot(labelTrue.T, hidActP)
biasLabelInc.add_sums(labelTrue, axis=0)
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidActP)
if cmp(method, "SML") == 0:
if np.logical_and(np.equal(epoch, 1), np.equal(batch, 1)):
pass # here does not need in practical use
elif cmp(method, "CD") == 0:
pass
# negative phase
cm.dot(hidActP, weight.T, target=visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
cm.dot(hidActP, weightLabel.T, target=labelTrue)
labelTrue.add_row_vec(biasLabel)
labelTrue = util.softmax(labelTrue)
# another positive phase
cm.dot(visActP, weight, target=hidActP)
hidActP.add_dot(labelTrue, weightLabel)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.subtract_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0, mult=-1)
biasHInc.add_sums(hidActP, axis=0, mult=-1)
weightLabelInc.subtract_dot(labelTrue.T, hidActP)
biasLabelInc.add_sums(labelTrue, axis=0, mult=-1)
# update weights and bias
weight.add_mult(weightInc, eta / batchSize)
biasV.add_mult(biasVInc, eta / batchSize)
biasH.add_mult(biasHInc, eta / batchSize)
weightLabel.add_mult(weightLabelInc, eta / batchSize)
biasLabel.add_mult(biasLabelInc, eta / batchSize)
# calculate reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm()**2)
# free memory
visTrue.free_device_memory()
labelTrue.free_device_memory()
if verbose:
print "Epoch %d/%d, reconstruction error is %f " % (
epoch + 1, maxEpoch, sum(error))
# save rbm model
weight.copy_to_host()
biasV.copy_to_host()
biasH.copy_to_host()
weightLabel.copy_to_host()
biasLabel.copy_to_host()
model_ = m.rbmModel(weight.numpy_array, biasV.numpy_array, biasH.numpy_array, \
weightLabel = weightLabel.numpy_array,\
biasLabel = biasLabel.numpy_array, labels = uniqueLabel)
# free device memory
deviceData.free_device_memory()
deviceLabel.free_device_memory()
weight.free_device_memory()
biasV.free_device_memory()
biasH.free_device_memory()
weightLabel.free_device_memory()
biasLabel.free_device_memory()
weightInc.free_device_memory()
biasVInc.free_device_memory()
biasHInc.free_device_memory()
weightLabelInc.free_device_memory()
biasLabelInc.free_device_memory()
hidActP.free_device_memory()
visActP.free_device_memory()
hidState.free_device_memory()
cm.shutdown()
if isSaveModel:
modelList = []
modelList.append(model_)
model = np.array(modelList)
np.save(name, model)
return model_
def rbm(X, numHid, **kwargs):
"""
rbm defination
data : when type is BB, should be binary, or in [0,1] to be interpreted as probabilities
when type is GB, should be continuous real value. data should have a format of *.npy
numHid : number nodes of and hidden layer
type : rbm type, can be set as BB or GB
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
batchsize The number of training instances per batch
verbose For printing progress
model.type Type of RBM (i.e. type of its visible and hidden units)
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.top The activity of the top layer, to be used when training
DBN's
errors The errors in reconstruction at every epoch
"""
# when compute the transpose of a matrix, using the method *.transpose() is much space consuming. I suggest we can use
# .T atrribute instead
arg = util.processOptions(kwargs, \
modelType = "BB", \
method = "CD", \
eta = 0.1, \
momentum = 0.5,\
maxEpoch = 500, \
avgLast = 0, \
penalty = 0, \
batchSize = 100, \
verbose = True)
[modelType, method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose] = [\
arg["modelType"], \
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"]
]
# from which step, we start to compute the average
# avgStart = maxEpoch - avgLast
# for weight decay use
# oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(X.shape)
if verbose:
print "processing data"
numVis = numDims
numBatch = util.ceil(numCases, batchSize)
# shuffle the data
data = copy.deepcopy(X)
np.random.shuffle(data)
# init CUDA
# cm.cuda_set_device()
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
# init weights
weight = cm.CUDAMatrix(0.1 * np.random.randn(numVis, numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis, numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1, numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1, numHid)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
for epoch in range(maxEpoch):
error = []
for batch in range(numBatch):
# train each data batch
if batchSize * (batch + 1) > numCases:
visTrue = deviceData.get_row_slice(batchSize * batch, numCases)
batchSize = visTrue.shape[0]
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
else:
visTrue = deviceData.get_row_slice(batchSize * batch,
batchSize * (batch + 1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
#apply momentum
weightInc.mult(momentum)
biasVInc.mult(momentum)
biasHInc.mult(momentum)
# positive phase
cm.dot(visActP, weight, target=hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.add_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0)
biasHInc.add_sums(hidActP, axis=0)
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidActP)
if cmp(method, "SML") == 0:
if np.logical_and(np.equal(epoch, 1), np.equal(batch, 1)):
pass # here does not need in practical use
elif cmp(method, "CD") == 0:
pass
# negetive phase
if cmp(modelType, "BB") == 0:
cm.dot(hidActP, weight.T, target=visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
elif cmp(modelType, "GB") == 0:
cm.dot(hidActP, weight.T, target=visActP)
visActP.add_row_vec(biasV)
visActP.add(np.random.randn(batchSize, numVis), target=visActP)
# another positive phase
cm.dot(visActP, weight, target=hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.subtract_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0, mult=-1)
biasHInc.add_sums(hidActP, axis=0, mult=-1)
#update weight and bias
weight.add_mult(weightInc, eta / batchSize)
biasV.add_mult(biasVInc, eta / batchSize)
biasH.add_mult(biasHInc, eta / batchSize)
# if epoch > avgStart :
# # apply average
# weightAgv.subtract(weightAgv.subtract(weight).mult(1.0/t))
# biasVAgv.subtract(biasVAgv.subtract(biasV).mult(1.0/t))
# biasHAgv.subtract(biasHAgv.subtract(biasH).mult(1.0/t))
# t = t+1
# else :
# weightAgv = weight
# biasVAgv = biasV
# biasHAgv = biasH
# reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm()**2)
# free device memory
visTrue.free_device_memory()
if verbose:
print "epoch %d/%d. Reconstruction error is %f " % (
epoch + 1, maxEpoch, sum(error))
# save rbm model
top = cm.CUDAMatrix(np.zeros((numCases, numHid)))
cm.dot(cm.CUDAMatrix(cm.reformat(X)), weight, target=top)
top.add_row_vec(biasH)
top.apply_sigmoid()
weight.copy_to_host()
biasV.copy_to_host()
biasH.copy_to_host()
top.copy_to_host()
model_ = m.rbmModel(weight.numpy_array, biasV.numpy_array, \
biasH.numpy_array, type = modelType, top = top.numpy_array)
# free device memory
deviceData.free_device_memory()
weight.free_device_memory()
biasV.free_device_memory()
biasH.free_device_memory()
weightInc.free_device_memory()
biasVInc.free_device_memory()
biasHInc.free_device_memory()
hidActP.free_device_memory()
visActP.free_device_memory()
hidState.free_device_memory()
cm.shutdown()
return model_
def rbm(X, numHid, **kwargs) :
"""
rbm defination
data : when type is BB, should be binary, or in [0,1] to be interpreted as probabilities
when type is GB, should be continuous real value. data should have a format of *.npy
numHid : number nodes of and hidden layer
type : rbm type, can be set as BB or GB
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
batchsize The number of training instances per batch
verbose For printing progress
model.type Type of RBM (i.e. type of its visible and hidden units)
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.top The activity of the top layer, to be used when training
DBN's
errors The errors in reconstruction at every epoch
"""
# when compute the transpose of a matrix, using the method *.transpose() is much space consuming. I suggest we can use
# .T atrribute instead
arg = util.processOptions(kwargs, \
modelType = "BB", \
method = "CD", \
eta = 0.1, \
momentum = 0.5,\
maxEpoch = 500, \
avgLast = 0, \
penalty = 0, \
batchSize = 100, \
verbose = True)
[modelType, method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose] = [\
arg["modelType"], \
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"]
]
# from which step, we start to compute the average
# avgStart = maxEpoch - avgLast
# for weight decay use
# oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(X.shape)
if verbose :
print "processing data"
numVis = numDims
numBatch = util.ceil(numCases, batchSize)
# shuffle the data
data = copy.deepcopy(X)
np.random.shuffle(data)
# init CUDA
# cm.cuda_set_device()
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
# init weights
weight = cm.CUDAMatrix(0.1*np.random.randn(numVis,numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis,numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1,numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1,numHid)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
for epoch in range(maxEpoch) :
error = []
for batch in range(numBatch) :
# train each data batch
if batchSize*(batch+1) > numCases :
visTrue = deviceData.get_row_slice(batchSize*batch, numCases)
batchSize = visTrue.shape[0]
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidState = cm.empty((batchSize, numHid))
else :
visTrue = deviceData.get_row_slice(batchSize*batch, batchSize*(batch+1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
#apply momentum
weightInc.mult(momentum)
biasVInc.mult(momentum)
biasHInc.mult(momentum)
# positive phase
cm.dot(visActP, weight, target = hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.add_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0)
biasHInc.add_sums(hidActP, axis=0)
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidActP)
if cmp(method, "SML") == 0 :
if np.logical_and(np.equal(epoch,1), np.equal(batch,1)) :
pass # here does not need in practical use
elif cmp(method, "CD") == 0 :
pass
# negetive phase
if cmp(modelType, "BB") == 0 :
cm.dot(hidActP, weight.T, target = visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
elif cmp(modelType, "GB") == 0 :
cm.dot(hidActP, weight.T, target = visActP)
visActP.add_row_vec(biasV)
visActP.add(np.random.randn(batchSize, numVis),target=visActP)
# another positive phase
cm.dot(visActP, weight, target = hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
weightInc.subtract_dot(visActP.T, hidActP)
biasVInc.add_sums(visActP, axis=0, mult=-1)
biasHInc.add_sums(hidActP, axis=0, mult=-1)
#update weight and bias
weight.add_mult(weightInc, eta/batchSize)
biasV.add_mult(biasVInc, eta/batchSize)
biasH.add_mult(biasHInc, eta/batchSize)
# if epoch > avgStart :
# # apply average
# weightAgv.subtract(weightAgv.subtract(weight).mult(1.0/t))
# biasVAgv.subtract(biasVAgv.subtract(biasV).mult(1.0/t))
# biasHAgv.subtract(biasHAgv.subtract(biasH).mult(1.0/t))
# t = t+1
# else :
# weightAgv = weight
# biasVAgv = biasV
# biasHAgv = biasH
# reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm() ** 2)
# free device memory
visTrue.free_device_memory()
if verbose :
print "epoch %d/%d. Reconstruction error is %f " % (epoch+1, maxEpoch, sum(error))
# save rbm model
top = cm.CUDAMatrix(np.zeros((numCases, numHid)))
cm.dot(cm.CUDAMatrix(cm.reformat(X)), weight, target = top)
top.add_row_vec(biasH)
top.apply_sigmoid()
weight.copy_to_host()
biasV.copy_to_host()
biasH.copy_to_host()
top.copy_to_host()
model_ = m.rbmModel(weight.numpy_array, biasV.numpy_array, \
biasH.numpy_array, type = modelType, top = top.numpy_array)
# free device memory
deviceData.free_device_memory()
weight.free_device_memory()
biasV.free_device_memory()
biasH.free_device_memory()
weightInc.free_device_memory()
biasVInc.free_device_memory()
biasHInc.free_device_memory()
hidActP.free_device_memory()
visActP.free_device_memory()
hidState.free_device_memory()
cm.shutdown()
return model_
def rbm(data, numHid, modelType="BB", **kwargs):
"""
rbm defination
data : when type is BB, should be binary, or in [0,1] to be interpreted as probabilities
when type is GB, should be continuous real value. data should have a format of *.npy
numHid : number nodes of and hidden layer
type : rbm type, can be set as BB or GB
additional inputs (specified as name value pairs or in struct)
method CD or SML
eta learning rate
momentum momentum for smoothness amd to prevent overfitting
NOTE: momentum is not recommended with SML
maxepoch # of epochs: each is a full pass through train data
avglast how many epochs before maxepoch to start averaging
before. Procedure suggested for faster convergence by
Kevin Swersky in his MSc thesis
penalty weight decay factor
batchsize The number of training instances per batch
verbose For printing progress
anneal Flag. If set true, the penalty is annealed linearly
through epochs to 10% of its original value
OUTPUTS:
model.type Type of RBM (i.e. type of its visible and hidden units)
model.weight The weights of the connections
model.biasH The biases of the hidden layer
model.biasV The biases of the visible layer
model.top The activity of the top layer, to be used when training
DBN's
errors The errors in reconstruction at every epoch
"""
arg = util.processOptions(kwargs, \
method = "CD", \
eta = 0.1, \
momentum = 0.9,\
maxEpoch = 50, \
avgLast = 0, \
penalty = 0, \
batchSize = 50, \
verbose = True, \
anneal = False)
[method, eta, momentum, maxEpoch, avgLast, penalty, batchSize, verbose, anneal] = [\
arg["method"],\
arg["eta"],\
arg["momentum"],\
arg["maxEpoch"],\
arg["avgLast"],\
arg["penalty"],\
arg["batchSize"],\
arg["verbose"],\
arg["anneal"]
]
# from which step, we start to compute the average
avgStart = maxEpoch - avgLast
# for weight decay use
oldPenalty = penalty
# numCases : number of example
# numDims : the length of each example
# each row is an example
[numCases, numDims] = list(data.shape)
if verbose:
print "processing data"
numVis = numDims
numBatch = util.ceil(numCases, batchSize)
# shuffle the data
np.random.shuffle(data)
# init CUDA
# cm.cuda_set_device()
cm.cublas_init()
cm.CUDAMatrix.init_random(100)
deviceData = cm.CUDAMatrix(cm.reformat(data))
# init weights
weight = cm.CUDAMatrix(0.1 * np.random.randn(numVis, numHid))
biasV = cm.CUDAMatrix(np.zeros((1, numVis)))
biasH = cm.CUDAMatrix(np.zeros((1, numHid)))
# init weight update
weightInc = cm.CUDAMatrix(np.zeros((numVis, numHid)))
biasVInc = cm.CUDAMatrix(np.zeros((1, numVis)))
biasHInc = cm.CUDAMatrix(np.zeros((1, numHid)))
#init temporary storage
visActP = cm.empty((batchSize, numVis))
hidActP = cm.empty((batchSize, numHid))
hidActP2 = cm.empty((batchSize, numHid))
visState = cm.empty((batchSize, numVis))
hidState = cm.empty((batchSize, numHid))
t = 1
for epoch in range(maxEpoch):
error = []
if anneal:
# apply linear weight decay
penalty = oldPenalty - 0.9 * epoch / maxEpoch * oldPenalty
for batch in range(numBatch):
# train each data batch
if batchSize * (batch + 1) > numCases:
visTrue = deviceData.get_row_slice(batchSize * batch, numCases)
batchSize = visTrue.shape[0]
else:
visTrue = deviceData.get_row_slice(batchSize * batch,
batchSize * (batch + 1))
batchSize = visTrue.shape[0]
visActP.assign(visTrue)
# positive phase
cm.dot(visActP, weight, target=hidActP)
hidActP.add_row_vec(biasH)
hidActP.apply_sigmoid()
hidState.fill_with_rand()
hidState.less_than(hidActP, target=hidState)
if cmp(method, "SML") == 0:
if np.logical_and(np.equal(epoch, 1), np.equal(batch, 1)):
pass # here does not need in practical use
elif cmp(method, "CD") == 0:
pass
# negetive phase
if cmp(modelType, "BB") == 0:
cm.dot(hidState, weight.transpose(), target=visActP)
visActP.add_row_vec(biasV)
visActP.apply_sigmoid()
visState.fill_with_rand()
visState.less_than(visActP, target=visState)
elif cmp(modelType, "GB") == 0:
cm.dot(hidState, weight.transpose(), target=visActP)
visActP.add_row_vec(biasV)
visActP.add(np.random.randn(batchSize, numVis),
target=visState)
# another positive phase
cm.dot(visState, weight, target=hidActP2)
hidActP2.add_row_vec(biasH)
hidActP2.apply_sigmoid()
hidState.fill_with_rand()
hidState.less_than(hidActP2, target=hidState)
#update weight and bias
dWeight = cm.dot(visTrue.transpose(), hidActP)
dWeight.subtract_dot(visState.transpose(), hidActP2)
dBiasV = visTrue.sum(axis=0).subtract(visState.sum(axis=0))
dBiasH = hidActP.sum(axis=0).subtract(hidActP2.sum(axis=0))
dWeight.divide(batchSize).subtract(weight.mult(penalty))
dBiasV.divide(batchSize)
dBiasH.divide(batchSize)
weightInc.mult(momentum).add_mult(dWeight, eta)
biasVInc.mult(momentum).add_mult(dBiasV, eta)
biasHInc.mult(momentum).add_mult(dBiasH, eta)
weight.add(weightInc)
biasV.add(biasVInc)
biasH.add(biasHInc)
if epoch > avgStart:
# apply average
weightAgv.subtract(weightAgv.subtract(weight).mult(1.0 / t))
biasVAgv.subtract(biasVAgv.subtract(biasV).mult(1.0 / t))
biasHAgv.subtract(biasHAgv.subtract(biasH).mult(1.0 / t))
t = t + 1
else:
weightAgv = weight
biasVAgv = biasV
biasHAgv = biasH
# reconstruction error
visTrue.subtract(visActP)
error.append(visTrue.euclid_norm()**2)
if verbose:
print "epoch %d/%d. Reconstruction error is %f " % (
epoch + 1, maxEpoch, sum(error))
# save rbm model
top = cm.CUDAMatrix(np.zeros((numCases, numHid)))
cm.dot(deviceData, weightAgv, target=top)
top.add_row_vec(biasHAgv)
top.apply_sigmoid()
model_ = m.rbmModel(weightAgv, biasVAgv, biasHAgv, type=modelType, top=top)
cm.shutdown()
return model_