def minibatchTrain(self, train, test, rtest, batch_size):
self.nn = self.RBM.nn
slearn_rate = self.RBM.modelArgs.learn_rate
max_iter = self.RBM.modelArgs.max_iter
CD = self.RBM.modelArgs.CD
lamda = self.RBM.modelArgs.lamda
momentum = self.RBM.modelArgs.momentum
min_learn_rate = self.RBM.modelArgs.min_learn_rate
dW_old = np.zeros(self.nn.weights[0].shape)
dv_old = np.zeros(self.nn.layers[0].bias.shape)
dh_old = np.zeros(self.nn.layers[1].bias.shape)
evaluate = EvaluateRBM(self.RBM)
m, n = train.shape
batches = range(0, m, batch_size)
if batches[-1] != m:
if (m - batches[-1]) < (batch_size / 2.0):
batches[-1] = m
else:
batches.append(m)
for i in range(max_iter):
if i > 50:
CD = 3
momentum = 0.9
for j in range(len(batches) - 1):
start = batches[j]
end = batches[j + 1]
learn_rate = slearn_rate / (end - start)
learn_rate = max(learn_rate, min_learn_rate)
vispos = train[start:end, :]
visneg = vispos.copy()
hidpos = self.RBM.getHiddenActivation(vispos)
hidneg = hidpos
for k in range(CD):
visneg_data = self.RBM.getVisibleActivation(hidneg, vispos)
visneg.data = visneg_data
hidneg = self.RBM.getHiddenActivation(visneg)
dW = momentum * dW_old + learn_rate *\
((vispos.T * hidpos) -
(visneg.T * hidneg) - lamda * self.nn.weights[0])
dvbias = momentum * dv_old + learn_rate *\
((vispos - visneg).sum(axis=0) -
lamda * self.nn.layers[0].bias)
dhbias = momentum * dh_old + 0.1 * learn_rate *\
((hidpos - hidneg).sum(axis=0) -
lamda * self.nn.layers[1].bias)
dW_old = dW
dv_old = dvbias
dh_old = dhbias
self.nn.weights[0] += dW
self.nn.layers[0].bias += dvbias
self.nn.layers[1].bias += dhbias
if i % 5 == 0:
slearn_rate *= 0.95
print evaluate.calculateRMSEandMAE(train, test, rtest)
def sgdTrain(self, train, test, rtest):
self.nn = self.RBM.nn
learn_rate = self.RBM.modelArgs.learn_rate
max_iter = self.RBM.modelArgs.max_iter
CD = self.RBM.modelArgs.CD
lamda = self.RBM.modelArgs.lamda
momentum = self.RBM.modelArgs.momentum
dW_old = np.zeros(self.nn.weights[0].shape)
dv_old = np.zeros(self.nn.layers[0].bias.shape)
dh_old = np.zeros(self.nn.layers[1].bias.shape)
evaluate = EvaluateRBM(self.RBM)
# traindata = train.data
# testdata = test.data
m, n = train.shape
for i in range(max_iter):
if i > 50:
CD = 3
momentum = 0.9
for j in range(m - 1):
vispos = train.getrow(j)
visneg = vispos.copy()
hidpos = self.RBM.getHiddenActivation(vispos)
hidneg = hidpos
for k in range(CD):
visneg_data = self.RBM.getVisibleActivation(hidneg, vispos)
visneg.data = visneg_data
hidneg = self.RBM.getHiddenActivation(visneg)
dW = momentum * dW_old + learn_rate *\
((vispos.T * hidpos) -
(visneg.T * hidneg) - lamda * self.nn.weights[0])
dvbias = momentum * dv_old + learn_rate *\
((vispos - visneg).sum(axis=0) -
lamda * self.nn.layers[0].bias)
dhbias = momentum * dh_old + 0.1 * learn_rate *\
((hidpos - hidneg).sum(axis=0) -
lamda * self.nn.layers[1].bias)
dW_old = dW
dv_old = dvbias
dh_old = dhbias
self.nn.weights[0] += dW
self.nn.layers[0].bias += dvbias
self.nn.layers[1].bias += dhbias
if i % 5 == 0:
slearn_rate *= 0.95
print evaluate.calculateRMSEandMAE(train, test, rtest)
def sgdTrain(self, train, test, rtest):
self.nn = self.RBM.nn
learn_rate = self.RBM.modelArgs.learn_rate
max_iter = self.RBM.modelArgs.max_iter
CD = self.RBM.modelArgs.CD
lamda = self.RBM.modelArgs.lamda
momentum = self.RBM.modelArgs.momentum
dW_old = np.zeros(self.nn.weights[0].shape)
dv_old = np.zeros(self.nn.layers[0].bias.shape)
dh_old = np.zeros(self.nn.layers[1].bias.shape)
evaluate = EvaluateRBM(self.RBM)
# traindata = train.data
# testdata = test.data
m, n = train.shape
for i in range(max_iter):
if i > 50:
CD = 3
momentum = 0.9
for j in range(m - 1):
vispos = train.getrow(j)
visneg = vispos.copy()
hidpos = self.RBM.getHiddenActivation(vispos)
hidneg = hidpos
for k in range(CD):
visneg_data = self.RBM.getVisibleActivation(hidneg, vispos)
visneg.data = visneg_data
hidneg = self.RBM.getHiddenActivation(visneg)
dW = momentum * dW_old + learn_rate *\
((vispos.T * hidpos) -
(visneg.T * hidneg) - lamda * self.nn.weights[0])
dvbias = momentum * dv_old + learn_rate *\
((vispos - visneg).sum(axis=0) -
lamda * self.nn.layers[0].bias)
dhbias = momentum * dh_old + 0.1 * learn_rate *\
((hidpos - hidneg).sum(axis=0) -
lamda * self.nn.layers[1].bias)
dW_old = dW
dv_old = dvbias
dh_old = dhbias
self.nn.weights[0] += dW
self.nn.layers[0].bias += dvbias
self.nn.layers[1].bias += dhbias
if i % 5 == 0:
slearn_rate *= 0.95
print evaluate.calculateRMSEandMAE(train, test, rtest)
def minibatchTrain(self, train, test, rtest, batch_size):
self.nn = self.RBM.nn
slearn_rate = self.RBM.modelArgs.learn_rate
max_iter = self.RBM.modelArgs.max_iter
CD = self.RBM.modelArgs.CD
lamda = self.RBM.modelArgs.lamda
momentum = self.RBM.modelArgs.momentum
min_learn_rate = self.RBM.modelArgs.min_learn_rate
dW_old = np.zeros(self.nn.weights[0].shape)
dv_old = np.zeros(self.nn.layers[0].bias.shape)
dh_old = np.zeros(self.nn.layers[1].bias.shape)
evaluate = EvaluateRBM(self.RBM)
m, n = train.shape
batches = range(0, m, batch_size)
if batches[-1] != m:
if (m - batches[-1]) < (batch_size / 2.0):
batches[-1] = m
else:
batches.append(m)
for i in range(max_iter):
if i > 50:
CD = 3
momentum = 0.9
for j in range(len(batches) - 1):
start = batches[j]
end = batches[j + 1]
learn_rate = slearn_rate / (end - start)
learn_rate = max(learn_rate, min_learn_rate)
vispos = train[start:end, :]
visneg = vispos.copy()
hidpos = self.RBM.getHiddenActivation(vispos)
hidneg = hidpos
for k in range(CD):
visneg_data = self.RBM.getVisibleActivation(hidneg, vispos)
visneg.data = visneg_data
hidneg = self.RBM.getHiddenActivation(visneg)
dW = momentum * dW_old + learn_rate *\
((vispos.T * hidpos) -
(visneg.T * hidneg) - lamda * self.nn.weights[0])
dvbias = momentum * dv_old + learn_rate *\
((vispos - visneg).sum(axis=0) -
lamda * self.nn.layers[0].bias)
dhbias = momentum * dh_old + 0.1 * learn_rate *\
((hidpos - hidneg).sum(axis=0) -
lamda * self.nn.layers[1].bias)
dW_old = dW
dv_old = dvbias
dh_old = dhbias
self.nn.weights[0] += dW
self.nn.layers[0].bias += dvbias
self.nn.layers[1].bias += dhbias
if i % 5 == 0:
slearn_rate *= 0.95
print evaluate.calculateRMSEandMAE(train, test, rtest)
