def test_fc_backprop(self):
for case in self.fc_cases:
weight = case['weight']
out_c, in_c = weight.shape
bias = case['bias']
x = case['x'].astype(np.float32)
out = case['out']
grad_output = case['grad_output']
grad_x = case['grad_x']
grad_w = case['grad_w']
grad_b = case['grad_b']
fc = FullyConnected(d_in=in_c, d_out=out_c)
fc.W = weight
fc.b = bias
test_out = fc(x)
dv_x, dv_W, dv_b = fc.backward(x, grad_output)
self.assertTrue(np.allclose(out, test_out, rtol=0.0001))
self.assertTrue(np.allclose(grad_x, dv_x, rtol=0.001))
self.assertTrue(np.allclose(grad_w, dv_W))
self.assertTrue(np.allclose(grad_b, dv_b))
def build_model(self, dataset_name):
if dataset_name is 'mnist':
self.add_layer(
Convolutional(name='conv1',
num_filters=8,
stride=2,
size=3,
activation='relu'))
self.add_layer(
Convolutional(name='conv2',
num_filters=8,
stride=2,
size=3,
activation='relu'))
self.add_layer(Dense(name='dense', nodes=8 * 6 * 6,
num_classes=10))
else:
self.add_layer(
Convolutional(name='conv1',
num_filters=32,
stride=1,
size=3,
activation='relu'))
self.add_layer(
Convolutional(name='conv2',
num_filters=32,
stride=1,
size=3,
activation='relu'))
self.add_layer(Pooling(name='pool1', stride=2, size=2))
self.add_layer(
Convolutional(name='conv3',
num_filters=64,
stride=1,
size=3,
activation='relu'))
self.add_layer(
Convolutional(name='conv4',
num_filters=64,
stride=1,
size=3,
activation='relu'))
self.add_layer(Pooling(name='pool2', stride=2, size=2))
self.add_layer(
FullyConnected(name='fullyconnected',
nodes1=64 * 5 * 5,
nodes2=256,
activation='relu'))
self.add_layer(Dense(name='dense', nodes=256, num_classes=10))
def __init__(self):
self.parameter = dict()
self.grad = dict()
self.gradInput = dict()
self.Fc1 = FullyConnected()
self.Fc2 = FullyConnected()
self.Fc3 = FullyConnected()
self.Relu1 = Relu()
self.Relu2 = Relu()
self.Relu3 = Relu()
self.softmaxloss = SoftMaxLoss()
self.optimizer = Optimizer()
class Net:
def __init__(self):
self.parameter = dict()
self.grad = dict()
self.gradInput = dict()
self.Fc1 = FullyConnected()
self.Fc2 = FullyConnected()
self.Fc3 = FullyConnected()
self.Relu1 = Relu()
self.Relu2 = Relu()
self.Relu3 = Relu()
self.softmaxloss = SoftMaxLoss()
self.optimizer = Optimizer()
def initWeigth(self):
# self.parameter['w1'] = np.zeros(shape=(28*28, 32), dtype=np.float32)
# self.parameter['w2'] = np.zeros(shape=(32, 10), dtype=np.float32)
# self.parameter['b1'] = np.zeros(shape=(32), dtype=np.float32)
# self.parameter['b2'] = np.zeros(shape=(10), dtype=np.float32)
self.parameter['w1'] = np.random.rand(28 * 28, 32) * np.sqrt(
1 / (28 * 28 + 32))
self.parameter['w2'] = np.random.rand(32, 16) * np.sqrt(1 / (32 + 16))
self.parameter['w3'] = np.random.rand(16, 10) * np.sqrt(1 / (16 + 10))
self.parameter['b1'] = np.random.rand(32)
self.parameter['b2'] = np.random.rand(16)
self.parameter['b3'] = np.random.rand(10)
# self.parameter['w1'] = np.random.uniform(low=0.5, high=15, size = (28 * 28, 32))
# self.parameter['w2'] = np.random.uniform(low=0.3, high=15, size = (32, 16))
# self.parameter['w3'] = np.random.uniform(low=0.1, high=2, size = (16, 10))
# self.parameter['b1'] = np.random.uniform(low=0, high=1, size = (32) )
# self.parameter['b2'] = np.random.uniform(low=0., high=1, size = (16) )
# self.parameter['b3'] = np.random.uniform(low=0., high=1, size = (10) )
def ForwardPass(self, image):
fc1, fca2 = self.Fc1.forward(
input=image,
weight=self.parameter["w1"],
bias=self.parameter["b1"]) # in = 1x1024, out = 1x32
relu1, _ = self.Relu1.forward(fc1)
fc2, _ = self.Fc2.forward(
input=relu1,
weight=self.parameter["w2"],
bias=self.parameter["b2"]) # in = 1x32 , out = 1x10
relu2, _ = self.Relu2.forward(fc2)
fc3, _ = self.Fc3.forward(
input=relu2,
weight=self.parameter["w3"],
bias=self.parameter["b3"]) # in = 1x32 , out = 1x10
relu3, _ = self.Relu3.forward(fc3)
out = Softmax.forward(relu3)
return out
def BackwordPass(self, prediction):
dout = self.softmaxloss.backward(prediction)
dout = self.Relu3.backward(dout)
delta3, dw3, db3 = self.Fc3.backward(dout=dout)
relu_delta2 = self.Relu2.backward(delta3)
delta2, dw2, db2 = self.Fc2.backward(dout=relu_delta2)
relu_delta1 = self.Relu1.backward(delta2)
delta1, dw1, db1 = self.Fc1.backward(relu_delta1)
self.grad['dw1'] = dw1
self.grad['dw2'] = dw2
self.grad['dw3'] = dw3
self.grad['db1'] = db1
self.grad['db2'] = db2
self.grad['db3'] = db3
def Train(self, image, label):
data = np.array(image, dtype=np.float32)
data = data.flatten()
input_data = np.reshape(data, (1, data.size))
label = np.reshape(label, (1, label.size))
output = self.ForwardPass(input_data)
# print(output)
loss = self.softmaxloss.forward(output, label)
# print(loss)
#print(loss)
self.BackwordPass(output)
return loss, output
def Test(self, testImages, testLabelsHotEncoding):
numberOfImages = testImages.shape[0]
# numberOfImages = 10
avgloss = 0
avgacc = 0
count = 0
for iter in range(numberOfImages):
image = testImages[iter, :, :]
labels = testLabelsHotEncoding[iter, :]
data = np.array(image, dtype=np.float32)
data = data.flatten()
input_data = np.reshape(data, (1, data.size))
labels = np.reshape(labels, (1, labels.size))
output = self.ForwardPass(input_data)
pred = np.argmax(output[0])
gt = np.argmax(labels[0])
loss = self.softmaxloss.forward(output, labels)
if pred == gt:
count += 1
#print("True")
avgloss += loss
avgacc = float(count) / float(numberOfImages)
avgloss = float(avgloss) / float(numberOfImages)
print("Test Accuracy: ", avgacc)
print("Test Loss: ", avgloss)
return avgloss, avgacc
def start_training_mnist(self,
data_folder,
batch_size,
learning_rate,
NumberOfEpoch,
display=1000):
self.initWeigth()
self.optimizer.initADAM(3, 3)
trainingImages, trainingLabels = Dataloader.loadMNIST(
'train', data_folder)
testImages, testLabels = Dataloader.loadMNIST('t10k', data_folder)
trainLabelsHotEncoding = Dataloader.toHotEncoding(trainingLabels)
testLabelsHotEncoding = Dataloader.toHotEncoding(testLabels)
numberOfImages = trainingImages.shape[0]
# numberOfImages = 10
pEpochTrainLoss = []
pEpochTrainAccuracy = []
pEpochTestLoss = []
pEpochTestAccuracy = []
print("Training started")
t = 0
for epoch in range(NumberOfEpoch):
avgLoss = 0
trainAcc = 0.0
count = 0.0
countacc = 0.0
pIterLoss = []
print("##############EPOCH : {}##################".format(epoch))
for iter in range(numberOfImages):
t += 1
image = trainingImages[iter, :, :]
labels = trainLabelsHotEncoding[iter, :]
loss, output = self.Train(image, labels)
self.parameter = self.optimizer.ADAM(self.parameter, self.grad,
learning_rate, t)
self.parameter = self.optimizer.l2_regularization(
self.parameter, 0.001)
output = output[0]
pred = np.argmax(output)
gt = np.argmax(labels)
if pred == gt:
count += 1.0
countacc += 1.0
#print("True")
# self.parameter = self.optimizer.SGD(self.parameter, self.grad, learning_rate)
pIterLoss.append(loss)
avgLoss += loss
if iter % display == 0:
print("Train Accuracy {} with prob : {}".format(
(countacc / float(display)), output[pred]))
print("Train Loss: ", loss)
countacc = 0.0
loss, acc = self.Test(testImages, testLabelsHotEncoding)
trainAcc = (float(count) / float(numberOfImages))
print("##################Overall Accuracy & Loss Calculation")
print("TrainAccuracy: ", trainAcc)
print("TrainLoss: ", (float(avgLoss) / float(numberOfImages)))
avgtestloss, avgtestacc = self.Test(testImages,
testLabelsHotEncoding)
totaloss = float(avgLoss) / float(numberOfImages)
pEpochTrainLoss.append(totaloss)
pEpochTrainAccuracy.append(trainAcc)
pEpochTestLoss.append(avgtestloss)
pEpochTestAccuracy.append(avgtestacc)
x_axis = np.linspace(0, epoch, len(pEpochTrainLoss), endpoint=True)
plt.semilogy(x_axis, pEpochTrainLoss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.draw()
file = open("Weightparameter1.pkl", "wb")
file.write(pickle.dumps(self.parameter))
file.close()
fill2 = open("parameter.pkl", "wb")
fill2.write(
pickle.dumps((pEpochTrainAccuracy, pEpochTrainLoss,
pEpochTestAccuracy, pEpochTestLoss)))
fill2.close()
def start_training(self, datafolder, batchsize, learning_rate,
number_of_epoch):
pass
def c_6layer_mnist_imputation(seed=0,
pertub_type=3,
pertub_prob=6,
pertub_prob1=14,
predir=None,
n_batch=144,
dataset='mnist.pkl.gz',
batch_size=500):
"""
Missing data imputation
"""
#cp->cd->cpd->cd->c
nkerns = [32, 32, 64, 64, 64]
drops = [0, 0, 0, 0, 0, 1]
#skerns=[5, 3, 3, 3, 3]
#pools=[2, 1, 1, 2, 1]
#modes=['same']*5
n_hidden = [500, 50]
drop_inverses = [
1,
]
# 28->12->12->5->5/5*5*64->500->50->500->5*5*64/5->5->12->12->28
if dataset == 'mnist.pkl.gz':
dim_input = (28, 28)
colorImg = False
train_set_x, test_set_x, test_set_x_pertub, pertub_label, pertub_number = datapy.load_pertub_data(
dirs='data_imputation/',
pertub_type=pertub_type,
pertub_prob=pertub_prob,
pertub_prob1=pertub_prob1)
datasets = datapy.load_data_gpu(dataset, have_matrix=True)
_, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
_, test_set_y, test_y_matrix = datasets[2]
# compute number of minibatches for training, validation and testing
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
#x_pertub = T.matrix('x_pertub') # the data is presented as rasterized images
#p_label = T.matrix('p_label')
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
y_matrix = T.imatrix('y_matrix')
drop = T.iscalar('drop')
drop_inverse = T.iscalar('drop_inverse')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x.reshape((batch_size, 1, 28, 28))
recg_layer = []
cnn_output = []
#1
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, 1, 28, 28),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x,
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
#2
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#3
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[1], 12, 12),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#4
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[2], 5, 5),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#5
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[3], 5, 5),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
mlp_input = cnn_output[-1].flatten(2)
recg_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=nkerns[4] * 5 * 5,
n_out=500,
activation=activation))
feature = recg_layer[-1].drop_output(mlp_input, drop=drop, rng=rng_share)
# classify the values of the fully-connected sigmoidal layer
classifier = Pegasos.Pegasos(input=feature,
rng=rng,
n_in=500,
n_out=10,
weight_decay=0,
loss=1)
# the cost we minimize during training is the NLL of the model
cost = classifier.hinge_loss(10, y, y_matrix) * batch_size
weight_decay = 1.0 / n_train_batches
# create a list of all model parameters to be fit by gradient descent
params = []
for r in recg_layer:
params += r.params
params += classifier.params
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
learning_rate = 3e-4
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
get_optimizer = optimizer.get_adam_optimizer_min(learning_rate=l_r,
decay1=0.1,
decay2=0.001,
weight_decay=weight_decay)
updates = get_optimizer(params, grads)
'''
Save parameters and activations
'''
parameters = theano.function(
inputs=[],
outputs=params,
)
# create a function to compute the mistakes that are made by the model
test_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
#y_matrix: test_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
})
test_pertub_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x_pertub[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
#y_matrix: test_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
})
validate_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
#y_matrix: valid_y_matrix[index * batch_size: (index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0)
})
##################
# Pretrain MODEL #
##################
model_epoch = 250
if os.environ.has_key('model_epoch'):
model_epoch = int(os.environ['model_epoch'])
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
if model_epoch == -1:
pre_train = np.load(predir + 'best-model.npz')
else:
pre_train = np.load(predir + 'model-' + str(model_epoch) + '.npz')
pre_train = pre_train['model']
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
else:
exit()
###############
# TRAIN MODEL #
###############
valid_losses = [validate_model(i) for i in xrange(n_valid_batches)]
valid_score = np.mean(valid_losses)
test_losses = [test_model(i) for i in xrange(n_test_batches)]
test_score = np.mean(test_losses)
test_losses_pertub = [test_pertub_model(i) for i in xrange(n_test_batches)]
test_score_pertub = np.mean(test_losses_pertub)
print valid_score, test_score, test_score_pertub
def cmmva_6layer_svhn(learning_rate=0.01,
n_epochs=600,
dataset='svhngcn_var',
batch_size=500,
dropout_flag=1,
seed=0,
predir=None,
activation=None,
n_batch=625,
weight_decay=1e-4,
super_predir=None,
super_preepoch=None):
"""
Implementation of convolutional MMVA
"""
'''
svhn
'''
n_channels = 3
colorImg = True
dim_w = 32
dim_h = 32
dim_input=(dim_h, dim_w)
n_classes = 10
D = 1.0
C = 1.0
if os.environ.has_key('C'):
C = np.cast['float32'](float((os.environ['C'])))
if os.environ.has_key('D'):
D = np.cast['float32'](float((os.environ['D'])))
color.printRed('D '+str(D)+' C '+str(C))
first_drop=0.5
if os.environ.has_key('first_drop'):
first_drop = float(os.environ['first_drop'])
last_drop=1
if os.environ.has_key('last_drop'):
last_drop = float(os.environ['last_drop'])
nkerns_1=96
if os.environ.has_key('nkerns_1'):
nkerns_1 = int(os.environ['nkerns_1'])
nkerns_2=96
if os.environ.has_key('nkerns_2'):
nkerns_2 = int(os.environ['nkerns_2'])
n_z=512
if os.environ.has_key('n_z'):
n_z = int(os.environ['n_z'])
opt_med='adam'
if os.environ.has_key('opt_med'):
opt_med = os.environ['opt_med']
train_logvar=True
if os.environ.has_key('train_logvar'):
train_logvar = bool(int(os.environ['train_logvar']))
std = 2e-2
if os.environ.has_key('std'):
std = os.environ['std']
Loss_L = 1
if os.environ.has_key('Loss_L'):
Loss_L = int(os.environ['Loss_L'])
pattern = 'hinge'
if os.environ.has_key('pattern'):
pattern = os.environ['pattern']
#cp->cd->cpd->cd->c
nkerns=[nkerns_1, nkerns_1, nkerns_1, nkerns_2, nkerns_2]
drops=[0, 1, 1, 1, 0, 1]
drop_p=[1, first_drop, first_drop, first_drop, 1, last_drop]
n_hidden=[n_z]
logdir = 'results/supervised/cmmva/svhn/cmmva_6layer_'+dataset+pattern+'_D_'+str(D)+'_C_'+str(C)+'_'#+str(nkerns)+str(n_hidden)+'_'+str(weight_decay)+'_'+str(learning_rate)+'_'
#if predir is not None:
# logdir +='pre_'
#if dropout_flag == 1:
# logdir += ('dropout_'+str(drops)+'_')
# logdir += ('drop_p_'+str(drop_p)+'_')
#logdir += ('trainvar_'+str(train_logvar)+'_')
#logdir += (opt_med+'_')
#logdir += (str(Loss_L)+'_')
#if super_predir is not None:
# logdir += (str(super_preepoch)+'_')
logdir += str(int(time.time()))+'/'
if not os.path.exists(logdir): os.makedirs(logdir)
print 'logdir:', logdir, 'predir', predir
print 'cmmva_6layer_svhn_fix', nkerns, n_hidden, seed, dropout_flag, drops, drop_p
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'logdir:', logdir, 'predir', predir
print >>f, 'cmmva_6layer_svhn_fix', nkerns, n_hidden, seed, dropout_flag, drops, drop_p
color.printRed('dataset '+dataset)
datasets = datapy.load_data_svhn(dataset, have_matrix=True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
test_set_x, test_set_y, test_y_matrix = datasets[1]
valid_set_x, valid_set_y, valid_y_matrix = datasets[2]
#datasets = datapy.load_data_svhn(dataset, have_matrix=False)
#train_set_x, train_set_y = datasets[0]
#test_set_x, test_set_y = datasets[1]
#valid_set_x, valid_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
random_z = T.matrix('random_z')
y_matrix = T.imatrix('y_matrix')
drop = T.iscalar('drop')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x.reshape((batch_size, n_channels, dim_h, dim_w))
recg_layer = []
cnn_output = []
l = []
d = []
#1
recg_layer.append(ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, n_channels, dim_h, dim_w),
filter_shape=(nkerns[0], n_channels, 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=activation,
std=std
))
if drops[0]==1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x, drop=drop, rng=rng_share, p=drop_p[0]))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
l+=[1, 2]
d+=[1, 0]
#2
recg_layer.append(ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[0], 16, 16),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation,
std=std
))
if drops[1]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share, p=drop_p[1]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l+=[1, 2]
d+=[1, 0]
#3
recg_layer.append(ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[1], 16, 16),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation,
std=std
))
if drops[2]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share, p=drop_p[2]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l+=[1, 2]
d+=[1, 0]
#4
recg_layer.append(ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[2], 8, 8),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation,
std=std
))
if drops[3]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share, p=drop_p[3]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l+=[1, 2]
d+=[1, 0]
#5
recg_layer.append(ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[3], 8, 8),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation,
std=std
))
if drops[4]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share, p=drop_p[4]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l+=[1, 2]
d+=[1, 0]
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
activations.append(mlp_input_x)
classifier = Pegasos.Pegasos(
input= activations[-1],
rng=rng,
n_in=nkerns[-1]*4*4,
n_out=n_classes,
weight_decay=0,
loss=Loss_L,
pattern=pattern
)
l+=[1, 2]
d+=[1, 0]
#stochastic layer
recg_layer.append(GaussianHidden.GaussianHidden(
rng=rng,
input=mlp_input_x,
n_in=4*4*nkerns[-1],
n_out=n_hidden[0],
activation=None
))
l+=[1, 2]
d+=[1, 0]
l+=[1, 2]
d+=[1, 0]
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in=n_hidden[-1],
n_out=4*4*nkerns[-1],
activation=activation
))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
l+=[1, 2]
d+=[1, 0]
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 4, 4))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 4, 4))
#1
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-1], 4, 4),
filter_shape=(nkerns[-2], nkerns[-1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation
))
l+=[1, 2]
d+=[1, 0]
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-2], 8, 8),
filter_shape=(nkerns[-3], nkerns[-2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
l+=[1, 2]
d+=[1, 0]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-3], 8, 8),
filter_shape=(nkerns[-4], nkerns[-3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation
))
l+=[1, 2]
d+=[1, 0]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-4], 16, 16),
filter_shape=(nkerns[-5], nkerns[-4], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
l+=[1, 2]
d+=[1, 0]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#5-1 stochastic layer
# for this layer, the activation is None to get a Guassian mean
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None
))
l+=[1, 2]
d+=[1, 0]
x_mean=gene_layer[-1].output(input=z_output[-1])
random_x_mean=gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch)
#5-2 stochastic layer
# for this layer, the activation is None to get logvar
if train_logvar:
gene_layer.append(UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None
))
l+=[1, 2]
d+=[1, 0]
x_logvar=gene_layer[-1].output(input=z_output[-1])
random_x_logvar=gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch)
else:
x_logvar = theano.shared(np.ones((batch_size, n_channels, dim_h, dim_w), dtype='float32'))
random_x_logvar = theano.shared(np.ones((n_batch, n_channels, dim_h, dim_w), dtype='float32'))
gene_layer.append(NoParamsGaussianVisiable.NoParamsGaussianVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=x_mean, logvar=x_logvar, data=input_x)
random_x = gene_layer[-1].sample_x(rng_share=rng_share, mean=random_x_mean, logvar=random_x_logvar)
#L = (logpx + logpz - logqz).sum()
lowerbound = (
(logpx + recg_layer[-1].logpz - recg_layer[-1].logqz).mean()
)
hinge_loss = classifier.hinge_loss(10, y, y_matrix)
cost = D * lowerbound - C * hinge_loss
px = (logpx.mean())
pz = (recg_layer[-1].logpz.mean())
qz = (- recg_layer[-1].logqz.mean())
super_params=[]
for r in recg_layer[:-1]:
super_params+=r.params
super_params+=classifier.params
params=[]
for g in gene_layer:
params+=g.params
for r in recg_layer:
params+=r.params
params+=classifier.params
grads = [T.grad(cost, param) for param in params]
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
#get_optimizer = optimizer.get_adam_optimizer(learning_rate=learning_rate)
if opt_med=='adam':
get_optimizer = optimizer_separated.get_adam_optimizer_max(learning_rate=l_r, decay1 = 0.1, decay2 = 0.001, weight_decay=weight_decay)
elif opt_med=='mom':
get_optimizer = optimizer_separated.get_momentum_optimizer_max(learning_rate=l_r, weight_decay=weight_decay)
updates = get_optimizer(w=params,g=grads, l=l, d=d)
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
test_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
#outputs=layer[-1].errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix: test_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
valid_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
#outputs=layer[-1].errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix: valid_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
valid_error = theano.function(
inputs=[index],
outputs=classifier.errors(y),
#outputs=layer[-1].errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
#y_matrix: valid_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
'''
Save parameters and activations
'''
pog = []
for (p,g) in zip(params, grads):
pog.append(p.max())
pog.append((p**2).mean())
pog.append((g**2).mean())
pog.append((T.sqrt(pog[-2] / pog[-1]))/ 1e3)
paramovergrad = theano.function(
inputs=[index],
outputs=pog,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag)
}
)
parameters = theano.function(
inputs=[],
outputs=params,
)
generation_check = theano.function(
inputs=[index],
outputs=[x, x_mean.flatten(2), x_logvar.flatten(2)],
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
#y: train_set_y[index * batch_size: (index + 1) * batch_size],
#y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: valid_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: valid_set_y[index * batch_size: (index + 1) * batch_size]
}
)
test_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: test_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: test_set_y[index * batch_size: (index + 1) * batch_size]
}
)
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
debug_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, px, pz, qz, hinge_loss, cost],
#updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag)
}
)
random_generation = theano.function(
inputs=[random_z],
outputs=[random_x_mean.flatten(2), random_x.flatten(2)],
givens={
#drop: np.cast['int32'](0)
}
)
train_bound_without_dropout = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
train_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost, px, pz, qz, z],
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag),
}
)
##################
# Pretrain MODEL #
##################
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
pre_train = np.load(predir+'model.npz')
pre_train = pre_train['model']
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0)
print '------------------', tmp
if super_predir is not None:
color.printBlue('... setting parameters')
color.printBlue(super_predir)
pre_train = np.load(super_predir+'svhn_model-'+str(super_preepoch)+'.npz')
pre_train = pre_train['model']
for (para, pre) in zip(super_params, pre_train):
para.set_value(pre)
this_test_losses = [test_model(i) for i in xrange(n_test_batches)]
this_test_score = np.mean(this_test_losses, axis=0)
#print predir
print 'preepoch', super_preepoch, 'pre_test_score', this_test_score
with open(logdir+'hook.txt', 'a') as f:
print >>f, predir
print >>f, 'preepoch', super_preepoch, 'pre_test_score', this_test_score
###############
# TRAIN MODEL #
###############
print '... training'
validation_frequency = n_train_batches
predy_valid_stats = [1, 1, 0]
start_time = time.clock()
NaN_count = 0
epoch = 0
threshold = 0
generatition_frequency = 1
if predir is not None:
threshold = 0
color.printRed('threshold, '+str(threshold) +
' generatition_frequency, '+str(generatition_frequency)
+' validation_frequency, '+str(validation_frequency))
done_looping = False
n_epochs = 80
decay_epochs = 40
record = 0
'''
print 'test initialization...'
pre_model = parameters()
for i in xrange(len(pre_model)):
pre_model[i] = np.asarray(pre_model[i])
print pre_model[i].shape, np.mean(pre_model[i]), np.var(pre_model[i])
print 'end test...'
'''
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
minibatch_avg_cost = 0
train_error = 0
train_lowerbound = 0
train_hinge_loss = 0
_____z = 0
pxx = 0
pzz = 0
qzz = 0
preW = None
currentW = None
tmp_start1 = time.clock()
if epoch == 30:
validation_frequency = n_train_batches/5
if epoch == 50:
validation_frequency = n_train_batches/10
if epoch == 30 or epoch == 50 or epoch == 70 or epoch == 90:
record = epoch
l_r.set_value(np.cast['float32'](l_r.get_value()/3.0))
print '---------', epoch, l_r.get_value()
with open(logdir+'hook.txt', 'a') as f:
print >>f,'---------', epoch, l_r.get_value()
'''
test_epoch = epoch - decay_epochs
if test_epoch > 0 and test_epoch % 5 == 0:
l_r.set_value(np.cast['float32'](l_r.get_value()/3.0))
print '---------------', l_r.get_value()
with open(logdir+'hook.txt', 'a') as f:
print >>f, '---------------', l_r.get_value()
'''
for minibatch_index in xrange(n_train_batches):
e, l, h, ttt, tpx, tpz, tqz, _z = train_model(minibatch_index)
pxx+=tpx
pzz+=tpz
qzz+=tqz
#_____z += (np.asarray(_z)**2).sum() / (n_hidden[-1] * batch_size)
train_error += e
train_lowerbound += l
train_hinge_loss += h
minibatch_avg_cost += ttt
'''
llll = debug_model(minibatch_index)
with open(logdir+'hook.txt', 'a') as f:
print >>f,'[]', llll
'''
if math.isnan(ttt):
color.printRed('--------'+str(epoch)+'--------'+str(minibatch_index))
exit()
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
'''
if (minibatch_index <11):
preW = currentW
currentW = parameters()
for i in xrange(len(currentW)):
currentW[i] = np.asarray(currentW[i]).astype(np.float32)
if preW is not None:
for (c,p) in zip(currentW, preW):
#print minibatch_index, (c**2).mean(), ((c-p)**2).mean(), np.sqrt((c**2).mean()/((c-p)**2).mean())
with open(logdir+'delta_w.txt', 'a') as f:
print >>f,minibatch_index, (c**2).mean(), ((c-p)**2).mean(), np.sqrt((c**2).mean()/((c-p)**2).mean())
'''
# check valid error only, to speed up
'''
if (iter + 1) % validation_frequency != 0 and (iter + 1) %(validation_frequency/10) == 0:
vt = [valid_error(i) for i in xrange(n_valid_batches)]
vt = np.mean(vt)
print 'quick valid error', vt
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'quick valid error', vt
print 'So far best model', predy_valid_stats
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'So far best model', predy_valid_stats
'''
if (iter + 1) % validation_frequency == 0:
print minibatch_index, 'stochastic training error', train_error/float(minibatch_index), train_lowerbound/float(minibatch_index), train_hinge_loss/float(minibatch_index), minibatch_avg_cost /float(minibatch_index), pxx/float(minibatch_index), pzz/float(minibatch_index), qzz/float(minibatch_index)#, 'z_norm', _____z/float(minibatch_index)
with open(logdir+'hook.txt', 'a') as f:
print >>f, minibatch_index, 'stochastic training error', train_error/float(minibatch_index), train_lowerbound/float(minibatch_index), train_hinge_loss/float(minibatch_index), minibatch_avg_cost /float(minibatch_index), pxx/float(minibatch_index), pzz/float(minibatch_index), qzz/float(minibatch_index)#, 'z_norm', _____z/float(minibatch_index)
valid_stats = [valid_model(i) for i in xrange(n_valid_batches)]
this_valid_stats = np.mean(valid_stats, axis=0)
print epoch, minibatch_index, 'validation stats', this_valid_stats
#print tmp
with open(logdir+'hook.txt', 'a') as f:
print >>f, epoch, minibatch_index, 'validation stats', this_valid_stats
print 'So far best model', predy_valid_stats
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'So far best model', predy_valid_stats
if this_valid_stats[0] < predy_valid_stats[0]:
test_stats = [test_model(i) for i in xrange(n_test_batches)]
this_test_stats = np.mean(test_stats, axis=0)
predy_valid_stats[0] = this_valid_stats[0]
predy_valid_stats[1] = this_test_stats[0]
predy_valid_stats[2] = epoch
record = epoch
print 'Update best model', this_test_stats
with open(logdir+'hook.txt', 'a') as f:
print >>f,'Update best model', this_test_stats
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
#print model[i].shape, np.mean(model[i]), np.var(model[i])
np.savez(logdir+'best-model', model=model)
genezero = generation_check(0)
with open(logdir+'gene_check.txt', 'a') as f:
print >>f, 'epoch-----------------------', epoch
print >>f, 'x', 'x_mean', 'x_logvar'
'''
for i in xrange(len(genezero)):
genezero[i] = np.asarray(genezero[i])
with open(logdir+'gene_check.txt', 'a') as f:
print >>f, genezero[i].max(), genezero[i].min(), genezero[i].mean()
with open(logdir+'gene_check.txt', 'a') as f:
print >>f, 'norm', np.sqrt(((genezero[0]- genezero[1])**2).sum())
'''
if epoch==1:
xxx = genezero[0]
image = paramgraphics.mat_to_img(xxx.T, dim_input, colorImg=colorImg, scale=True)
image.save(logdir+'data.png', 'PNG')
if epoch%1==0:
tail='-'+str(epoch)+'.png'
xxx_now = genezero[1]
image = paramgraphics.mat_to_img(xxx_now.T, dim_input, colorImg=colorImg, scale=True)
image.save(logdir+'data_re'+tail, 'PNG')
if math.isnan(minibatch_avg_cost):
NaN_count+=1
color.printRed("NaN detected. Reverting to saved best parameters")
print '---------------NaN_count:', NaN_count
with open(logdir+'hook.txt', 'a') as f:
print >>f, '---------------NaN_count:', NaN_count
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0)
print '------------------NaN check:', tmp
with open(logdir+'hook.txt', 'a') as f:
print >>f, '------------------NaN check:', tmp
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
print model[i].shape, np.mean(model[i]), np.var(model[i])
print np.max(model[i]), np.min(model[i])
print np.all(np.isfinite(model[i])), np.any(np.isnan(model[i]))
with open(logdir+'hook.txt', 'a') as f:
print >>f, model[i].shape, np.mean(model[i]), np.var(model[i])
print >>f, np.max(model[i]), np.min(model[i])
print >>f, np.all(np.isfinite(model[i])), np.any(np.isnan(model[i]))
best_before = np.load(logdir+'model.npz')
best_before = best_before['model']
for (para, pre) in zip(params, best_before):
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0)
print '------------------', tmp
return
if epoch%1==0:
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
np.savez(logdir+'model-'+str(epoch), model=model)
tmp_start4=time.clock()
if epoch % generatition_frequency == 0:
tail='-'+str(epoch)+'.png'
random_z = np.random.standard_normal((n_batch, n_hidden[-1])).astype(np.float32)
_x_mean, _x = random_generation(random_z)
#print _x.shape
#print _x_mean.shape
image = paramgraphics.mat_to_img(_x.T, dim_input, colorImg=colorImg, scale=True)
image.save(logdir+'samples'+tail, 'PNG')
image = paramgraphics.mat_to_img(_x_mean.T, dim_input, colorImg=colorImg, scale=True)
image.save(logdir+'mean_samples'+tail, 'PNG')
#print 'generation_time', time.clock() - tmp_start4
#print 'one epoch time', time.clock() - tmp_start1
end_time = time.clock()
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
if NaN_count > 0:
print '---------------NaN_count:', NaN_count
with open(logdir+'hook.txt', 'a') as f:
print >>f, '---------------NaN_count:', NaN_count
def mnist_model():
model = network.Network()
model.add(FullyConnected(in_feature=784, out_feature=512), name='fc1')
model.add(LeakyReLU(), name='leaky_relu1')
model.add(FullyConnected(in_feature=512, out_feature=256), name='fc2')
model.add(LeakyReLU(), name='leaky_relu2')
model.add(FullyConnected(in_feature=256, out_feature=256), name='fc3')
model.add(LeakyReLU(), name='leaky_relu3')
model.add(FullyConnected(in_feature=256, out_feature=128), name='fc4')
model.add(LeakyReLU(), name='leaky_relu4')
model.add(FullyConnected(in_feature=128, out_feature=10), name='fc5')
model.add(Softmax(), name='softmax')
model.add_loss(CrossEntropyLoss())
optimizer = SGD(lr=1e-4)
print(model)
traingset = fetch_traingset()
train_images, train_labels = traingset['images'], traingset['labels']
batch_size = 256
training_size = len(train_images)
loss_list = np.zeros((50, int(training_size / batch_size)))
for epoch in range(50):
for i in range(int(training_size / batch_size)):
batch_images = np.array(train_images[i * batch_size:(i + 1) *
batch_size])
batch_labels = np.array(train_labels[i * batch_size:(i + 1) *
batch_size])
batch_labels = one_hot(batch_labels, 10)
_, loss = model.forward(batch_images, batch_labels)
if i % 50 == 0:
loss_list[epoch][i] = loss
print("e:{}, i:{} loss: {}".format(epoch, i, loss))
model.backward()
model.optimize(optimizer)
filename = 'model.data'
f = open(filename, 'wb')
pickle.dump(model, f)
f.close()
loss_fname = 'loss.data'
f = open(loss_fname, 'wb')
pickle.dump(loss_list, f)
f.close()
testset = fetch_testingset()
test_images, test_labels = testset['images'], testset['labels']
test_images = np.array(test_images[:])
test_labels_one_hot = one_hot(test_labels, 10)
y_, test_loss = model.forward(test_images, test_labels_one_hot)
test_labels_pred = testResult2labels(y_)
test_labels = np.array(test_labels)
right_num = np.sum(test_labels == test_labels_pred)
accuracy = 1.0 * right_num / test_labels.shape[0]
print('test accuracy is: ', accuracy)
a = 0
def cva_6layer_dropout_mnist_60000(seed=0,
dropout_flag=1,
drop_inverses_flag=0,
learning_rate=3e-4,
predir=None,
n_batch=144,
dataset='mnist.pkl.gz',
batch_size=500,
nkerns=[20, 50],
n_hidden=[500, 50]):
"""
Implementation of convolutional VA
"""
#cp->cd->cpd->cd->c
nkerns = [32, 32, 64, 64, 64]
drops = [1, 0, 1, 0, 0]
#skerns=[5, 3, 3, 3, 3]
#pools=[2, 1, 1, 2, 1]
#modes=['same']*5
n_hidden = [500, 50]
drop_inverses = [
1,
]
# 28->12->12->5->5/5*5*64->500->50->500->5*5*64/5->5->12->12->28
if dataset == 'mnist.pkl.gz':
dim_input = (28, 28)
colorImg = False
logdir = 'results/supervised/cva/mnist/cva_6layer_mnist_60000' + str(
nkerns) + str(n_hidden) + '_' + str(learning_rate) + '_'
if predir is not None:
logdir += 'pre_'
if dropout_flag == 1:
logdir += ('dropout_' + str(drops) + '_')
if drop_inverses_flag == 1:
logdir += ('inversedropout_' + str(drop_inverses) + '_')
logdir += str(int(time.time())) + '/'
if not os.path.exists(logdir): os.makedirs(logdir)
print 'logdir:', logdir, 'predir', predir
print 'cva_6layer_mnist_60000', nkerns, n_hidden, seed, drops, drop_inverses, dropout_flag, drop_inverses_flag
with open(logdir + 'hook.txt', 'a') as f:
print >> f, 'logdir:', logdir, 'predir', predir
print >> f, 'cva_6layer_mnist_60000', nkerns, n_hidden, seed, drops, drop_inverses, dropout_flag, drop_inverses_flag
datasets = datapy.load_data_gpu_60000(dataset, have_matrix=True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
test_set_x, test_set_y, test_y_matrix = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
random_z = T.matrix('random_z')
drop = T.iscalar('drop')
drop_inverse = T.iscalar('drop_inverse')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x.reshape((batch_size, 1, 28, 28))
recg_layer = []
cnn_output = []
#1
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, 1, 28, 28),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x,
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
#2
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#3
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[1], 12, 12),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#4
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[2], 5, 5),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#5
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[3], 5, 5),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
#1
recg_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=5 * 5 * nkerns[-1],
n_out=n_hidden[0],
activation=activation))
if drops[-1] == 1:
activations.append(recg_layer[-1].drop_output(input=mlp_input_x,
drop=drop,
rng=rng_share))
else:
activations.append(recg_layer[-1].output(input=mlp_input_x))
#stochastic layer
recg_layer.append(
GaussianHidden.GaussianHidden(rng=rng,
input=activations[-1],
n_in=n_hidden[0],
n_out=n_hidden[1],
activation=None))
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[1],
n_out=n_hidden[0],
activation=activation))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
#2
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[0],
n_out=5 * 5 * nkerns[-1],
activation=activation))
if drop_inverses[0] == 1:
z_output.append(gene_layer[-1].drop_output(input=z_output[-1],
drop=drop_inverse,
rng=rng_share))
random_z_output.append(gene_layer[-1].drop_output(
input=random_z_output[-1], drop=drop_inverse, rng=rng_share))
else:
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(
gene_layer[-1].output(input=random_z_output[-1]))
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 5, 5))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 5, 5))
#1
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-1], 5, 5),
filter_shape=(nkerns[-2], nkerns[-1], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(
input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-2], 5, 5),
filter_shape=(nkerns[-3], nkerns[-2], 3,
3),
poolsize=(2, 2),
border_mode='full',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-3], 12,
12),
filter_shape=(nkerns[-4], nkerns[-3], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-4], 12,
12),
filter_shape=(nkerns[-5], nkerns[-4], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#5 stochastic layer
# for the last layer, the nonliearity should be sigmoid to achieve mean of Bernoulli
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-5], 12,
12),
filter_shape=(1, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='full',
activation=nonlinearity.sigmoid))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
gene_layer.append(
NoParamsBernoulliVisiable.NoParamsBernoulliVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=z_output[-1], data=input_x)
# 4-D tensor of random generation
random_x_mean = random_z_output[-1]
random_x = gene_layer[-1].sample_x(rng_share, random_x_mean)
#L = (logpx + logpz - logqz).sum()
cost = ((logpx + recg_layer[-1].logpz - recg_layer[-1].logqz).sum())
px = (logpx.sum())
pz = (recg_layer[-1].logpz.sum())
qz = (-recg_layer[-1].logqz.sum())
params = []
for g in gene_layer:
params += g.params
for r in recg_layer:
params += r.params
gparams = [T.grad(cost, param) for param in params]
weight_decay = 1.0 / n_train_batches
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
#get_optimizer = optimizer.get_adam_optimizer(learning_rate=learning_rate)
get_optimizer = optimizer.get_adam_optimizer_max(learning_rate=l_r,
decay1=0.1,
decay2=0.001,
weight_decay=weight_decay,
epsilon=1e-8)
with open(logdir + 'hook.txt', 'a') as f:
print >> f, 'AdaM', learning_rate, weight_decay
updates = get_optimizer(params, gparams)
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
test_model = theano.function(
inputs=[index],
outputs=cost,
#outputs=layer[-1].errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
#y: test_set_y[index * batch_size:(index + 1) * batch_size],
#y_matrix: test_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
})
validate_model = theano.function(
inputs=[index],
outputs=cost,
#outputs=layer[-1].errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
#y: valid_set_y[index * batch_size:(index + 1) * batch_size],
#y_matrix: valid_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
})
'''
Save parameters and activations
'''
parameters = theano.function(
inputs=[],
outputs=params,
)
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: train_set_y[index * batch_size: (index + 1) * batch_size]
})
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: valid_set_y[index * batch_size: (index + 1) * batch_size]
})
test_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: test_set_y[index * batch_size: (index + 1) * batch_size]
})
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
debug_model = theano.function(
inputs=[index],
outputs=[cost, px, pz, qz],
#updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
#y: train_set_y[index * batch_size: (index + 1) * batch_size],
#y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag),
drop_inverse: np.cast['int32'](drop_inverses_flag)
})
random_generation = theano.function(
inputs=[random_z],
outputs=[random_x_mean.flatten(2),
random_x.flatten(2)],
givens={
#drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
})
train_bound_without_dropout = theano.function(
inputs=[index],
outputs=cost,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
#y: train_set_y[index * batch_size: (index + 1) * batch_size],
#y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
})
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
#y: train_set_y[index * batch_size: (index + 1) * batch_size],
#y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag),
drop_inverse: np.cast['int32'](drop_inverses_flag)
})
##################
# Pretrain MODEL #
##################
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
pre_train = np.load(predir + 'model.npz')
pre_train = pre_train['model']
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
print '------------------', tmp
###############
# TRAIN MODEL #
###############
print '... training'
# early-stopping parameters
patience = 10000 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_bound = -1000000.0
best_iter = 0
test_score = 0.
start_time = time.clock()
NaN_count = 0
epoch = 0
threshold = 0
validation_frequency = 1
generatition_frequency = 10
if predir is not None:
threshold = 0
color.printRed('threshold, ' + str(threshold) +
' generatition_frequency, ' + str(generatition_frequency) +
' validation_frequency, ' + str(validation_frequency))
done_looping = False
n_epochs = 600
decay_epochs = 500
'''
print 'test initialization...'
pre_model = parameters()
for i in xrange(len(pre_model)):
pre_model[i] = np.asarray(pre_model[i])
print pre_model[i].shape, np.mean(pre_model[i]), np.var(pre_model[i])
print 'end test...'
'''
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
minibatch_avg_cost = 0
tmp_start1 = time.clock()
test_epoch = epoch - decay_epochs
if test_epoch > 0 and test_epoch % 10 == 0:
print l_r.get_value()
with open(logdir + 'hook.txt', 'a') as f:
print >> f, l_r.get_value()
l_r.set_value(np.cast['float32'](l_r.get_value() / 3.0))
for minibatch_index in xrange(n_train_batches):
#print minibatch_index
'''
color.printRed('lalala')
xxx = dims(minibatch_index)
print xxx.shape
'''
#print n_train_batches
minibatch_avg_cost += train_model(minibatch_index)
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if math.isnan(minibatch_avg_cost):
NaN_count += 1
color.printRed("NaN detected. Reverting to saved best parameters")
print '---------------NaN_count:', NaN_count
with open(logdir + 'hook.txt', 'a') as f:
print >> f, '---------------NaN_count:', NaN_count
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
print '------------------NaN check:', tmp
with open(logdir + 'hook.txt', 'a') as f:
print >> f, '------------------NaN check:', tmp
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
print model[i].shape, np.mean(model[i]), np.var(model[i])
print np.max(model[i]), np.min(model[i])
print np.all(np.isfinite(model[i])), np.any(np.isnan(model[i]))
with open(logdir + 'hook.txt', 'a') as f:
print >> f, model[i].shape, np.mean(model[i]), np.var(
model[i])
print >> f, np.max(model[i]), np.min(model[i])
print >> f, np.all(np.isfinite(model[i])), np.any(
np.isnan(model[i]))
best_before = np.load(logdir + 'model.npz')
best_before = best_before['model']
for (para, pre) in zip(params, best_before):
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
print '------------------', tmp
return
#print 'optimization_time', time.clock() - tmp_start1
print epoch, 'stochastic training error', minibatch_avg_cost / float(
n_train_batches * batch_size)
with open(logdir + 'hook.txt', 'a') as f:
print >> f, epoch, 'stochastic training error', minibatch_avg_cost / float(
n_train_batches * batch_size)
if epoch % validation_frequency == 0:
tmp_start2 = time.clock()
test_losses = [test_model(i) for i in xrange(n_test_batches)]
this_test_bound = np.mean(test_losses) / float(batch_size)
#tmp = [debug_model(i) for i
# in xrange(n_train_batches)]
#tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
print epoch, 'test bound', this_test_bound
#print tmp
with open(logdir + 'hook.txt', 'a') as f:
print >> f, epoch, 'test bound', this_test_bound
if epoch % 100 == 0:
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
np.savez(logdir + 'model-' + str(epoch), model=model)
for i in xrange(n_train_batches):
if i == 0:
train_features = np.asarray(train_activations(i))
else:
train_features = np.vstack(
(train_features, np.asarray(train_activations(i))))
for i in xrange(n_valid_batches):
if i == 0:
valid_features = np.asarray(valid_activations(i))
else:
valid_features = np.vstack(
(valid_features, np.asarray(valid_activations(i))))
for i in xrange(n_test_batches):
if i == 0:
test_features = np.asarray(test_activations(i))
else:
test_features = np.vstack(
(test_features, np.asarray(test_activations(i))))
np.save(logdir + 'train_features', train_features)
np.save(logdir + 'valid_features', valid_features)
np.save(logdir + 'test_features', test_features)
tmp_start4 = time.clock()
if epoch % generatition_frequency == 0:
tail = '-' + str(epoch) + '.png'
random_z = np.random.standard_normal(
(n_batch, n_hidden[-1])).astype(np.float32)
_x_mean, _x = random_generation(random_z)
#print _x.shape
#print _x_mean.shape
image = paramgraphics.mat_to_img(_x.T,
dim_input,
colorImg=colorImg)
image.save(logdir + 'samples' + tail, 'PNG')
image = paramgraphics.mat_to_img(_x_mean.T,
dim_input,
colorImg=colorImg)
image.save(logdir + 'mean_samples' + tail, 'PNG')
#print 'generation_time', time.clock() - tmp_start4
end_time = time.clock()
print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
if NaN_count > 0:
print '---------------NaN_count:', NaN_count
with open(logdir + 'hook.txt', 'a') as f:
print >> f, '---------------NaN_count:', NaN_count
def c_6layer_svhn_features(learning_rate=0.01,
n_epochs=600,
dataset='svhngcn_var',
batch_size=1000,
dropout_flag=1,
seed=0,
predir=None,
activation=None,
n_batch=625,
weight_decay=1e-4,
super_predir=None,
super_preepoch=None):
"""
Missing data imputation
"""
'''
svhn
'''
n_channels = 3
colorImg = True
dim_w = 32
dim_h = 32
dim_input = (dim_h, dim_w)
n_classes = 10
first_drop = 0.6
if os.environ.has_key('first_drop'):
first_drop = float(os.environ['first_drop'])
last_drop = 1
if os.environ.has_key('last_drop'):
last_drop = float(os.environ['last_drop'])
nkerns_1 = 96
if os.environ.has_key('nkerns_1'):
nkerns_1 = int(os.environ['nkerns_1'])
nkerns_2 = 96
if os.environ.has_key('nkerns_2'):
nkerns_2 = int(os.environ['nkerns_2'])
opt_med = 'mom'
if os.environ.has_key('opt_med'):
opt_med = os.environ['opt_med']
train_logvar = True
if os.environ.has_key('train_logvar'):
train_logvar = bool(int(os.environ['train_logvar']))
dataset = 'svhnlcn'
if os.environ.has_key('dataset'):
dataset = os.environ['dataset']
n_z = 256
if os.environ.has_key('n_z'):
n_z = int(os.environ['n_z'])
#cp->cd->cpd->cd->c
nkerns = [nkerns_1, nkerns_1, nkerns_1, nkerns_2, nkerns_2]
drops = [0, 1, 1, 1, 0, 1]
drop_p = [1, first_drop, first_drop, first_drop, 1, last_drop]
n_hidden = [n_z]
logdir = 'results/supervised/cva/svhn_features/cva_6layer_svhn'
if not os.path.exists(logdir): os.makedirs(logdir)
print 'logdir:', logdir, 'predir', predir
color.printRed('dataset ' + dataset)
datasets = datapy.load_data_svhn(dataset, have_matrix=False)
train_set_x, train_set_y = datasets[0]
test_set_x, test_set_y = datasets[1]
valid_set_x, valid_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
random_z = T.matrix('random_z')
p_label = T.matrix('p_label')
drop = T.iscalar('drop')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x.reshape((batch_size, n_channels, dim_h, dim_w))
recg_layer = []
cnn_output = []
l = []
d = []
#1
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, n_channels, dim_h, dim_w),
filter_shape=(nkerns[0], n_channels, 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x,
drop=drop,
rng=rng_share,
p=drop_p[0]))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
l += [1, 2]
d += [1, 1]
#2
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[0], 16, 16),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[1]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#3
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[1], 16, 16),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[2]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#4
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[2], 8, 8),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[3]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#5
'''
--------------------- (2,2) or (4,4)
'''
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[3], 8, 8),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[4]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
activations.append(mlp_input_x)
#1
'''
---------------------No MLP
'''
'''
recg_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in= 4 * 4 * nkerns[-1],
n_out=n_hidden[0],
activation=activation
))
if drops[-1]==1:
activations.append(recg_layer[-1].drop_output(input=mlp_input_x, drop=drop, rng=rng_share, p=drop_p[-1]))
else:
activations.append(recg_layer[-1].output(input=mlp_input_x))
'''
#stochastic layer
recg_layer.append(
GaussianHidden.GaussianHidden(rng=rng,
input=activations[-1],
n_in=4 * 4 * nkerns[-1],
n_out=n_hidden[0],
activation=None))
l += [1, 2]
d += [1, 1]
l += [1, 2]
d += [1, 1]
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[0],
n_out=4 * 4 * nkerns[-1],
activation=activation))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
l += [1, 2]
d += [1, 1]
#2
'''
gene_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in=n_hidden[0],
n_out = 4*4*nkerns[-1],
activation=activation
))
if drop_inverses[0]==1:
z_output.append(gene_layer[-1].drop_output(input=z_output[-1], drop=drop_inverse, rng=rng_share))
random_z_output.append(gene_layer[-1].drop_output(input=random_z_output[-1], drop=drop_inverse, rng=rng_share))
else:
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output(input=random_z_output[-1]))
'''
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 4, 4))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 4, 4))
#1
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-1], 4, 4),
filter_shape=(nkerns[-2], nkerns[-1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(
input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-2], 8, 8),
filter_shape=(nkerns[-3], nkerns[-2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-3], 8, 8),
filter_shape=(nkerns[-4], nkerns[-3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-4], 16, 16),
filter_shape=(nkerns[-5], nkerns[-4], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#5-1 stochastic layer
# for this layer, the activation is None to get a Guassian mean
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None))
l += [1, 2]
d += [1, 1]
x_mean = gene_layer[-1].output(input=z_output[-1])
random_x_mean = gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch)
#5-2 stochastic layer
# for this layer, the activation is None to get logvar
if train_logvar:
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None))
l += [1, 2]
d += [1, 1]
x_logvar = gene_layer[-1].output(input=z_output[-1])
random_x_logvar = gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch)
else:
x_logvar = theano.shared(
np.ones((batch_size, n_channels, dim_h, dim_w), dtype='float32'))
random_x_logvar = theano.shared(
np.ones((n_batch, n_channels, dim_h, dim_w), dtype='float32'))
gene_layer.append(
NoParamsGaussianVisiable.NoParamsGaussianVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=x_mean, logvar=x_logvar, data=input_x)
random_x = gene_layer[-1].sample_x(rng_share=rng_share,
mean=random_x_mean,
logvar=random_x_logvar)
params = []
for g in gene_layer:
params += g.params
for r in recg_layer:
params += r.params
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: train_set_y[index * batch_size: (index + 1) * batch_size]
})
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: valid_set_y[index * batch_size: (index + 1) * batch_size]
})
test_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#y: test_set_y[index * batch_size: (index + 1) * batch_size]
})
##################
# Pretrain MODEL #
##################
model_epoch = 100
ctype = 'cva'
if os.environ.has_key('model_epoch'):
model_epoch = int(os.environ['model_epoch'])
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
if model_epoch == -1:
pre_train = np.load(predir + 'best-model.npz')
else:
pre_train = np.load(predir + 'model-' + str(model_epoch) + '.npz')
pre_train = pre_train['model']
if ctype == 'cva':
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
elif ctype == 'cmmva':
for (para, pre) in zip(params, pre_train[:-2]):
para.set_value(pre)
else:
exit()
else:
exit()
###############
# TRAIN MODEL #
###############
print 'extract features: valid'
for i in xrange(n_valid_batches):
if i == 0:
valid_features = np.asarray(valid_activations(i))
else:
valid_features = np.vstack(
(valid_features, np.asarray(valid_activations(i))))
#print 'valid'
print 'extract features: test'
for i in xrange(n_test_batches):
if i == 0:
test_features = np.asarray(test_activations(i))
else:
test_features = np.vstack(
(test_features, np.asarray(test_activations(i))))
f = file(logdir + "svhn_features.bin", "wb")
np.save(f, valid_features)
np.save(f, test_features)
f.close()
#print 'test'
print 'extract features: train'
f = file(logdir + "svhn_train_features.bin", "wb")
for i in xrange(n_train_batches):
#print n_train_batches
#print i
train_features = np.asarray(train_activations(i))
np.save(f, train_features)
f.close()
def c_6layer_svhn_imputation(seed=0,
ctype='cva',
pertub_type=5,
pertub_prob=0,
pertub_prob1=16,
visualization_times=20,
denoise_times=200,
predir=None,
n_batch=900,
batch_size=500):
"""
Missing data imputation
"""
'''
svhn
'''
n_channels = 3
colorImg = True
dim_w = 32
dim_h = 32
dim_input = (dim_h, dim_w)
n_classes = 10
first_drop = 0.6
if os.environ.has_key('first_drop'):
first_drop = float(os.environ['first_drop'])
last_drop = 1
if os.environ.has_key('last_drop'):
last_drop = float(os.environ['last_drop'])
nkerns_1 = 96
if os.environ.has_key('nkerns_1'):
nkerns_1 = int(os.environ['nkerns_1'])
nkerns_2 = 96
if os.environ.has_key('nkerns_2'):
nkerns_2 = int(os.environ['nkerns_2'])
opt_med = 'mom'
if os.environ.has_key('opt_med'):
opt_med = os.environ['opt_med']
train_logvar = True
if os.environ.has_key('train_logvar'):
train_logvar = bool(int(os.environ['train_logvar']))
dataset = 'svhnlcn'
if os.environ.has_key('dataset'):
dataset = os.environ['dataset']
n_z = 256
if os.environ.has_key('n_z'):
n_z = int(os.environ['n_z'])
#cp->cd->cpd->cd->c
nkerns = [nkerns_1, nkerns_1, nkerns_1, nkerns_2, nkerns_2]
drops = [0, 1, 1, 1, 0, 1]
drop_p = [1, first_drop, first_drop, first_drop, 1, last_drop]
n_hidden = [n_z]
logdir = 'results/imputation/' + ctype + '/svhn/' + ctype + '_6layer_' + dataset + '_'
logdir += str(int(time.time())) + '/'
if not os.path.exists(logdir): os.makedirs(logdir)
print predir
with open(logdir + 'hook.txt', 'a') as f:
print >> f, predir
color.printRed('dataset ' + dataset)
test_set_x, test_set_x_pertub, pertub_label, pertub_number = datapy.load_pertub_data_svhn(
dirs='data_imputation/',
dataset=dataset,
pertub_type=pertub_type,
pertub_prob=pertub_prob,
pertub_prob1=pertub_prob1)
pixel_max, pixel_min = datapy.load_max_min(dirs='data_imputation/',
dataset=dataset,
pertub_prob=pertub_prob)
# compute number of minibatches for training, validation and testing
#n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
#n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
random_z = T.matrix('random_z')
x_pertub = T.matrix(
'x_pertub') # the data is presented as rasterized images
p_label = T.matrix('p_label')
drop = T.iscalar('drop')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x_pertub.reshape((batch_size, n_channels, dim_h, dim_w))
recg_layer = []
cnn_output = []
l = []
d = []
#1
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, n_channels, dim_h, dim_w),
filter_shape=(nkerns[0], n_channels, 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x,
drop=drop,
rng=rng_share,
p=drop_p[0]))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
l += [1, 2]
d += [1, 1]
#2
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[0], 16, 16),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[1]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#3
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[1], 16, 16),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[2]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#4
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[2], 8, 8),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[3]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
#5
'''
--------------------- (2,2) or (4,4)
'''
recg_layer.append(
ConvMaxPool_GauInit_DNN.ConvMaxPool_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[3], 8, 8),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share,
p=drop_p[4]))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
l += [1, 2]
d += [1, 1]
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
activations.append(mlp_input_x)
#1
'''
---------------------No MLP
'''
'''
recg_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in= 4 * 4 * nkerns[-1],
n_out=n_hidden[0],
activation=activation
))
if drops[-1]==1:
activations.append(recg_layer[-1].drop_output(input=mlp_input_x, drop=drop, rng=rng_share, p=drop_p[-1]))
else:
activations.append(recg_layer[-1].output(input=mlp_input_x))
'''
#stochastic layer
recg_layer.append(
GaussianHidden.GaussianHidden(rng=rng,
input=activations[-1],
n_in=4 * 4 * nkerns[-1],
n_out=n_hidden[0],
activation=None))
l += [1, 2]
d += [1, 1]
l += [1, 2]
d += [1, 1]
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[0],
n_out=4 * 4 * nkerns[-1],
activation=activation))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
l += [1, 2]
d += [1, 1]
#2
'''
gene_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in=n_hidden[0],
n_out = 4*4*nkerns[-1],
activation=activation
))
if drop_inverses[0]==1:
z_output.append(gene_layer[-1].drop_output(input=z_output[-1], drop=drop_inverse, rng=rng_share))
random_z_output.append(gene_layer[-1].drop_output(input=random_z_output[-1], drop=drop_inverse, rng=rng_share))
else:
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output(input=random_z_output[-1]))
'''
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 4, 4))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 4, 4))
#1
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-1], 4, 4),
filter_shape=(nkerns[-2], nkerns[-1], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(
input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-2], 8, 8),
filter_shape=(nkerns[-3], nkerns[-2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-3], 8, 8),
filter_shape=(nkerns[-4], nkerns[-3], 3, 3),
poolsize=(2, 2),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-4], 16, 16),
filter_shape=(nkerns[-5], nkerns[-4], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
l += [1, 2]
d += [1, 1]
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#5-1 stochastic layer
# for this layer, the activation is None to get a Guassian mean
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None))
l += [1, 2]
d += [1, 1]
x_mean = gene_layer[-1].output(input=z_output[-1])
random_x_mean = gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch)
#5-2 stochastic layer
# for this layer, the activation is None to get logvar
if train_logvar:
gene_layer.append(
UnpoolConvNon_GauInit_DNN.UnpoolConvNon_GauInit_DNN(
rng,
image_shape=(batch_size, nkerns[-5], 16, 16),
filter_shape=(n_channels, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='same',
activation=None))
l += [1, 2]
d += [1, 1]
x_logvar = gene_layer[-1].output(input=z_output[-1])
random_x_logvar = gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch)
else:
x_logvar = theano.shared(
np.ones((batch_size, n_channels, dim_h, dim_w), dtype='float32'))
random_x_logvar = theano.shared(
np.ones((n_batch, n_channels, dim_h, dim_w), dtype='float32'))
gene_layer.append(
NoParamsGaussianVisiable.NoParamsGaussianVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=x_mean, logvar=x_logvar, data=input_x)
random_x = gene_layer[-1].sample_x(rng_share=rng_share,
mean=random_x_mean,
logvar=random_x_logvar)
x_denoised = p_label * x + (1 - p_label) * x_mean.flatten(2)
mse = ((x - x_denoised)**2).sum() / pertub_number
params = []
for g in gene_layer:
params += g.params
for r in recg_layer:
params += r.params
'''
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x_pertub: train_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
'''
'''
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x_pertub: valid_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0)
}
)
'''
test_activations = theano.function(inputs=[x_pertub],
outputs=T.concatenate(activations,
axis=1),
givens={drop: np.cast['int32'](0)})
imputation_model = theano.function(
inputs=[index, x_pertub],
outputs=[x_denoised, mse],
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
p_label: pertub_label[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
})
##################
# Pretrain MODEL #
##################
model_epoch = 100
if os.environ.has_key('model_epoch'):
model_epoch = int(os.environ['model_epoch'])
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
if model_epoch == -1:
pre_train = np.load(predir + 'best-model.npz')
else:
pre_train = np.load(predir + 'model-' + str(model_epoch) + '.npz')
pre_train = pre_train['model']
if ctype == 'cva':
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
elif ctype == 'cmmva':
for (para, pre) in zip(params, pre_train[:-2]):
para.set_value(pre)
else:
exit()
else:
exit()
###############
# TRAIN MODEL #
###############
print '... training'
scale = False
epoch = 0
n_visualization = 900
pixel_max = pixel_max[:n_visualization]
pixel_min = pixel_min[:n_visualization]
output = np.ones((n_visualization, visualization_times + 2,
n_channels * dim_input[0] * dim_input[1]))
output[:, 0, :] = test_set_x.get_value()[:n_visualization, :]
output[:, 1, :] = test_set_x_pertub.get_value()[:n_visualization, :]
image = paramgraphics.mat_to_img(paramgraphics.scale_max_min(
output[:, 0, :].T, pixel_max, pixel_min),
dim_input,
colorImg=colorImg,
scale=scale)
image.save(logdir + 'data.png', 'PNG')
image = paramgraphics.mat_to_img(paramgraphics.scale_max_min(
output[:, 1, :].T, pixel_max, pixel_min),
dim_input,
colorImg=colorImg,
scale=scale)
image.save(logdir + 'data_pertub.png', 'PNG')
tmp = test_set_x_pertub.get_value()
while epoch < denoise_times:
epoch = epoch + 1
for i in xrange(n_test_batches):
d, m = imputation_model(i,
tmp[i * batch_size:(i + 1) * batch_size])
tmp[i * batch_size:(i + 1) * batch_size] = np.asarray(d)
if epoch <= visualization_times:
output[:, epoch + 1, :] = tmp[:n_visualization, :]
image = paramgraphics.mat_to_img(paramgraphics.scale_max_min(
tmp[:n_visualization, :].T, pixel_max, pixel_min),
dim_input,
colorImg=colorImg,
scale=scale)
image.save(logdir + 'procedure-' + str(epoch) + '.png', 'PNG')
np.savez(logdir + 'procedure-' + str(epoch), tmp=tmp)
'''
image = paramgraphics.mat_to_img((output.reshape(-1,32*32*3)).T, dim_input, colorImg=colorImg, tile_shape=(n_visualization,22), scale=scale)
image.save(logdir+'output.png', 'PNG')
np.savez(logdir+'output', output=output)
'''
'''
def deep_cnn_6layer_mnist_50000(learning_rate=3e-4,
n_epochs=250,
dataset='mnist.pkl.gz',
batch_size=500,
dropout_flag=0,
seed=0,
activation=None):
#cp->cd->cpd->cd->c
nkerns = [32, 32, 64, 64, 64]
drops = [1, 0, 1, 0, 0]
#skerns=[5, 3, 3, 3, 3]
#pools=[2, 1, 1, 2, 1]
#modes=['same']*5
n_hidden = [500]
logdir = 'results/supervised/cnn/mnist/deep_cnn_6layer_50000_' + str(
nkerns) + str(drops) + str(n_hidden) + '_' + str(
learning_rate) + '_' + str(int(time.time())) + '/'
if dropout_flag == 1:
logdir = 'results/supervised/cnn/mnist/deep_cnn_6layer_50000_' + str(
nkerns) + str(drops) + str(n_hidden) + '_' + str(
learning_rate) + '_dropout_' + str(int(time.time())) + '/'
if not os.path.exists(logdir): os.makedirs(logdir)
print 'logdir:', logdir
print 'deep_cnn_6layer_mnist_50000_', nkerns, n_hidden, drops, seed, dropout_flag
with open(logdir + 'hook.txt', 'a') as f:
print >> f, 'logdir:', logdir
print >> f, 'deep_cnn_6layer_mnist_50000_', nkerns, n_hidden, drops, seed, dropout_flag
rng = np.random.RandomState(0)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
'''
'''
datasets = datapy.load_data_gpu_60000(dataset, have_matrix=True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
test_set_x, test_set_y, test_y_matrix = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0]
n_test_batches = test_set_x.get_value(borrow=True).shape[0]
n_train_batches /= batch_size
n_valid_batches /= batch_size
n_test_batches /= batch_size
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
# start-snippet-1
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
'''
dropout
'''
drop = T.iscalar('drop')
y_matrix = T.imatrix(
'y_matrix') # labels, presented as 2D matrix of int labels
print '... building the model'
layer0_input = x.reshape((batch_size, 1, 28, 28))
if activation == 'nonlinearity.relu':
activation = nonlinearity.relu
elif activation == 'nonlinearity.tanh':
activation = nonlinearity.tanh
elif activation == 'nonlinearity.softplus':
activation = nonlinearity.softplus
recg_layer = []
cnn_output = []
#1
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, 1, 28, 28),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(layer0_input,
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(layer0_input))
#2
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#3
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[1], 12, 12),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#4
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[2], 5, 5),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#5
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[3], 5, 5),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
mlp_input = cnn_output[-1].flatten(2)
recg_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=nkerns[4] * 5 * 5,
n_out=500,
activation=activation))
feature = recg_layer[-1].drop_output(mlp_input, drop=drop, rng=rng_share)
# classify the values of the fully-connected sigmoidal layer
classifier = Pegasos.Pegasos(input=feature,
rng=rng,
n_in=500,
n_out=10,
weight_decay=0,
loss=1)
# the cost we minimize during training is the NLL of the model
cost = classifier.hinge_loss(10, y, y_matrix) * batch_size
weight_decay = 1.0 / n_train_batches
# create a list of all model parameters to be fit by gradient descent
params = []
for r in recg_layer:
params += r.params
params += classifier.params
# create a list of gradients for all model parameters
grads = T.grad(cost, params)
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
get_optimizer = optimizer.get_adam_optimizer_min(learning_rate=l_r,
decay1=0.1,
decay2=0.001,
weight_decay=weight_decay)
updates = get_optimizer(params, grads)
'''
Save parameters and activations
'''
parameters = theano.function(
inputs=[],
outputs=params,
)
# create a function to compute the mistakes that are made by the model
test_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0)
})
validate_model = theano.function(
inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0)
})
train_model_average = theano.function(
inputs=[index],
outputs=[cost, classifier.errors(y)],
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:
train_y_matrix[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag)
})
train_model = theano.function(
inputs=[index],
outputs=[cost, classifier.errors(y)],
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:
train_y_matrix[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag)
})
print '... training'
# early-stopping parameters
patience = n_train_batches * 100 # look as this many examples regardless
patience_increase = 2 # wait this much longer when a new best is
# found
improvement_threshold = 0.995 # a relative improvement of this much is
# considered significant
validation_frequency = min(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
best_validation_loss = np.inf
best_test_score = np.inf
test_score = 0.
start_time = time.clock()
epoch = 0
decay_epochs = 150
while (epoch < n_epochs):
epoch = epoch + 1
tmp1 = time.clock()
minibatch_avg_cost = 0
train_error = 0
for minibatch_index in xrange(n_train_batches):
co, te = train_model(minibatch_index)
minibatch_avg_cost += co
train_error += te
#print minibatch_avg_cost
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
test_epoch = epoch - decay_epochs
if test_epoch > 0 and test_epoch % 10 == 0:
print l_r.get_value()
with open(logdir + 'hook.txt', 'a') as f:
print >> f, l_r.get_value()
l_r.set_value(np.cast['float32'](l_r.get_value() / 3.0))
# compute zero-one loss on validation set
validation_losses = [
validate_model(i) for i in xrange(n_valid_batches)
]
this_validation_loss = np.mean(validation_losses)
this_test_losses = [
test_model(i) for i in xrange(n_test_batches)
]
this_test_score = np.mean(this_test_losses)
train_thing = [
train_model_average(i) for i in xrange(n_train_batches)
]
train_thing = np.mean(train_thing, axis=0)
print epoch, 'hinge loss and training error', train_thing
with open(logdir + 'hook.txt', 'a') as f:
print >> f, epoch, 'hinge loss and training error', train_thing
if this_test_score < best_test_score:
best_test_score = this_test_score
print(
'epoch %i, minibatch %i/%i, validation error %f %%, test error %f %%'
% (epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100, this_test_score * 100.))
with open(logdir + 'hook.txt', 'a') as f:
print >> f, (
'epoch %i, minibatch %i/%i, validation error %f %%, test error %f %%'
% (epoch, minibatch_index + 1, n_train_batches,
this_validation_loss * 100, this_test_score * 100.))
# if we got the best validation score until now
if this_validation_loss < best_validation_loss:
#improve patience if loss improvement is good enough
if this_validation_loss < best_validation_loss * \
improvement_threshold:
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
# test it on the test set
test_losses = [
test_model(i) for i in xrange(n_test_batches)
]
test_score = np.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of'
' best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
with open(logdir + 'hook.txt', 'a') as f:
print >> f, (
(' epoch %i, minibatch %i/%i, test error of'
' best model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if epoch % 50 == 0:
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
np.savez(logdir + 'model-' + str(epoch), model=model)
print 'hinge loss and training error', minibatch_avg_cost / float(
n_train_batches), train_error / float(n_train_batches)
print 'time', time.clock() - tmp1
with open(logdir + 'hook.txt', 'a') as f:
print >> f, 'hinge loss and training error', minibatch_avg_cost / float(
n_train_batches), train_error / float(n_train_batches)
print >> f, 'time', time.clock() - tmp1
end_time = time.clock()
print 'The code run for %d epochs, with %f epochs/sec' % (
epoch, 1. * epoch / (end_time - start_time))
print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] +
' ran for %.1fs' % ((end_time - start_time)))
def cmmva_6layer_dropout_mnist_60000(seed=0, start_layer=0, end_layer=1, dropout_flag=1, drop_inverses_flag=0, learning_rate=3e-5, predir=None, n_batch=144,
dataset='mnist.pkl.gz', batch_size=500, nkerns=[20, 50], n_hidden=[500, 50]):
"""
Implementation of convolutional MMVA
"""
#cp->cd->cpd->cd->c
nkerns=[32, 32, 64, 64, 64]
drops=[1, 0, 1, 0, 0, 1]
#skerns=[5, 3, 3, 3, 3]
#pools=[2, 1, 1, 2, 1]
#modes=['same']*5
n_hidden=[500, 50]
drop_inverses=[1,]
# 28->12->12->5->5/5*5*64->500->50->500->5*5*64/5->5->12->12->28
if dataset=='mnist.pkl.gz':
dim_input=(28, 28)
colorImg=False
D = 1.0
C = 1.0
if os.environ.has_key('C'):
C = np.cast['float32'](float((os.environ['C'])))
if os.environ.has_key('D'):
D = np.cast['float32'](float((os.environ['D'])))
color.printRed('D '+str(D)+' C '+str(C))
logdir = 'results/supervised/cmmva/mnist/cmmva_6layer_60000_'+str(nkerns)+str(n_hidden)+'_D_'+str(D)+'_C_'+str(C)+'_'+str(learning_rate)+'_'
if predir is not None:
logdir +='pre_'
if dropout_flag == 1:
logdir += ('dropout_'+str(drops)+'_')
if drop_inverses_flag==1:
logdir += ('inversedropout_'+str(drop_inverses)+'_')
logdir += str(int(time.time()))+'/'
if not os.path.exists(logdir): os.makedirs(logdir)
print 'logdir:', logdir, 'predir', predir
print 'cmmva_6layer_mnist_60000', nkerns, n_hidden, seed, drops, drop_inverses, dropout_flag, drop_inverses_flag
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'logdir:', logdir, 'predir', predir
print >>f, 'cmmva_6layer_mnist_60000', nkerns, n_hidden, seed, drops, drop_inverses, dropout_flag, drop_inverses_flag
datasets = datapy.load_data_gpu_60000(dataset, have_matrix=True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
test_set_x, test_set_y, test_y_matrix = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
y_matrix = T.imatrix('y_matrix')
random_z = T.matrix('random_z')
drop = T.iscalar('drop')
drop_inverse = T.iscalar('drop_inverse')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x.reshape((batch_size, 1, 28, 28))
recg_layer = []
cnn_output = []
#1
recg_layer.append(ConvMaxPool.ConvMaxPool(
rng,
image_shape=(batch_size, 1, 28, 28),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
border_mode='valid',
activation=activation
))
if drops[0]==1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x, drop=drop, rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
#2
recg_layer.append(ConvMaxPool.ConvMaxPool(
rng,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
if drops[1]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#3
recg_layer.append(ConvMaxPool.ConvMaxPool(
rng,
image_shape=(batch_size, nkerns[1], 12, 12),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='valid',
activation=activation
))
if drops[2]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#4
recg_layer.append(ConvMaxPool.ConvMaxPool(
rng,
image_shape=(batch_size, nkerns[2], 5, 5),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
if drops[3]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#5
recg_layer.append(ConvMaxPool.ConvMaxPool(
rng,
image_shape=(batch_size, nkerns[3], 5, 5),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
if drops[4]==1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1], drop=drop, rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
#1
recg_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in= 5 * 5 * nkerns[-1],
n_out=n_hidden[0],
activation=activation
))
if drops[-1]==1:
activations.append(recg_layer[-1].drop_output(input=mlp_input_x, drop=drop, rng=rng_share))
else:
activations.append(recg_layer[-1].output(input=mlp_input_x))
features = T.concatenate(activations[start_layer:end_layer], axis=1)
color.printRed('feature dimension: '+str(np.sum(n_hidden[start_layer:end_layer])))
classifier = Pegasos.Pegasos(
input= features,
rng=rng,
n_in=np.sum(n_hidden[start_layer:end_layer]),
n_out=10,
weight_decay=0,
loss=1,
std=1e-2
)
recg_layer.append(GaussianHidden.GaussianHidden(
rng=rng,
input=activations[-1],
n_in=n_hidden[0],
n_out = n_hidden[1],
activation=None
))
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in=n_hidden[1],
n_out = n_hidden[0],
activation=activation
))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
#2
gene_layer.append(FullyConnected.FullyConnected(
rng=rng,
n_in=n_hidden[0],
n_out = 5*5*nkerns[-1],
activation=activation
))
if drop_inverses[0]==1:
z_output.append(gene_layer[-1].drop_output(input=z_output[-1], drop=drop_inverse, rng=rng_share))
random_z_output.append(gene_layer[-1].drop_output(input=random_z_output[-1], drop=drop_inverse, rng=rng_share))
else:
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output(input=random_z_output[-1]))
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 5, 5))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 5, 5))
#1
gene_layer.append(UnpoolConvNon.UnpoolConvNon(
rng,
image_shape=(batch_size, nkerns[-1], 5, 5),
filter_shape=(nkerns[-2], nkerns[-1], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(UnpoolConvNon.UnpoolConvNon(
rng,
image_shape=(batch_size, nkerns[-2], 5, 5),
filter_shape=(nkerns[-3], nkerns[-2], 3, 3),
poolsize=(2, 2),
border_mode='full',
activation=activation
))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(UnpoolConvNon.UnpoolConvNon(
rng,
image_shape=(batch_size, nkerns[-3], 12, 12),
filter_shape=(nkerns[-4], nkerns[-3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(UnpoolConvNon.UnpoolConvNon(
rng,
image_shape=(batch_size, nkerns[-4], 12, 12),
filter_shape=(nkerns[-5], nkerns[-4], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation
))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
#5 stochastic layer
# for the last layer, the nonliearity should be sigmoid to achieve mean of Bernoulli
gene_layer.append(UnpoolConvNon.UnpoolConvNon(
rng,
image_shape=(batch_size, nkerns[-5], 12, 12),
filter_shape=(1, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='full',
activation=nonlinearity.sigmoid
))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(input=random_z_output[-1], n_batch=n_batch))
gene_layer.append(NoParamsBernoulliVisiable.NoParamsBernoulliVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=z_output[-1], data=input_x)
# 4-D tensor of random generation
random_x_mean = random_z_output[-1]
random_x = gene_layer[-1].sample_x(rng_share, random_x_mean)
#L = (logpx + logpz - logqz).sum()
lowerbound = (
(logpx + recg_layer[-1].logpz - recg_layer[-1].logqz).sum()
)
hinge_loss = classifier.hinge_loss(10, y, y_matrix) * batch_size
#
# D is redundent, you could just set D = 1 and tune C and weight decay parameters
# beacuse AdaM is scale-invariant
#
cost = D * lowerbound - C * hinge_loss #- classifier.L2_reg
px = (logpx.sum())
pz = (recg_layer[-1].logpz.sum())
qz = (- recg_layer[-1].logqz.sum())
params=[]
for g in gene_layer:
params+=g.params
for r in recg_layer:
params+=r.params
params+=classifier.params
gparams = [T.grad(cost, param) for param in params]
weight_decay=1.0/n_train_batches
epsilon=1e-8
#get_optimizer = optimizer.get_adam_optimizer(learning_rate=learning_rate)
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
get_optimizer = optimizer.get_adam_optimizer_max(learning_rate=l_r,
decay1=0.1, decay2=0.001, weight_decay=weight_decay, epsilon=epsilon)
with open(logdir+'hook.txt', 'a') as f:
print >>f, 'AdaM', learning_rate, weight_decay, epsilon
updates = get_optimizer(params,gparams)
# compiling a Theano function that computes the mistakes that are made
# by the model on a minibatch
test_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
#outputs=layer[-1].errors(y),
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
y: test_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix: test_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
}
)
validate_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
#outputs=layer[-1].errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix: valid_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
}
)
'''
Save parameters and activations
'''
parameters = theano.function(
inputs=[],
outputs=params,
)
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: valid_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: valid_set_y[index * batch_size: (index + 1) * batch_size]
}
)
test_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x: test_set_x[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
#drop_inverse: np.cast['int32'](0)
#y: test_set_y[index * batch_size: (index + 1) * batch_size]
}
)
# compiling a Theano function `train_model` that returns the cost, but
# in the same time updates the parameter of the model based on the rules
# defined in `updates`
debug_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, px, pz, qz, hinge_loss, cost],
#updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag),
drop_inverse: np.cast['int32'](drop_inverses_flag)
}
)
random_generation = theano.function(
inputs=[random_z],
outputs=[random_x_mean.flatten(2), random_x.flatten(2)],
givens={
#drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
}
)
train_bound_without_dropout = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
}
)
train_model = theano.function(
inputs=[index],
outputs=[classifier.errors(y), lowerbound, hinge_loss, cost],
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
y: train_set_y[index * batch_size: (index + 1) * batch_size],
y_matrix: train_y_matrix[index * batch_size: (index + 1) * batch_size],
drop: np.cast['int32'](dropout_flag),
drop_inverse: np.cast['int32'](drop_inverses_flag)
}
)
# end-snippet-5
##################
# Pretrain MODEL #
##################
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
pre_train = np.load(predir+'model.npz')
pre_train = pre_train['model']
# params include w and b, exclude it
for (para, pre) in zip(params[:-2], pre_train):
#print pre.shape
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
print '------------------', tmp[1:5]
# valid_error test_error epochs
predy_test_stats = [1, 1, 0]
predy_valid_stats = [1, 1, 0]
best_validation_bound = -1000000.0
best_iter = 0
test_score = 0.
start_time = time.clock()
NaN_count = 0
epoch = 0
threshold = 0
validation_frequency = 1
generatition_frequency = 10
if predir is not None:
threshold = 0
color.printRed('threshold, '+str(threshold) +
' generatition_frequency, '+str(generatition_frequency)
+' validation_frequency, '+str(validation_frequency))
done_looping = False
decay_epochs=500
n_epochs=600
'''
print 'test initialization...'
pre_model = parameters()
for i in xrange(len(pre_model)):
pre_model[i] = np.asarray(pre_model[i])
print pre_model[i].shape, np.mean(pre_model[i]), np.var(pre_model[i])
print 'end test...'
'''
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
train_error = 0
train_lowerbound = 0
train_hinge_loss = 0
train_obj = 0
test_epoch = epoch - decay_epochs
if test_epoch > 0 and test_epoch % 10 == 0:
print l_r.get_value()
with open(logdir+'hook.txt', 'a') as f:
print >>f,l_r.get_value()
l_r.set_value(np.cast['float32'](l_r.get_value()/3.0))
tmp_start1 = time.clock()
for minibatch_index in xrange(n_train_batches):
#print n_train_batches
e, l, h, o = train_model(minibatch_index)
train_error += e
train_lowerbound += l
train_hinge_loss += h
train_obj += o
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if math.isnan(train_lowerbound):
NaN_count+=1
color.printRed("NaN detected. Reverting to saved best parameters")
print '---------------NaN_count:', NaN_count
with open(logdir+'hook.txt', 'a') as f:
print >>f, '---------------NaN_count:', NaN_count
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
tmp[0]*=batch_size
print '------------------NaN check:', tmp
with open(logdir+'hook.txt', 'a') as f:
print >>f, '------------------NaN check:', tmp
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
print model[i].shape, np.mean(model[i]), np.var(model[i])
print np.max(model[i]), np.min(model[i])
print np.all(np.isfinite(model[i])), np.any(np.isnan(model[i]))
with open(logdir+'hook.txt', 'a') as f:
print >>f, model[i].shape, np.mean(model[i]), np.var(model[i])
print >>f, np.max(model[i]), np.min(model[i])
print >>f, np.all(np.isfinite(model[i])), np.any(np.isnan(model[i]))
best_before = np.load(logdir+'model.npz')
best_before = best_before['model']
for (para, pre) in zip(params, best_before):
para.set_value(pre)
tmp = [debug_model(i) for i in xrange(n_train_batches)]
tmp = (np.asarray(tmp)).mean(axis=0) / float(batch_size)
tmp[0]*=batch_size
print '------------------', tmp
continue
n_train=n_train_batches*batch_size
#print 'optimization_time', time.clock() - tmp_start1
print epoch, 'stochastic training error', train_error / float(batch_size), train_lowerbound / float(n_train), train_hinge_loss / float(n_train), train_obj / float(n_train)
with open(logdir+'hook.txt', 'a') as f:
print >>f, epoch, 'stochastic training error', train_error / float(batch_size), train_lowerbound / float(n_train), train_hinge_loss / float(n_train), train_obj / float(n_train)
if epoch % validation_frequency == 0:
tmp_start2 = time.clock()
# compute zero-one loss on validation set
#train_stats = [train_bound_without_dropout(i) for i
# in xrange(n_train_batches)]
#this_train_stats = np.mean(train_stats, axis=0)
#this_train_stats[1:] = this_train_stats[1:]/ float(batch_size)
test_stats = [test_model(i) for i in xrange(n_test_batches)]
this_test_stats = np.mean(test_stats, axis=0)
this_test_stats[1:] = this_test_stats[1:]/ float(batch_size)
print epoch, 'test error', this_test_stats
with open(logdir+'hook.txt', 'a') as f:
print >>f, epoch, 'test error', this_test_stats
if epoch%100==0:
model = parameters()
for i in xrange(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
#print model[i].shape, np.mean(model[i]), np.var(model[i])
np.savez(logdir+'model-'+str(epoch), model=model)
tmp_start4=time.clock()
if epoch % generatition_frequency == 0:
tail='-'+str(epoch)+'.png'
random_z = np.random.standard_normal((n_batch, n_hidden[-1])).astype(np.float32)
_x_mean, _x = random_generation(random_z)
#print _x.shape
#print _x_mean.shape
image = paramgraphics.mat_to_img(_x.T, dim_input, colorImg=colorImg)
image.save(logdir+'samples'+tail, 'PNG')
image = paramgraphics.mat_to_img(_x_mean.T, dim_input, colorImg=colorImg)
image.save(logdir+'mean_samples'+tail, 'PNG')
#print 'generation_time', time.clock() - tmp_start4
end_time = time.clock()
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
if NaN_count > 0:
print '---------------NaN_count:', NaN_count
with open(logdir+'hook.txt', 'a') as f:
print >>f, '---------------NaN_count:', NaN_count
def c_6layer_mnist_imputation(seed=0,
ctype='cva',
pertub_type=3,
pertub_prob=6,
pertub_prob1=14,
visualization_times=20,
denoise_times=200,
predir=None,
n_batch=144,
dataset='mnist.pkl.gz',
batch_size=500):
"""
Missing data imputation
"""
#cp->cd->cpd->cd->c
nkerns = [32, 32, 64, 64, 64]
drops = [0, 0, 0, 0, 0, 1]
#skerns=[5, 3, 3, 3, 3]
#pools=[2, 1, 1, 2, 1]
#modes=['same']*5
n_hidden = [500, 50]
drop_inverses = [
1,
]
# 28->12->12->5->5/5*5*64->500->50->500->5*5*64/5->5->12->12->28
if dataset == 'mnist.pkl.gz':
dim_input = (28, 28)
colorImg = False
logdir = 'results/imputation/' + ctype + '/mnist/' + ctype + '_6layer_mnist_' + str(
pertub_type) + '_' + str(pertub_prob) + '_' + str(
pertub_prob1) + '_' + str(denoise_times) + '_'
logdir += str(int(time.time())) + '/'
if not os.path.exists(logdir): os.makedirs(logdir)
print predir
with open(logdir + 'hook.txt', 'a') as f:
print >> f, predir
train_set_x, test_set_x, test_set_x_pertub, pertub_label, pertub_number = datapy.load_pertub_data(
dirs='data_imputation/',
pertub_type=pertub_type,
pertub_prob=pertub_prob,
pertub_prob1=pertub_prob1)
datasets = datapy.load_data_gpu(dataset, have_matrix=True)
_, _, _ = datasets[0]
valid_set_x, _, _ = datasets[1]
_, _, _ = datasets[2]
# compute number of minibatches for training, validation and testing
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x')
x_pertub = T.matrix(
'x_pertub') # the data is presented as rasterized images
p_label = T.matrix('p_label')
random_z = T.matrix('random_z')
drop = T.iscalar('drop')
drop_inverse = T.iscalar('drop_inverse')
activation = nonlinearity.relu
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
input_x = x_pertub.reshape((batch_size, 1, 28, 28))
recg_layer = []
cnn_output = []
#1
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, 1, 28, 28),
filter_shape=(nkerns[0], 1, 5, 5),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[0] == 1:
cnn_output.append(recg_layer[-1].drop_output(input=input_x,
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(input=input_x))
#2
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[0], 12, 12),
filter_shape=(nkerns[1], nkerns[0], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[1] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#3
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[1], 12, 12),
filter_shape=(nkerns[2], nkerns[1], 3, 3),
poolsize=(2, 2),
border_mode='valid',
activation=activation))
if drops[2] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#4
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[2], 5, 5),
filter_shape=(nkerns[3], nkerns[2], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[3] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
#5
recg_layer.append(
ConvMaxPool.ConvMaxPool(rng,
image_shape=(batch_size, nkerns[3], 5, 5),
filter_shape=(nkerns[4], nkerns[3], 3, 3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
if drops[4] == 1:
cnn_output.append(recg_layer[-1].drop_output(cnn_output[-1],
drop=drop,
rng=rng_share))
else:
cnn_output.append(recg_layer[-1].output(cnn_output[-1]))
mlp_input_x = cnn_output[-1].flatten(2)
activations = []
#1
recg_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=5 * 5 * nkerns[-1],
n_out=n_hidden[0],
activation=activation))
if drops[-1] == 1:
activations.append(recg_layer[-1].drop_output(input=mlp_input_x,
drop=drop,
rng=rng_share))
else:
activations.append(recg_layer[-1].output(input=mlp_input_x))
#stochastic layer
recg_layer.append(
GaussianHidden.GaussianHidden(rng=rng,
input=activations[-1],
n_in=n_hidden[0],
n_out=n_hidden[1],
activation=None))
z = recg_layer[-1].sample_z(rng_share)
gene_layer = []
z_output = []
random_z_output = []
#1
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[1],
n_out=n_hidden[0],
activation=activation))
z_output.append(gene_layer[-1].output(input=z))
random_z_output.append(gene_layer[-1].output(input=random_z))
#2
gene_layer.append(
FullyConnected.FullyConnected(rng=rng,
n_in=n_hidden[0],
n_out=5 * 5 * nkerns[-1],
activation=activation))
if drop_inverses[0] == 1:
z_output.append(gene_layer[-1].drop_output(input=z_output[-1],
drop=drop_inverse,
rng=rng_share))
random_z_output.append(gene_layer[-1].drop_output(
input=random_z_output[-1], drop=drop_inverse, rng=rng_share))
else:
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(
gene_layer[-1].output(input=random_z_output[-1]))
input_z = z_output[-1].reshape((batch_size, nkerns[-1], 5, 5))
input_random_z = random_z_output[-1].reshape((n_batch, nkerns[-1], 5, 5))
#1
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-1], 5, 5),
filter_shape=(nkerns[-2], nkerns[-1], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=input_z))
random_z_output.append(gene_layer[-1].output_random_generation(
input=input_random_z, n_batch=n_batch))
#2
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-2], 5, 5),
filter_shape=(nkerns[-3], nkerns[-2], 3,
3),
poolsize=(2, 2),
border_mode='full',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#3
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-3], 12,
12),
filter_shape=(nkerns[-4], nkerns[-3], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#4
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-4], 12,
12),
filter_shape=(nkerns[-5], nkerns[-4], 3,
3),
poolsize=(1, 1),
border_mode='same',
activation=activation))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
#5 stochastic layer
# for the last layer, the nonliearity should be sigmoid to achieve mean of Bernoulli
gene_layer.append(
UnpoolConvNon.UnpoolConvNon(rng,
image_shape=(batch_size, nkerns[-5], 12,
12),
filter_shape=(1, nkerns[-5], 5, 5),
poolsize=(2, 2),
border_mode='full',
activation=nonlinearity.sigmoid))
z_output.append(gene_layer[-1].output(input=z_output[-1]))
random_z_output.append(gene_layer[-1].output_random_generation(
input=random_z_output[-1], n_batch=n_batch))
gene_layer.append(
NoParamsBernoulliVisiable.NoParamsBernoulliVisiable(
#rng=rng,
#mean=z_output[-1],
#data=input_x,
))
logpx = gene_layer[-1].logpx(mean=z_output[-1], data=input_x)
# 4-D tensor of random generation
random_x_mean = random_z_output[-1]
random_x = gene_layer[-1].sample_x(rng_share, random_x_mean)
x_denoised = z_output[-1].flatten(2)
x_denoised = p_label * x + (1 - p_label) * x_denoised
mse = ((x - x_denoised)**2).sum() / pertub_number
params = []
for g in gene_layer:
params += g.params
for r in recg_layer:
params += r.params
train_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x_pertub: train_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0)
})
valid_activations = theano.function(
inputs=[index],
outputs=T.concatenate(activations, axis=1),
givens={
x_pertub: valid_set_x[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0)
})
test_activations = theano.function(inputs=[x_pertub],
outputs=T.concatenate(activations,
axis=1),
givens={drop: np.cast['int32'](0)})
imputation_model = theano.function(
inputs=[index, x_pertub],
outputs=[x_denoised, mse],
givens={
x: test_set_x[index * batch_size:(index + 1) * batch_size],
p_label: pertub_label[index * batch_size:(index + 1) * batch_size],
drop: np.cast['int32'](0),
drop_inverse: np.cast['int32'](0)
})
##################
# Pretrain MODEL #
##################
model_epoch = 600
if os.environ.has_key('model_epoch'):
model_epoch = int(os.environ['model_epoch'])
if predir is not None:
color.printBlue('... setting parameters')
color.printBlue(predir)
if model_epoch == -1:
pre_train = np.load(predir + 'best-model.npz')
else:
pre_train = np.load(predir + 'model-' + str(model_epoch) + '.npz')
pre_train = pre_train['model']
if ctype == 'cva':
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
elif ctype == 'cmmva':
for (para, pre) in zip(params, pre_train[:-2]):
para.set_value(pre)
else:
exit()
else:
exit()
###############
# TRAIN MODEL #
###############
print '... training'
epoch = 0
n_visualization = 100
output = np.ones((n_visualization, visualization_times + 2, 784))
output[:, 0, :] = test_set_x.get_value()[:n_visualization, :]
output[:, 1, :] = test_set_x_pertub.get_value()[:n_visualization, :]
image = paramgraphics.mat_to_img(output[:, 0, :].T,
dim_input,
colorImg=colorImg)
image.save(logdir + 'data.png', 'PNG')
image = paramgraphics.mat_to_img(output[:, 1, :].T,
dim_input,
colorImg=colorImg)
image.save(logdir + 'data_pertub.png', 'PNG')
tmp = test_set_x_pertub.get_value()
while epoch < denoise_times:
epoch = epoch + 1
this_mse = 0
for i in xrange(n_test_batches):
d, m = imputation_model(i,
tmp[i * batch_size:(i + 1) * batch_size])
tmp[i * batch_size:(i + 1) * batch_size] = np.asarray(d)
this_mse += m
if epoch <= visualization_times:
output[:, epoch + 1, :] = tmp[:n_visualization, :]
print epoch, this_mse
with open(logdir + 'hook.txt', 'a') as f:
print >> f, epoch, this_mse
image = paramgraphics.mat_to_img(tmp[:n_visualization, :].T,
dim_input,
colorImg=colorImg)
image.save(logdir + 'procedure-' + str(epoch) + '.png', 'PNG')
np.savez(logdir + 'procedure-' + str(epoch), tmp=tmp)
image = paramgraphics.mat_to_img((output.reshape(-1, 784)).T,
dim_input,
colorImg=colorImg,
tile_shape=(n_visualization, 22))
image.save(logdir + 'output.png', 'PNG')
np.savez(logdir + 'output', output=output)
# save original train features and denoise test features
for i in xrange(n_train_batches):
if i == 0:
train_features = np.asarray(train_activations(i))
else:
train_features = np.vstack(
(train_features, np.asarray(train_activations(i))))
for i in xrange(n_valid_batches):
if i == 0:
valid_features = np.asarray(valid_activations(i))
else:
valid_features = np.vstack(
(valid_features, np.asarray(valid_activations(i))))
for i in xrange(n_test_batches):
if i == 0:
test_features = np.asarray(
test_activations(tmp[i * batch_size:(i + 1) * batch_size]))
else:
test_features = np.vstack(
(test_features,
np.asarray(
test_activations(tmp[i * batch_size:(i + 1) *
batch_size]))))
np.save(logdir + 'train_features', train_features)
np.save(logdir + 'valid_features', valid_features)
np.save(logdir + 'test_features', test_features)
def cmmd(dataset='mnist.pkl.gz',
batch_size=100,
layer_num=3,
hidden_dim=5,
seed=0,
layer_size=[64, 256, 256, 512]):
validation_frequency = 1
test_frequency = 1
pre_train = 1
dim_input = (28, 28)
colorImg = False
print "Loading data ......."
#datasets = datapy.load_data_gpu_60000_with_noise(dataset, have_matrix = True)
datasets = datapy.load_data_gpu_60000(dataset, have_matrix=True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
test_set_x, test_set_y, test_y_matrix = datasets[2]
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
n_train_batches = train_set_x.get_value().shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
aImage = paramgraphics.mat_to_img(train_set_x.get_value()[0:169].T,
dim_input,
colorImg=colorImg)
aImage.save('mnist_sample', 'PNG')
################################
## build model ##
################################
print "Building model ......."
index = T.lscalar()
x = T.matrix('x') ##### batch_size * 28^2
y = T.vector('y')
y_matrix = T.matrix('y_matrix')
random_z = T.matrix('random_z') ### batch_size * hidden_dim
Inv_K_d = T.matrix('Inv_K_d')
layers = []
layer_output = []
activation = nonlinearity.relu
#activation = Tnn.sigmoid
#### first layer
layers.append(
FullyConnected.FullyConnected(
rng=rng,
n_in=10 + hidden_dim,
#n_in = 10,
n_out=layer_size[0],
activation=activation))
layer_output.append(layers[-1].output_mix(input=[y_matrix, random_z]))
#layer_output.append(layers[-1].output_mix2(input=[y_matrix,random_z]))
#layer_output.append(layers[-1].output(input=x))
#layer_output.append(layers[-1].output(input=random_z))
#### middle layer
for i in range(layer_num):
layers.append(
FullyConnected.FullyConnected(rng=rng,
n_in=layer_size[i],
n_out=layer_size[i + 1],
activation=activation))
layer_output.append(layers[-1].output(input=layer_output[-1]))
#### last layer
activation = Tnn.sigmoid
#activation = nonlinearity.relu
layers.append(
FullyConnected.FullyConnected(rng=rng,
n_in=layer_size[-1],
n_out=28 * 28,
activation=activation))
x_gen = layers[-1].output(input=layer_output[-1])
lambda1_ = 100
lambda_ = theano.shared(np.asarray(lambda1_, dtype=np.float32))
K_d = kernel_gram_for_y(y_matrix, y_matrix, batch_size, 10)
K_s = K_d
K_sd = K_d
Invv_1 = T.sum(y_matrix, axis=0) / batch_size
Invv = NL.alloc_diag(1 / Invv_1)
Inv_K_d = Invv
#Inv_K_d = NL.matrix_inverse(K_d +lambda_ * T.identity_like(K_d))
Inv_K_s = Inv_K_d
L_d = kernel_gram_for_x(x, x, batch_size, 28 * 28)
L_s = kernel_gram_for_x(x_gen, x_gen, batch_size, 28 * 28)
L_ds = kernel_gram_for_x(x, x_gen, batch_size, 28 * 28)
'''
cost = -(NL.trace(T.dot(T.dot(T.dot(K_d, Inv_K_d), L_d), Inv_K_d)) +\
NL.trace(T.dot(T.dot(T.dot(K_s, Inv_K_s), L_s),Inv_K_s))- \
2 * NL.trace(T.dot(T.dot(T.dot(K_sd, Inv_K_d) ,L_ds ), Inv_K_s)))
'''
'''
cost = -(NL.trace(T.dot(L_d, T.ones_like(L_d) )) +\
NL.trace(T.dot(L_s,T.ones_like(L_s)))- \
2 * NL.trace(T.dot(L_ds,T.ones_like(L_ds) )))
cost2 = 2 * T.sum(L_ds) - T.sum(L_s) + NL.trace(T.dot(L_s, T.ones_like(L_s)))\
- 2 * NL.trace( T.dot(L_ds , T.ones_like(L_ds)))
cost2 = T.dot(T.dot(Inv_K_d, K_d),Inv_K_d)
'''
cost2 = K_d
#cost2 = T.dot(T.dot(Inv_K_d,K_d),Inv_K_d)
#cost = - T.sum(L_d) +2 * T.sum(L_ds) - T.sum(L_s)
cost2 = K_d
cost2 = T.dot(T.dot(T.dot(y_matrix, Inv_K_d), Inv_K_d), y_matrix.T)
cost = -(NL.trace(T.dot(T.dot(T.dot(T.dot(L_d, y_matrix),Inv_K_d), Inv_K_d),y_matrix.T)) +\
NL.trace(T.dot(T.dot(T.dot(T.dot(L_s, y_matrix),Inv_K_s), Inv_K_s),y_matrix.T))- \
2 * NL.trace(T.dot(T.dot(T.dot(T.dot(L_ds, y_matrix),Inv_K_d), Inv_K_s),y_matrix.T)))
'''
cost = - T.sum(L_d) +2 * T.sum(L_ds) - T.sum(L_s)
cost = - NL.trace(K_s * Inv_K_s * L_s * Inv_K_s)+ \
2 * NL.trace(K_sd * Inv_K_d * L_ds * Inv_K_s)
'''
################################
## updates ##
################################
params = []
for aLayer in layers:
params += aLayer.params
gparams = [T.grad(cost, param) for param in params]
learning_rate = 3e-4
weight_decay = 1.0 / n_train_batches
epsilon = 1e-8
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
get_optimizer = optimizer.get_adam_optimizer_max(learning_rate=l_r,
decay1=0.1,
decay2=0.001,
weight_decay=weight_decay,
epsilon=epsilon)
updates = get_optimizer(params, gparams)
################################
## pretrain model ##
################################
parameters = theano.function(
inputs=[],
outputs=params,
)
gen_fig = theano.function(
inputs=[y_matrix, random_z],
outputs=x_gen,
on_unused_input='warn',
)
if pre_train == 1:
print "pre-training model....."
pre_train = np.load('./result/MMD-100-5-64-256-256-512.npz')['model']
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
s = 8
for jj in range(10):
a = np.zeros((s, 10), dtype=np.float32)
for ii in range(s):
kk = random.randint(0, 9)
a[ii, kk] = 1
x_gen = gen_fig(a, gen_random_z(s, hidden_dim))
ttt = train_set_x.get_value()
for ll in range(s):
minn = 1000000
ss = 0
for kk in range(ttt.shape[0]):
tt = np.linalg.norm(x_gen[ll] - ttt[kk])
if tt < minn:
minn = tt
ss = kk
#np.concatenate(x_gen,ttt[ss])
x_gen = np.vstack((x_gen, ttt[ss]))
aImage = paramgraphics.mat_to_img(x_gen.T,
dim_input,
colorImg=colorImg)
aImage.save('samples_' + str(jj) + '_similar', 'PNG')
################################
## prepare data ##
################################
#### compute matrix inverse
#print "Preparing data ...."
#Invv = NL.matrix_inverse(K_d +lambda_ * T.identity_like(K_d))
'''
Invv_1 = T.sum(y_matrix,axis=0)/batch_size
Invv = NL.alloc_diag(1/Invv_1)
Inv_K_d = Invv
prepare_data = theano.function(
inputs = [index],
outputs = [Invv,K_d],
givens = {
#x:train_set_x[index * batch_size:(index + 1) * batch_size],
y_matrix:train_y_matrix[index * batch_size:(index + 1) * batch_size],
}
)
Inv_K_d_l, K_d_l = prepare_data(0)
print Inv_K_d_l
for minibatch_index in range(1, n_train_batches):
if minibatch_index % 10 == 0:
print 'minibatch_index:', minibatch_index
Inv_pre_mini, K_d_pre_mini = prepare_data(minibatch_index)
Inv_K_d_l = np.vstack((Inv_K_d_l,Inv_pre_mini))
K_d_l = np.vstack((K_d_l,K_d_pre_mini))
Inv_K_d_g = theano.shared(Inv_K_d_l,borrow=True)
K_d_g = theano.shared(K_d_l, borrow=True)
'''
################################
## train model ##
################################
train_model = theano.function(
inputs=[index, random_z],
outputs=[cost, x_gen, cost2],
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:
train_y_matrix[index * batch_size:(index + 1) * batch_size],
#K_d:K_d_g[index * batch_size:(index + 1) * batch_size],
#Inv_K_d:Inv_K_d_g[index * batch_size:(index + 1) * batch_size],
},
on_unused_input='warn')
n_epochs = 500
cur_epoch = 0
print "Training model ......"
while (cur_epoch < n_epochs):
cur_epoch = cur_epoch + 1
cor = 0
for minibatch_index in xrange(n_train_batches):
print minibatch_index,
print " : ",
cost, x_gen, cost2 = train_model(
minibatch_index, gen_random_z(batch_size, hidden_dim))
print 'cost: ', cost
print 'cost2: ', cost2
if minibatch_index % 30 == 0:
aImage = paramgraphics.mat_to_img(x_gen[0:1].T,
dim_input,
colorImg=colorImg)
aImage.save(
'samples_epoch_' + str(cur_epoch) + '_mini_' +
str(minibatch_index), 'PNG')
if cur_epoch % 1 == 0:
model = parameters()
for i in range(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
np.savez('model-' + str(cur_epoch), model=model)
def cmmd(dataset='mnist.pkl.gz',batch_size=500, layer_num = 2, hidden_dim = 20,seed = 0,layer_size=[500,200,100]):
validation_frequency = 1
test_frequency = 1
pre_train = 0
pre_train_epoch = 30
print "Loading data ......."
datasets = datapy.load_data_gpu_60000(dataset, have_matrix = True)
train_set_x, train_set_y, train_y_matrix = datasets[0]
valid_set_x, valid_set_y, valid_y_matrix = datasets[1]
test_set_x, test_set_y, test_y_matrix = datasets[2]
n_train_batches = train_set_x.get_value().shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
rng = np.random.RandomState(seed)
rng_share = theano.tensor.shared_randomstreams.RandomStreams(0)
################################
## build model ##
################################
print "Building model ......."
index = T.lscalar()
x = T.matrix('x') ##### batch_size * 28^2
y = T.vector('y')
y_matrix = T.matrix('y_matrix')
random_z = T.matrix('random_z') ### batch_size * hidden_dim
Inv_K_d = T.matrix('Inv_K_d')
layers = []
layer_output= []
activation = nonlinearity.relu
#activation = Tnn.sigmoid
#### first layer
layers.append(FullyConnected.FullyConnected(
rng = rng,
n_in = 28*28 + hidden_dim,
#n_in = 28*28,
n_out = layer_size[0],
activation = activation
))
layer_output.append(layers[-1].output_mix(input=[x,random_z]))
#layer_output.append(layers[-1].output(input=x))
#### middle layer
for i in range(layer_num):
layers.append(FullyConnected.FullyConnected(
rng = rng,
n_in = layer_size[i],
n_out = layer_size[i+1],
activation = activation
))
layer_output.append(layers[-1].output(input= layer_output[-1]))
#### last layer
activation = Tnn.sigmoid
layers.append(FullyConnected.FullyConnected(
rng = rng,
n_in = layer_size[-1],
n_out = 10,
activation = activation
))
y_gen = layers[-1].output(input = layer_output[-1])
lambda1_ = 1e-3
lambda_= theano.shared(np.asarray(lambda1_, dtype=np.float32))
K_d = kernel_gram_for_x(x,x,batch_size,28*28)
K_s = K_d
K_sd = K_d
#Inv_K_d = NL.matrix_inverse(K_d +lambda_ * T.identity_like(K_d))
Inv_K_s = Inv_K_d
L_d = kernel_gram(y_matrix,y_matrix,batch_size,10)
L_s = kernel_gram(y_gen,y_gen,batch_size,10)
L_ds = kernel_gram(y_matrix,y_gen,batch_size,10)
cost = -(NL.trace(K_d * Inv_K_d * L_d * Inv_K_d) +\
NL.trace(K_s * Inv_K_s * L_s * Inv_K_s)- \
NL.trace(K_sd * Inv_K_d * L_ds * Inv_K_s))
cost_pre = -T.sum(T.sqr(y_matrix - y_gen))
cc = T.argmax(y_gen,axis=1)
correct = T.sum(T.eq(T.cast(T.argmax(y_gen,axis=1),'int32'),T.cast(y,'int32')))
################################
## updates ##
################################
params = []
for aLayer in layers:
params += aLayer.params
gparams = [T.grad(cost,param) for param in params]
gparams_pre = [T.grad(cost_pre,param) for param in params]
learning_rate = 3e-4
weight_decay=1.0/n_train_batches
epsilon=1e-8
l_r = theano.shared(np.asarray(learning_rate, dtype=np.float32))
get_optimizer = optimizer.get_adam_optimizer_max(learning_rate=l_r,
decay1=0.1, decay2=0.001, weight_decay=weight_decay, epsilon=epsilon)
updates = get_optimizer(params,gparams)
updates_pre = get_optimizer(params,gparams_pre)
################################
## pretrain model ##
################################
parameters = theano.function(
inputs = [],
outputs = params,
)
'''
pre_train_model = theano.function(
inputs = [index,random_z],
outputs = [cost_pre, correct],
updates=updates_pre,
givens={
x:train_set_x[index * batch_size:(index + 1) * batch_size],
y:train_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:train_y_matrix[index * batch_size:(index + 1) * batch_size],
},
on_unused_input='warn'
)
cur_epoch = 0
if pre_train == 1:
for cur_epoch in range(pre_train_epoch):
print 'cur_epoch: ', cur_epoch,
cor = 0
for minibatch_index in range(n_train_batches):
cost_pre_mini,correct_pre_mini = pre_train_model(minibatch_index,gen_random_z(batch_size,hidden_dim))
cor = cor + correct_pre_mini
print 'correct number: ' , cor
#np.savez(,model = model)
'''
if pre_train == 1:
print "pre-training model....."
pre_train = np.load('model.npz')['model']
for (para, pre) in zip(params, pre_train):
para.set_value(pre)
################################
## prepare data ##
################################
#### compute matrix inverse
print "Preparing data ...."
Invv = NL.matrix_inverse(K_d +lambda_ * T.identity_like(K_d))
prepare_data = theano.function(
inputs = [index],
outputs = [Invv,K_d],
givens = {
x:train_set_x[index * batch_size:(index + 1) * batch_size],
}
)
Inv_K_d_l, K_d_l = prepare_data(0)
for minibatch_index in range(1, n_train_batches):
if minibatch_index % 10 == 0:
print 'minibatch_index:', minibatch_index
Inv_pre_mini, K_d_pre_mini = prepare_data(minibatch_index)
Inv_K_d_l = np.vstack((Inv_K_d_l,Inv_pre_mini))
K_d_l = np.vstack((K_d_l,K_d_pre_mini))
Inv_K_d_g = theano.shared(Inv_K_d_l,borrow=True)
K_d_g = theano.shared(K_d_l, borrow=True)
################################
## train model ##
################################
train_model = theano.function(
inputs = [index,random_z],
outputs = [correct,cost,y,cc,y_gen],
updates=updates,
givens={
x:train_set_x[index * batch_size:(index + 1) * batch_size],
y:train_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:train_y_matrix[index * batch_size:(index + 1) * batch_size],
#K_d:K_d_g[index * batch_size:(index + 1) * batch_size],
Inv_K_d:Inv_K_d_g[index * batch_size:(index + 1) * batch_size],
},
on_unused_input='warn'
)
valid_model = theano.function(
inputs = [index,random_z],
outputs = correct,
#updates=updates,
givens={
x:valid_set_x[index * batch_size:(index + 1) * batch_size],
y:valid_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:valid_y_matrix[index * batch_size:(index + 1) * batch_size],
},
on_unused_input='warn'
)
test_model = theano.function(
inputs = [index,random_z],
outputs = [correct,y_gen],
#updates=updates,
givens={
x:test_set_x[index * batch_size:(index + 1) * batch_size],
y:test_set_y[index * batch_size:(index + 1) * batch_size],
y_matrix:test_y_matrix[index * batch_size:(index + 1) * batch_size],
},
on_unused_input='warn'
)
n_epochs = 500
cur_epoch = 0
print "Training model ......"
while (cur_epoch < n_epochs) :
cur_epoch = cur_epoch + 1
cor = 0
for minibatch_index in xrange(n_train_batches):
print minibatch_index,
print " : ",
correct,cost,a,b,y_gen = train_model(minibatch_index,gen_random_z(batch_size,hidden_dim))
cor = cor + correct
print correct
print b
print y_gen
with open('log.txt','a') as f:
print >>f , "epoch: " , cur_epoch, "training_correct: " , cor
if cur_epoch % validation_frequency == 0:
cor2 = 0
for minibatch_index in xrange(n_valid_batches):
correct = valid_model(minibatch_index,gen_random_z(batch_size,hidden_dim))
cor2 = cor2 + correct
with open('log.txt','a') as f:
print >>f , " validation_correct: " , cor2
if cur_epoch % test_frequency == 0:
cor2 = 0
for minibatch_index in xrange(n_test_batches):
correct,y_gen = test_model(minibatch_index,gen_random_z(batch_size,hidden_dim))
with open('log.txt','a') as f:
for index in range(batch_size):
if not np.argmax(y_gen[index]) == test_set_y[minibatch_index * batch_size + index]:
print >>f , "index: " , minibatch_index * batch_size + index, 'true Y: ', test_set_y[minibatch_index * batch_size + index]
print >>f , 'gen_y: ' , y_gen[index]
cor2 = cor2 + correct
with open('log.txt','a') as f:
print >>f , " test_correct: " , cor2
if epoch %1 == 0:
model = parameters()
for i in range(len(model)):
model[i] = np.asarray(model[i]).astype(np.float32)
np.savez('model-'+str(epoch),model=model)
def mnist_model():
initializer = parameter.GaussianInitializer(std=0.1)
bias_initializer = parameter.ConstantInitializer(0.1)
model = network.Network()
model.add(
FullyConnected(
name='fc1',
in_feature=784,
out_feature=512,
weight_initializer=initializer,
bias_initializer=bias_initializer))
model.add(
BatchNorm(name='bn1',
num_features=512)
)
model.add(ReLU(name='relu1'))
model.add(
FullyConnected(
name='fc2',
in_feature=512,
out_feature=256,
weight_initializer=initializer,
bias_initializer=bias_initializer))
model.add(ReLU(name='relu2'))
model.add(
FullyConnected(
name='fc3',
in_feature=256,
out_feature=10,
weight_initializer=initializer,
bias_initializer=bias_initializer))
model.add(Softmax())
model.add_loss(CrossEntropyLoss())
lr = 0.1
optimizer = Momentum(lr=lr)
print(model)
for k, v in model.parameters().items():
print(k, v)
traingset = fetch_traingset()
testset = fetch_testingset()
test_images, test_labels = testset['images'], testset['labels']
test_labels = np.array(test_labels)
images, labels = traingset['images'], traingset['labels']
batch_size = 512
training_size = len(images)
pbar = tqdm.tqdm(range(100))
for epoch in pbar:
losses = []
model.train_mode()
for i in range(int(training_size/batch_size)):
batch_images = np.array(images[i*batch_size:(i+1)*batch_size])
batch_labels = np.array(labels[i*batch_size:(i+1)*batch_size])
batch_labels = one_hot(batch_labels, 10)
_, loss = model.forward(batch_images, batch_labels)
losses.append(loss)
model.optimize(optimizer)
model.eval_mode()
predicts = np.zeros((len(test_labels)))
for i in range(int(len(test_labels)/1000)):
batch_images = np.array(test_images[i*1000:(i+1)*1000])
pred = model.forward(batch_images)
pred = np.argmax(pred, 1)
predicts[i*1000:(i+1)*1000] = pred
acc = np.sum(test_labels == predicts) * 100 / len(test_labels)
pbar.set_description('e:{} loss:{} acc:{}%'.format(epoch, float(np.mean(losses)), acc))