def train():
"""
This examples implements the variational autoencoder from the paper
Auto-Encoding Variational Bayes by Diederik P Kingma, Max Welling, arXiv:1312.6114
"""
# build dataset
data = Mnist(batch_size=100, binary=False, train_valid_test_ratio=[5,1,1])
# for autoencoder, the output will be equal to input
data.set_train(X=data.get_train().X, y=data.get_train().X)
data.set_valid(X=data.get_valid().X, y=data.get_valid().X)
# build model
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(VariationalAutoencoder(input_dim=28*28, bottlenet_dim=200, z_dim=20))
# build learning method
learning_method = SGD(learning_rate=0.0001, momentum=0.9,
lr_decay_factor=0.9, decay_batch=10000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = SGVB_bin,
valid_cost = SGVB_bin,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
def train():
data = Cifar10(batch_size=32, train_valid_test_ratio=[4,1,1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Convolution2D(input_channels=3, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='full'))
model.add(RELU())
model.add(Convolution2D(input_channels=8, filters=16, kernel_size=(3,3), stride=(1,1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(4, 4), stride=(4,4), mode='max'))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Linear(16*8*8, 512))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(512, 10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# Build Logger
log = Log(experiment_name = 'cifar10_cnn',
description = 'This is a tutorial',
save_outputs = True, # log all the outputs from the screen
save_model = True, # save the best model
save_epoch_error = True, # log error at every epoch
save_to_database = {'name': 'hyperparam.sqlite3',
'records': {'Batch_Size': data.batch_size,
'Learning_Rate': learning_method.learning_rate,
'Momentum': learning_method.momentum}}
) # end log
# put everything into the train object
train_object = TrainObject(model = model,
log = log,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 30,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
# test the model on test set
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
batch_size = 128
data = Cifar10(batch_size=batch_size, train_valid_test_ratio=[4,1,1])
_, c, h, w = data.train.X.shape
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Convolution2D(input_channels=c, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='full'))
h, w = full(h, w, kernel=3, stride=1)
model.add(RELU())
model.add(Convolution2D(input_channels=8, filters=16, kernel_size=(3,3), stride=(1,1), border_mode='valid'))
h, w = valid(h, w, kernel=3, stride=1)
model.add(RELU())
model.add(Pooling2D(poolsize=(4, 4), stride=(4,4), mode='max'))
h, w = valid(h, w, kernel=4, stride=4)
model.add(Flatten())
model.add(Linear(16*h*w, 512))
model.add(BatchNormalization((512,), short_memory=0.9))
model.add(RELU())
model.add(Linear(512, 10))
model.add(Softmax())
learning_method = RMSprop(learning_rate=0.01)
# Build Logger
log = Log(experiment_name = 'cifar10_cnn_example',
description = 'This is a tutorial',
save_outputs = True, # log all the outputs from the screen
save_model = True, # save the best model
save_epoch_error = True, # log error at every epoch
save_to_database = {'name': 'hyperparam.sqlite3',
'records': {'Batch_Size': batch_size,
'Learning_Rate': learning_method.learning_rate}}
) # end log
# put everything into the train object
train_object = TrainObject(model = model,
log = log,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 30,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
# test the model on test set
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
# build dataset
batch_size = 64
data = Mnist(batch_size=batch_size, train_valid_test_ratio=[5,1,1])
# build model
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Linear(prev_dim=28*28, this_dim=200))
model.add(RELU())
model.add(Linear(prev_dim=200, this_dim=100))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(prev_dim=100, this_dim=10))
model.add(Softmax())
# build learning method
decay_batch = int(data.train.X.shape[0] * 2 / batch_size)
learning_method = SGD(learning_rate=0.1, momentum=0.9,
lr_decay_factor=0.9, decay_batch=decay_batch)
# Build Logger
log = Log(experiment_name = 'MLP',
description = 'This is a tutorial',
save_outputs = True, # log all the outputs from the screen
save_model = True, # save the best model
save_epoch_error = True, # log error at every epoch
save_to_database = {'name': 'Example.sqlite3',
'records': {'Batch_Size': batch_size,
'Learning_Rate': learning_method.learning_rate,
'Momentum': learning_method.momentum}}
) # end log
# put everything into the train object
train_object = TrainObject(model = model,
log = log,
dataset = data,
train_cost = mse,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 100,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
# build dataset
data = Mnist(batch_size=64, train_valid_test_ratio=[5,1,1])
# for autoencoder, the output will be equal to input
data.set_train(X=data.get_train().X, y=data.get_train().X)
data.set_valid(X=data.get_valid().X, y=data.get_valid().X)
# build model
model = Sequential()
# build encoder
model.add(Gaussian(input_var=T.matrix()))
encode_layer1 = Linear(prev_dim=28*28, this_dim=200)
model.add(encode_layer1)
model.add(RELU())
encode_layer2 = Linear(prev_dim=200, this_dim=50)
model.add(encode_layer2)
model.add(Tanh())
# build decoder
decode_layer1 = Linear(prev_dim=50, this_dim=200, W=encode_layer2.W.T)
model.add(decode_layer1)
model.add(RELU())
decode_layer2 = Linear(prev_dim=200, this_dim=28*28, W=encode_layer1.W.T)
model.add(decode_layer2)
model.add(Sigmoid())
# build learning method
learning_method = AdaGrad(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=10000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = entropy,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
def train():
_TEXT_INPUT_DIM_ = 10
_NUM_EXP_ = 1000
_IMG_INPUT_DIM_ = (3, 32, 32)
_OUTPUT_DIM_ = 100
_TEXT_OUTPUT_DIM_ = 100
_IMG_OUTPUT_DIM_ = 80
# build dataset
txt = np.random.rand(_NUM_EXP_, _TEXT_INPUT_DIM_)
img = np.random.rand(_NUM_EXP_, *_IMG_INPUT_DIM_)
y = np.random.rand(_NUM_EXP_, _OUTPUT_DIM_)
data = MultiInputsData(X=(txt, img), y=y)
# build left and right model
left_model = _left_model(_TEXT_INPUT_DIM_, _TEXT_OUTPUT_DIM_)
right_model = _right_model(_IMG_INPUT_DIM_, _IMG_OUTPUT_DIM_)
# build the master model
model = Sequential(input_var=(T.matrix(), T.tensor4()), output_var=T.matrix())
model.add(Parallel(left_model, right_model))
model.add(FlattenAll())
model.add(Concate(_TEXT_OUTPUT_DIM_ + _IMG_OUTPUT_DIM_, _OUTPUT_DIM_))
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
def train():
data = VOC(batch_size=32, train_valid_test_ratio=[5, 1, 1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Alexnet(input_shape=(3, 222, 222), output_dim=11))
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9, lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(
model=model,
log=None,
dataset=data,
train_cost=error,
valid_cost=error,
learning_method=learning_method,
stop_criteria={"max_epoch": 10, "epoch_look_back": 5, "percent_decrease": 0.01},
)
# finally run the code
train_object.setup()
train_object.run()
def train():
"""
This examples implements the variational autoencoder from the paper
Auto-Encoding Variational Bayes by Diederik P Kingma, Max Welling, arXiv:1312.6114
"""
# build dataset
data = Mnist(batch_size=100, binary=False, train_valid_test_ratio=[5,1,1])
# for autoencoder, the output will be equal to input
data.set_train(X=data.get_train().X, y=data.get_train().X)
data.set_valid(X=data.get_valid().X, y=data.get_valid().X)
# build model
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(VariationalAutoencoder(input_dim=28*28, bottlenet_dim=200, z_dim=20))
# build learning method
learning_method = SGD(learning_rate=0.0001, momentum=0.9,
lr_decay_factor=0.9, decay_batch=10000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = SGVB_bin,
valid_cost = SGVB_bin,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
def train():
# build dataset
batch_size = 64
data = Mnist(batch_size=batch_size, train_valid_test_ratio=[5, 1, 1])
# build model
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Linear(prev_dim=28 * 28, this_dim=200))
model.add(RELU())
model.add(Linear(prev_dim=200, this_dim=100))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(prev_dim=100, this_dim=10))
model.add(Softmax())
# build learning method
decay_batch = int(data.train.X.shape[0] * 2 / batch_size)
learning_method = SGD(learning_rate=0.1,
momentum=0.9,
lr_decay_factor=0.9,
decay_batch=decay_batch)
# Build Logger
log = Log(
experiment_name='MLP',
description='This is a tutorial',
save_outputs=True, # log all the outputs from the screen
save_model=True, # save the best model
save_epoch_error=True, # log error at every epoch
save_to_database={
'name': 'Example.sqlite3',
'records': {
'Batch_Size': batch_size,
'Learning_Rate': learning_method.learning_rate,
'Momentum': learning_method.momentum
}
}) # end log
# put everything into the train object
train_object = TrainObject(model=model,
log=log,
dataset=data,
train_cost=mse,
valid_cost=error,
learning_method=learning_method,
stop_criteria={
'max_epoch': 100,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print('test accuracy:', accuracy)
def train():
data = VOC(batch_size=32, train_valid_test_ratio=[5, 1, 1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Alexnet(input_shape=(3, 222, 222), output_dim=11))
# build learning method
learning_method = SGD(learning_rate=0.01,
momentum=0.9,
lr_decay_factor=0.9,
decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model=model,
log=None,
dataset=data,
train_cost=error,
valid_cost=error,
learning_method=learning_method,
stop_criteria={
'max_epoch': 10,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
def train():
# build dataset
data = Mnist(batch_size=64, train_valid_test_ratio=[5,1,1])
# build model
model = Sequential()
model.add(Linear(prev_dim=28*28, this_dim=200))
model.add(RELU())
model.add(Linear(prev_dim=200, this_dim=100))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(prev_dim=100, this_dim=10))
model.add(Softmax())
# build learning method
learning_method = AdaGrad(learning_rate=0.1, momentum=0.9,
lr_decay_factor=0.9, decay_batch=10000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = mse,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
print 'test accuracy:', accuracy_score(y, ypred)
def train():
data = Cifar10(batch_size=32, train_valid_test_ratio=[4,1,1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Convolution2D(input_channels=3, filters=32, kernel_size=(3,3), stride=(1,1), border_mode='full'))
model.add(RELU())
model.add(Convolution2D(input_channels=32, filters=32, kernel_size=(3,3), stride=(1,1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(2, 2), mode='max'))
model.add(Dropout(0.25))
model.add(Convolution2D(input_channels=32, filters=64, kernel_size=(3,3), stride=(1,1), border_mode='full'))
model.add(RELU())
model.add(Convolution2D(input_channels=64, filters=64, kernel_size=(3,3), stride=(1,1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(2, 2), mode='max'))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Linear(64*8*8, 512))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(512, 10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
# test the model on test set
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
# build dataset
data = Mnist(batch_size=64, train_valid_test_ratio=[5, 1, 1])
# for autoencoder, the output will be equal to input
data.set_train(X=data.get_train().X, y=data.get_train().X)
data.set_valid(X=data.get_valid().X, y=data.get_valid().X)
# build model
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
# build encoder
model.add(Gaussian())
encode_layer1 = Linear(prev_dim=28 * 28, this_dim=200)
model.add(encode_layer1)
model.add(RELU())
encode_layer2 = Linear(prev_dim=200, this_dim=50)
model.add(encode_layer2)
model.add(Tanh())
# build decoder
decode_layer1 = Linear(prev_dim=50, this_dim=200, W=encode_layer2.W.T)
model.add(decode_layer1)
model.add(RELU())
decode_layer2 = Linear(prev_dim=200, this_dim=28 * 28, W=encode_layer1.W.T)
model.add(decode_layer2)
model.add(Sigmoid())
# build learning method
learning_method = AdaGrad(learning_rate=0.01,
momentum=0.9,
lr_decay_factor=0.9,
decay_batch=10000)
# put everything into the train object
train_object = TrainObject(model=model,
log=None,
dataset=data,
train_cost=entropy,
valid_cost=entropy,
learning_method=learning_method,
stop_criteria={
'max_epoch': 10,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
def train():
_TEXT_INPUT_DIM_ = 10
_NUM_EXP_ = 1000
_IMG_INPUT_DIM_ = (3, 32, 32)
_OUTPUT_DIM_ = 100
_TEXT_OUTPUT_DIM_ = 100
_IMG_OUTPUT_DIM_ = 80
# build dataset
txt = np.random.rand(_NUM_EXP_, _TEXT_INPUT_DIM_)
img = np.random.rand(_NUM_EXP_, *_IMG_INPUT_DIM_)
y = np.random.rand(_NUM_EXP_, _OUTPUT_DIM_)
data = MultiInputsData(X=(txt, img), y=y)
# build left and right model
left_model = _left_model(_TEXT_INPUT_DIM_, _TEXT_OUTPUT_DIM_)
right_model = _right_model(_IMG_INPUT_DIM_, _IMG_OUTPUT_DIM_)
# build the master model
model = Sequential(input_var=(T.matrix(), T.tensor4()),
output_var=T.matrix())
model.add(Parallel(left_model, right_model))
model.add(FlattenAll())
model.add(Concate(_TEXT_OUTPUT_DIM_ + _IMG_OUTPUT_DIM_, _OUTPUT_DIM_))
# build learning method
learning_method = SGD(learning_rate=0.01,
momentum=0.9,
lr_decay_factor=0.9,
decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model=model,
log=None,
dataset=data,
train_cost=entropy,
valid_cost=error,
learning_method=learning_method,
stop_criteria={
'max_epoch': 10,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
def train():
# create a fake dataset
X1 = np.random.rand(100000, 1000)
y1 = np.random.rand(100000, 10)
with open('X1.npy', 'wb') as xin, open('y1.npy', 'wb') as yin:
np.save(xin, X1)
np.save(yin, y1)
X2 = np.random.rand(100000, 1000)
y2 = np.random.rand(100000, 10)
with open('X2.npy', 'wb') as xin, open('y2.npy', 'wb') as yin:
np.save(xin, X2)
np.save(yin, y2)
X3 = np.random.rand(100000, 1000)
y3 = np.random.rand(100000, 10)
with open('X3.npy', 'wb') as xin, open('y3.npy', 'wb') as yin:
np.save(xin, X3)
np.save(yin, y3)
# now we can create the data by putting the paths
# ('X1.npy', 'y1.npy') and ('X2.npy', 'y2.npy') into DataBlocks
data = DataBlocks(data_paths=[('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy'), ('X3.npy', 'y3.npy')],
batch_size=100, train_valid_test_ratio=[3,2,0], allow_preload=False)
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Linear(prev_dim=1000, this_dim=200))
model.add(RELU())
model.add(Linear(prev_dim=200, this_dim=100))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(prev_dim=100, this_dim=10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
for X_path, y_path in [('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy')]:
with open(X_path) as Xin, open(y_path) as yin:
# test the model on test set
ypred = model.fprop(np.load(Xin))
ypred = np.argmax(ypred, axis=1)
y = np.argmax(np.load(yin), axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print('combined accuracy for blk %s:'%X_path, accuracy)
def train():
batch_size = 256
short_memory = 0.9
learning_rate = 0.005
data = Cifar10(batch_size=batch_size, train_valid_test_ratio=[4, 1, 1])
_, c, h, w = data.train.X.shape
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(
Convolution2D(input_channels=c,
filters=8,
kernel_size=(3, 3),
stride=(1, 1),
border_mode='full'))
h, w = full(h, w, kernel=3, stride=1)
model.add(
BatchNormalization(dim=8, layer_type='conv',
short_memory=short_memory))
model.add(RELU())
model.add(
Convolution2D(input_channels=8,
filters=16,
kernel_size=(3, 3),
stride=(1, 1),
border_mode='valid'))
h, w = valid(h, w, kernel=3, stride=1)
model.add(
BatchNormalization(dim=16,
layer_type='conv',
short_memory=short_memory))
model.add(RELU())
model.add(Pooling2D(poolsize=(4, 4), stride=(4, 4), mode='max'))
h, w = valid(h, w, kernel=4, stride=4)
model.add(Flatten())
model.add(Linear(16 * h * w, 512))
model.add(
BatchNormalization(dim=512, layer_type='fc',
short_memory=short_memory))
model.add(RELU())
model.add(Linear(512, 10))
model.add(Softmax())
# learning_method = RMSprop(learning_rate=learning_rate)
learning_method = Adam(learning_rate=learning_rate)
# learning_method = SGD(learning_rate=0.001)
# Build Logger
log = Log(
experiment_name='cifar10_cnn_tutorial',
description='This is a tutorial',
save_outputs=True, # log all the outputs from the screen
save_model=True, # save the best model
save_epoch_error=True, # log error at every epoch
save_to_database={
'name': 'hyperparam.sqlite3',
'records': {
'Batch_Size': batch_size,
'Learning_Rate': learning_method.learning_rate
}
}) # end log
# put everything into the train object
train_object = TrainObject(model=model,
log=log,
dataset=data,
train_cost=entropy,
valid_cost=error,
learning_method=learning_method,
stop_criteria={
'max_epoch': 100,
'epoch_look_back': 10,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
# test the model on test set
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
max_features=20000
maxseqlen = 100 # cut texts after this number of words (among top max_features most common words)
batch_size = 16
word_vec_len = 256
iter_class = 'SequentialRecurrentIterator'
seq_len = 10
data = IMDB(pad_zero=True, maxlen=100, nb_words=max_features, batch_size=batch_size,
train_valid_test_ratio=[8,2,0], iter_class=iter_class, seq_len=seq_len)
print('Build model...')
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Embedding(max_features, word_vec_len))
# MLP layers
model.add(Transform((word_vec_len,))) # transform from 3d dimensional input to 2d input for mlp
model.add(Linear(word_vec_len, 100))
model.add(RELU())
model.add(BatchNormalization(dim=100, layer_type='fc'))
model.add(Linear(100,100))
model.add(RELU())
model.add(BatchNormalization(dim=100, layer_type='fc'))
model.add(Linear(100, word_vec_len))
model.add(RELU())
model.add(Transform((maxseqlen, word_vec_len))) # transform back from 2d to 3d for recurrent input
# Stacked up BiLSTM layers
model.add(BiLSTM(word_vec_len, 50, output_mode='concat', return_sequences=True))
model.add(BiLSTM(100, 24, output_mode='sum', return_sequences=True))
model.add(LSTM(24, 24, return_sequences=True))
# MLP layers
model.add(Reshape((24 * maxseqlen,)))
model.add(BatchNormalization(dim=24 * maxseqlen, layer_type='fc'))
model.add(Linear(24 * maxseqlen, 50))
model.add(RELU())
model.add(Dropout(0.2))
model.add(Linear(50, 1))
model.add(Sigmoid())
# build learning method
decay_batch = int(data.train.X.shape[0] * 5 / batch_size)
learning_method = SGD(learning_rate=0.1, momentum=0.9,
lr_decay_factor=1.0, decay_batch=decay_batch)
# Build Logger
log = Log(experiment_name = 'MLP',
description = 'This is a tutorial',
save_outputs = True, # log all the outputs from the screen
save_model = True, # save the best model
save_epoch_error = True, # log error at every epoch
save_to_database = {'name': 'Example.sqlite3',
'records': {'Batch_Size': batch_size,
'Learning_Rate': learning_method.learning_rate,
'Momentum': learning_method.momentum}}
) # end log
# put everything into the train object
train_object = TrainObject(model = model,
log = log,
dataset = data,
train_cost = mse,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 100,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
def train(args):
# build dataset
xpath = os.environ['MOZI_DATA_PATH'] + '/X_{}_augment_{}.npy'.format(
'_'.join([str(d) for d in _IMG_INPUT_DIM_]), str(img_augment))
ypath = os.environ['MOZI_DATA_PATH'] + '/y_{}_augment_{}.npy'.format(
'_'.join([str(d) for d in _IMG_INPUT_DIM_]), str(img_augment))
if not os.path.exists(xpath) or not os.path.exists(ypath):
X, y = make_Xy(args, img_augment)
with open(xpath, 'wb') as fout:
np.save(fout, X)
print '..saved to', xpath
with open(ypath, 'wb') as fout:
np.save(fout, y)
print '..saved to', ypath
else:
with open(xpath) as xin, open(ypath) as yin:
X = np.load(xin)
y = np.load(yin)
print '..data loaded'
if img_preprocess:
X = img_preprocess.apply(X)
# import pdb; pdb.set_trace()
idxs = np.arange(len(X))
np.random.shuffle(idxs)
data = MultiInputsData(X=X[idxs][:10000],
y=y[idxs][:10000],
train_valid_test_ratio=train_valid_test_ratio,
batch_size=batch_size)
if load_model:
print '..loading model', load_model
model_path = os.environ[
'MOZI_SAVE_PATH'] + '/' + load_model + '/model.pkl'
with open(model_path) as fin:
model = cPickle.load(fin)
else:
# c, h, w = _IMG_INPUT_DIM_
# build the master model
model = Sequential(input_var=T.tensor4(),
output_var=T.tensor4(),
verbose=verbose)
ks = 11
model.add(
Convolution2D(input_channels=3,
filters=16,
kernel_size=(ks, ks),
stride=(1, 1),
border_mode='full'))
model.add(Crop(border=(ks / 2, ks / 2)))
model.add(
BatchNormalization(dim=16,
layer_type='conv',
short_memory=short_memory))
model.add(RELU())
model.add(
Pooling2D(poolsize=(3, 3),
stride=(1, 1),
padding=(1, 1),
mode='max'))
# model.add(RELU())
# h, w = full(h, w, 5, 1)
ks = 9
model.add(
Convolution2D(input_channels=16,
filters=32,
kernel_size=(ks, ks),
stride=(1, 1),
border_mode='full'))
model.add(Crop(border=(ks / 2, ks / 2)))
model.add(
BatchNormalization(dim=32,
layer_type='conv',
short_memory=short_memory))
model.add(RELU())
model.add(
Pooling2D(poolsize=(3, 3),
stride=(1, 1),
padding=(1, 1),
mode='max'))
ks = 5
model.add(
Convolution2D(input_channels=32,
filters=1,
kernel_size=(ks, ks),
stride=(1, 1),
border_mode='full'))
# model.add(BatchNormalization(dim=1, layer_type='conv', short_memory=short_memory))
model.add(Crop(border=(ks / 2, ks / 2)))
model.add(Sigmoid())
# build learning method
# learning_method = SGD(learning_rate=lr, momentum=momentum,
# lr_decay_factor=lr_decay_factor, decay_batch=decay_batch)
learning_method = Adam(learning_rate=lr)
# learning_method = RMSprop(learning_rate=lr)
# Build Logger
log = Log(
experiment_name=experiment_name,
description=desc,
save_outputs=True, # log all the outputs from the screen
save_model=save_model, # save the best model
save_epoch_error=True, # log error at every epoch
save_to_database={
'name': 'skin_segment.sqlite3',
'records': {
'learning_rate': lr,
'valid_cost_func': valid_cost,
'train_cost_func': train_cost
}
}) # end log
os.system('cp {} {}'.format(__file__, log.exp_dir))
dname = os.path.dirname(os.path.realpath(__file__))
# put everything into the train object
train_object = TrainObject(model=model,
log=log,
dataset=data,
train_cost=train_cost,
valid_cost=valid_cost,
learning_method=learning_method,
stop_criteria={
'max_epoch': 100,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
def train():
data = Cifar10(batch_size=32, train_valid_test_ratio=[4, 1, 1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(
Convolution2D(input_channels=3,
filters=8,
kernel_size=(3, 3),
stride=(1, 1),
border_mode='full'))
model.add(RELU())
model.add(
Convolution2D(input_channels=8,
filters=16,
kernel_size=(3, 3),
stride=(1, 1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(4, 4), stride=(4, 4), mode='max'))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Linear(16 * 8 * 8, 512))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(512, 10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01,
momentum=0.9,
lr_decay_factor=0.9,
decay_batch=5000)
# Build Logger
log = Log(
experiment_name='cifar10_cnn',
description='This is a tutorial',
save_outputs=True, # log all the outputs from the screen
save_model=True, # save the best model
save_epoch_error=True, # log error at every epoch
save_to_database={
'name': 'hyperparam.sqlite3',
'records': {
'Batch_Size': data.batch_size,
'Learning_Rate': learning_method.learning_rate,
'Momentum': learning_method.momentum
}
}) # end log
# put everything into the train object
train_object = TrainObject(model=model,
log=log,
dataset=data,
train_cost=entropy,
valid_cost=error,
learning_method=learning_method,
stop_criteria={
'max_epoch': 30,
'epoch_look_back': 5,
'percent_decrease': 0.01
})
# finally run the code
train_object.setup()
train_object.run()
# test the model on test set
ypred = model.fprop(data.get_test().X)
ypred = np.argmax(ypred, axis=1)
y = np.argmax(data.get_test().y, axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'test accuracy:', accuracy
def train():
max_features = 20000
maxseqlen = 100 # cut texts after this number of words (among top max_features most common words)
batch_size = 16
word_vec_len = 256
iter_class = "SequentialRecurrentIterator"
seq_len = 10
data = IMDB(
pad_zero=True,
maxlen=100,
nb_words=max_features,
batch_size=batch_size,
train_valid_test_ratio=[8, 2, 0],
iter_class=iter_class,
seq_len=seq_len,
)
print("Build model...")
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Embedding(max_features, word_vec_len))
# MLP layers
model.add(Transform((word_vec_len,))) # transform from 3d dimensional input to 2d input for mlp
model.add(Linear(word_vec_len, 100))
model.add(RELU())
model.add(BatchNormalization(dim=100, layer_type="fc"))
model.add(Linear(100, 100))
model.add(RELU())
model.add(BatchNormalization(dim=100, layer_type="fc"))
model.add(Linear(100, word_vec_len))
model.add(RELU())
model.add(Transform((maxseqlen, word_vec_len))) # transform back from 2d to 3d for recurrent input
# Stacked up BiLSTM layers
model.add(BiLSTM(word_vec_len, 50, output_mode="concat", return_sequences=True))
model.add(BiLSTM(100, 24, output_mode="sum", return_sequences=True))
model.add(LSTM(24, 24, return_sequences=True))
# MLP layers
model.add(Reshape((24 * maxseqlen,)))
model.add(BatchNormalization(dim=24 * maxseqlen, layer_type="fc"))
model.add(Linear(24 * maxseqlen, 50))
model.add(RELU())
model.add(Dropout(0.2))
model.add(Linear(50, 1))
model.add(Sigmoid())
# build learning method
decay_batch = int(data.train.X.shape[0] * 5 / batch_size)
learning_method = SGD(learning_rate=0.1, momentum=0.9, lr_decay_factor=1.0, decay_batch=decay_batch)
# Build Logger
log = Log(
experiment_name="MLP",
description="This is a tutorial",
save_outputs=True, # log all the outputs from the screen
save_model=True, # save the best model
save_epoch_error=True, # log error at every epoch
save_to_database={
"name": "Example.sqlite3",
"records": {
"Batch_Size": batch_size,
"Learning_Rate": learning_method.learning_rate,
"Momentum": learning_method.momentum,
},
},
) # end log
# put everything into the train object
train_object = TrainObject(
model=model,
log=log,
dataset=data,
train_cost=mse,
valid_cost=error,
learning_method=learning_method,
stop_criteria={"max_epoch": 100, "epoch_look_back": 5, "percent_decrease": 0.01},
)
# finally run the code
train_object.setup()
train_object.run()
def train():
# create a fake dataset
X1 = np.random.rand(100000, 1000)
y1 = np.random.rand(100000, 10)
with open('X1.npy', 'wb') as xin, open('y1.npy', 'wb') as yin:
np.save(xin, X1)
np.save(yin, y1)
X2 = np.random.rand(100000, 1000)
y2 = np.random.rand(100000, 10)
with open('X2.npy', 'wb') as xin, open('y2.npy', 'wb') as yin:
np.save(xin, X2)
np.save(yin, y2)
X3 = np.random.rand(100000, 1000)
y3 = np.random.rand(100000, 10)
with open('X3.npy', 'wb') as xin, open('y3.npy', 'wb') as yin:
np.save(xin, X3)
np.save(yin, y3)
# now we can create the data by putting the paths
# ('X1.npy', 'y1.npy') and ('X2.npy', 'y2.npy') into DataBlocks
data = DataBlocks(data_paths=[('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy'), ('X3.npy', 'y3.npy')],
batch_size=100, train_valid_test_ratio=[3,2,0])
model = Sequential(input_var=T.matrix(), output_var=T.matrix())
model.add(Linear(prev_dim=1000, this_dim=200))
model.add(RELU())
model.add(Linear(prev_dim=200, this_dim=100))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(prev_dim=100, this_dim=10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
for X_path, y_path in [('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy')]:
with open(X_path) as Xin, open(y_path) as yin:
# test the model on test set
ypred = model.fprop(np.load(Xin))
ypred = np.argmax(ypred, axis=1)
y = np.argmax(np.load(yin), axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'combined accuracy for blk %s:'%X_path, accuracy
def train():
X1 = np.random.rand(1000, 3, 32, 32)
y1 = np.random.rand(1000, 10)
with open('X1.npy', 'wb') as xin, open('y1.npy', 'wb') as yin:
np.save(xin, X1)
np.save(yin, y1)
X2 = np.random.rand(1000, 3, 32, 32)
y2 = np.random.rand(1000, 10)
with open('X2.npy', 'wb') as xin, open('y2.npy', 'wb') as yin:
np.save(xin, X1)
np.save(yin, y1)
# now we can create the data by putting the paths
# ('X1.npy', 'y1.npy') and ('X2.npy', 'y2.npy') into DataBlocks
data = DataBlocks(data_paths=[('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy')],
batch_size=100, train_valid_test_ratio=[3,1,1])
model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
model.add(Convolution2D(input_channels=3, filters=32, kernel_size=(3,3), stride=(1,1), border_mode='full'))
model.add(RELU())
model.add(Convolution2D(input_channels=32, filters=32, kernel_size=(3,3), stride=(1,1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(2, 2), mode='max'))
model.add(Dropout(0.25))
model.add(Convolution2D(input_channels=32, filters=64, kernel_size=(3,3), stride=(1,1), border_mode='full'))
model.add(RELU())
model.add(Convolution2D(input_channels=64, filters=64, kernel_size=(3,3), stride=(1,1)))
model.add(RELU())
model.add(Pooling2D(poolsize=(2, 2), mode='max'))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Linear(64*8*8, 512))
model.add(RELU())
model.add(Dropout(0.5))
model.add(Linear(512, 10))
model.add(Softmax())
# build learning method
learning_method = SGD(learning_rate=0.01, momentum=0.9,
lr_decay_factor=0.9, decay_batch=5000)
# put everything into the train object
train_object = TrainObject(model = model,
log = None,
dataset = data,
train_cost = entropy,
valid_cost = error,
learning_method = learning_method,
stop_criteria = {'max_epoch' : 10,
'epoch_look_back' : 5,
'percent_decrease' : 0.01}
)
# finally run the code
train_object.setup()
train_object.run()
for X_path, y_path in [('X1.npy', 'y1.npy'), ('X2.npy', 'y2.npy')]:
with open(X_path) as Xin, open(y_path) as yin:
# test the model on test set
ypred = model.fprop(np.load(Xin))
ypred = np.argmax(ypred, axis=1)
y = np.argmax(np.load(yin), axis=1)
accuracy = np.equal(ypred, y).astype('f4').sum() / len(y)
print 'combined accuracy for blk %s:'%X_path, accuracy