def __init__(self, input_shape, output_dim):
'''
FIELDS:
self.params: any params from the layer that needs to be updated
by backpropagation can be put inside self.params
PARAMS:
input_shape: tuple
shape of the input image with format (channel, height, width)
output_dim: int
the output dimension of the model
'''
assert len(input_shape) == 3, 'input_shape must be a tuple or list of dim (channel, height, width)'
c, h, w = input_shape
valid = lambda x, y, kernel, stride : ((x-kernel)/stride + 1, (y-kernel)/stride + 1)
full = lambda x, y, kernel, stride : ((x+kernel)/stride - 1, (y+kernel)/stride - 1)
self.layers = []
self.layers.append(Convolution2D(input_channels=3, filters=96, kernel_size=(11,11),
stride=(4,4), border_mode='valid'))
nh, nw = valid(h, w, 11, 4)
self.layers.append(RELU())
self.layers.append(LRN())
self.layers.append(Pooling2D(poolsize=(3,3), stride=(2,2), mode='max'))
nh, nw = valid(nh, nw, 3, 2)
self.layers.append(Convolution2D(input_channels=96, filters=256, kernel_size=(5,5),
stride=(1,1), border_mode='full'))
nh, nw = full(nh, nw, 5, 1)
self.layers.append(RELU())
self.layers.append(LRN())
self.layers.append(Pooling2D(poolsize=(3,3), stride=(2,2), mode='max'))
nh, nw = valid(nh, nw, 3, 2)
self.layers.append(Convolution2D(input_channels=256, filters=384, kernel_size=(3,3),
stride=(1,1), border_mode='full'))
nh, nw = full(nh, nw, 3, 1)
self.layers.append(RELU())
self.layers.append(Convolution2D(input_channels=384, filters=384, kernel_size=(3,3),
stride=(1,1), border_mode='full'))
nh, nw = full(nh, nw, 3, 1)
self.layers.append(RELU())
self.layers.append(Convolution2D(input_channels=384, filters=256, kernel_size=(3,3),
stride=(1,1), border_mode='full'))
nh, nw = full(nh, nw, 3, 1)
self.layers.append(RELU())
self.layers.append(Pooling2D(poolsize=(3,3), stride=(2,2), mode='max'))
nh, nw = valid(nh, nw, 3, 2)
self.layers.append(Flatten())
self.layers.append(Linear(256*nh*nw,4096))
self.layers.append(RELU())
self.layers.append(Dropout(0.5))
self.layers.append(Linear(4096,4096))
self.layers.append(RELU())
self.layers.append(Dropout(0.5))
self.layers.append(Linear(4096,output_dim))
self.layers.append(Softmax())
self.params = []
for layer in self.layers:
self.params += layer.params
def _right_model(img_input_dim, merged_dim):
c, h, w = img_input_dim
valid = lambda x, y, kernel, stride : ((x-kernel)/stride + 1, (y-kernel)/stride + 1)
full = lambda x, y, kernel, stride : ((x+kernel)/stride - 1, (y+kernel)/stride - 1)
right_model = Sequential(input_var=T.tensor4(), output_var=T.matrix())
right_model.add(Convolution2D(input_channels=3, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='full'))
h, w = full(h, w, 3, 1)
right_model.add(RELU())
right_model.add(Convolution2D(input_channels=8, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='valid'))
h, w = valid(h, w, 3, 1)
right_model.add(RELU())
right_model.add(Pooling2D(poolsize=(2, 2), stride=(1,1), mode='max'))
h, w = valid(h, w, 2, 1)
right_model.add(Dropout(0.25))
right_model.add(Convolution2D(input_channels=8, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='full'))
h, w = full(h, w, 3, 1)
right_model.add(RELU())
right_model.add(Convolution2D(input_channels=8, filters=8, kernel_size=(3,3), stride=(1,1), border_mode='valid'))
h, w = valid(h, w, 3, 1)
right_model.add(RELU())
right_model.add(Pooling2D(poolsize=(2, 2), stride=(1,1), mode='max'))
h, w = valid(h, w, 2, 1)
right_model.add(Dropout(0.25))
right_model.add(Flatten())
right_model.add(Linear(8*h*w, 512))
right_model.add(Linear(512, 512))
right_model.add(RELU())
right_model.add(Dropout(0.5))
right_model.add(Linear(512, merged_dim))
return right_model
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
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():
# 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():
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():
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