def main(max_iter):
seed = 100
nb_data = 1000
print("loading data ....")
mnist = fetch_mldata('MNIST original',
data_home=os.path.join(os.path.dirname(__file__),
'./data'))
X_train = mnist.data.reshape((-1, 1, 28, 28)) / 255.0
np.random.seed(seed)
X_train = np.random.permutation(X_train)[:nb_data]
y_train = mnist.target
np.random.seed(seed)
y_train = np.random.permutation(y_train)[:nb_data]
n_classes = np.unique(y_train).size
print("building model ...")
net = xeno.Model()
net.add(xeno.layers.Convolution(1, (3, 3), input_shape=(None, 1, 28, 28)))
net.add(xeno.layers.MeanPooling((2, 2)))
net.add(xeno.layers.Convolution(2, (4, 4)))
net.add(xeno.layers.MeanPooling((2, 2)))
net.add(xeno.layers.Flatten())
net.add(xeno.layers.Softmax(n_out=n_classes))
net.compile()
print("train model ... ")
net.fit(X_train,
xeno.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1,
batch_size=100)
def main2(max_iter):
nb_batch = 30
nb_seq = 20
xs, ys, x_size, y_size = prepare_data(nb_seq)
net = xeno.Model()
net.add(xeno.layers.Embedding(nb_batch=nb_batch, nb_seq=nb_seq,
n_out=200, input_size=x_size,
static=False))
net.add(xeno.layers.BatchLSTM(n_out=400, return_sequence=True))
net.add(xeno.layers.BatchLSTM(n_out=200, return_sequence=True))
net.add(xeno.layers.MeanPooling((nb_seq, 1)))
net.add(xeno.layers.Flatten())
net.add(xeno.layers.Softmax(n_out=y_size))
net.compile(loss='scce', optimizer=xeno.optimizers.RMSprop())
net.fit(xs, ys, batch_size=nb_batch, validation_split=0.1, max_iter=max_iter)
def main4(max_iter):
# test Adagrad optimizer
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = xeno.Model()
model.add(xeno.layers.Dense(n_out=100, n_in=784, activation='relu'))
model.add(xeno.layers.Softmax(n_out=n_classes))
model.compile(loss='scce', optimizer='adagrad')
# train
print("train model ... ")
model.fit(X_train,
xeno.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main2(max_iter):
batch_size, vocab_size, time_steps, batch_in, batch_out = get_data()
print("Building model ...")
net = xeno.Model()
net.add(
xeno.layers.BatchLSTM(n_out=300,
n_in=vocab_size,
return_sequence=False,
nb_batch=batch_size,
nb_seq=time_steps))
# net.add(xeno.layers.MeanPooling(pool_size=(time_steps, 1)))
# net.add(xeno.layers.Flatten())
net.add(xeno.layers.Softmax(n_out=vocab_size))
net.compile(loss=xeno.objectives.SCCE(),
optimizer=xeno.optimizers.SGD(lr=0.00001, clip=5))
print("Train model ...")
net.fit(batch_in, batch_out, max_iter=max_iter, batch_size=batch_size)
def main3(max_iter):
# test NesterovMomentum optimizer
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = xeno.Model()
model.add(xeno.layers.Dense(n_out=200, n_in=784, activation='relu'))
model.add(xeno.layers.Softmax(n_out=n_classes))
model.compile(loss=xeno.objectives.SCCE(),
optimizer=xeno.optimizers.NesterovMomentum())
# train
print("train model ... ")
model.fit(X_train,
xeno.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main(max_iter):
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = xeno.Model()
model.add(
xeno.layers.Dense(n_out=200,
n_in=784,
activation=xeno.activations.ReLU()))
model.add(
xeno.layers.Dense(n_out=n_classes,
activation=xeno.activations.Softmax()))
model.compile(loss=xeno.objectives.SCCE(), optimizer=xeno.optimizers.SGD())
# train
print("train model ... ")
model.fit(X_train,
xeno.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
