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 = npdl.Model()
net.add(npdl.layers.Convolution(1, (3, 3), input_shape=(None, 1, 28, 28)))
net.add(npdl.layers.MeanPooling((2, 2)))
net.add(npdl.layers.Flatten())
net.add(npdl.layers.Softmax(n_out=n_classes))
net.compile()
print("train model ... ")
net.fit(X_train,
npdl.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main(max_iter):
# data
print("loading data ...")
mnist = fetch_mldata('MNIST original',
data_home=os.path.join(os.path.dirname(__file__),
'./data'))
X_train = mnist.data / 255.0
y_train = mnist.target
n_classes = np.unique(y_train).size
# model
print("building model ...")
model = npdl.Model()
model.add(
npdl.layers.Dense(n_out=500,
n_in=784,
activation=npdl.activation.ReLU()))
model.add(
npdl.layers.Dense(n_out=n_classes,
activation=npdl.activation.Softmax()))
model.compile(loss=npdl.objectives.SCCE(),
optimizer=npdl.optimizers.SGD(lr=0.001))
# train
print("train model ... ")
model.fit(X_train,
npdl.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main(max_iter, corpus_path=os.path.join(os.path.dirname(__file__), 'data/lm/tiny_shakespeare.txt')):
raw_text = open(corpus_path, 'r').read()
chars = list(set(raw_text))
data_size, vocab_size = len(raw_text), len(chars)
print("data has %s charactres, % unique." % (data_size, vocab_size))
char_to_index = {ch: i for i, ch in enumerate(chars)}
index_to_char = {i: ch for i, ch in enumerate(chars)}
time_steps, batch_size = 30, 40
length = batch_size * 20
text_pointers = np.random.randint(data_size - time_steps - 1, size=length)
batch_in = np.zeros([length, time_steps, vocab_size])
batch_out = np.zeros([length, vocab_size], dtype=np.uint8)
for i in range(length):
b_ = [char_to_index[c] for c in raw_text[text_pointers[i]:text_pointers[i] + time_steps + 1]]
batch_in[i, range(time_steps), b_[:-1]] = 1
batch_out[i, b_[-1]] = 1
print("Building model ...")
net = npdl.Model()
# net.add(model.layers.SimpleRNN(n_out=500, return_sequence=True))
net.add(npdl.layers.SimpleRNN(n_out=500, n_in=vocab_size))
net.add(npdl.layers.Softmax(n_out=vocab_size))
net.compile(loss=npdl.objectives.SCCE(), optimizer=npdl.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 main2(max_iter):
batch_size, vocab_size, time_steps, batch_in, batch_out = get_data()
print("Building model ...")
net = npdl.Model()
net.add(npdl.layers.SimpleRNN(n_out=200, n_in=vocab_size, return_sequence=True,
nb_batch=batch_size, nb_seq=time_steps))
net.add(npdl.layers.SimpleRNN(n_out=200, n_in=200))
net.add(npdl.layers.Softmax(n_out=vocab_size))
net.compile(loss=npdl.objectives.SCCE(), optimizer=npdl.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 main4(max_iter):
# test Adagrad optimizer
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = npdl.Model()
model.add(npdl.layers.Dense(n_out=500, n_in=784, activation='relu'))
model.add(npdl.layers.Softmax(n_out=n_classes))
model.compile(loss='scce', optimizer='adagrad')
# train
print("train model ... ")
model.fit(X_train,
npdl.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main3(max_iter):
# test NesterovMomentum optimizer
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = npdl.Model()
model.add(npdl.layers.Dense(n_out=500, n_in=784, activation='relu'))
model.add(npdl.layers.Softmax(n_out=n_classes))
model.compile(loss=npdl.objectives.SCCE(),
optimizer=npdl.optimizers.NesterovMomentum())
# train
print("train model ... ")
model.fit(X_train,
npdl.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 = npdl.Model()
net.add(
npdl.layers.BatchLSTM(n_out=300,
n_in=vocab_size,
return_sequence=False,
nb_batch=batch_size,
nb_seq=time_steps))
# net.add(npdl.layers.MeanPooling(pool_size=(time_steps, 1)))
# net.add(npdl.layers.Flatten())
net.add(npdl.layers.Softmax(n_out=vocab_size))
net.compile(loss=npdl.objectives.SCCE(),
optimizer=npdl.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 main(max_iter):
n_classes, X_train, y_train = get_data()
# model
print("building model ...")
model = npdl.Model()
model.add(
npdl.layers.Dense(n_out=500,
n_in=784,
activation=npdl.activations.ReLU()))
model.add(
npdl.layers.Dense(n_out=n_classes,
activation=npdl.activations.Softmax()))
model.compile(loss=npdl.objectives.SCCE(), optimizer=npdl.optimizers.SGD())
# train
print("train model ... ")
model.fit(X_train,
npdl.utils.data.one_hot(y_train),
max_iter=max_iter,
validation_split=0.1)
def main(max_iter):
nb_batch = 30
nb_seq = 20
xs, ys, x_size, y_size = prepare_data(nb_seq)
net = npdl.Model()
net.add(
npdl.layers.Embedding(nb_batch=nb_batch,
nb_seq=nb_seq,
n_out=200,
input_size=x_size,
static=True))
net.add(npdl.layers.BatchLSTM(n_out=400, return_sequence=True))
net.add(npdl.layers.BatchLSTM(n_out=200, return_sequence=True))
net.add(npdl.layers.MeanPooling((nb_seq, 1)))
net.add(npdl.layers.Flatten())
net.add(npdl.layers.Softmax(n_out=y_size))
net.compile(loss='scce', optimizer=npdl.optimizers.SGD(lr=0.005))
net.fit(xs,
ys,
batch_size=nb_batch,
validation_split=0.1,
max_iter=max_iter)
