def train(X_train, X_test, y_train, y_test):
xp = self.xp
model = self._create_model()
optimizer = SGD(lr=0.001)
optimizer.setup(model)
if self.cv:
y_test = xp.array(y_test).reshape((-1, 1))
for epoch in range(n_epoch):
print('epoch:', epoch)
chainer.using_config('train', True)
self._train_once(model, optimizer, X_train, y_train)
with chainer.no_backprop_mode():
chainer.using_config('train', False)
pred = self._predict(model, X_test)
if self.cv:
loss = F.sigmoid_cross_entropy(pred, y_test)
print("test loss:", loss.data)
if self.gpu is not None:
pred = chainer.cuda.to_cpu(pred)
return pred
def setup_optimizer(cfg):
if cfg.solver.optimizer == 'SGD':
optimizer = SGD(cfg.optimizer.base_lr)
elif cfg.solver.optimizer == 'MomentumSGD':
optimizer = MomentumSGD(cfg.solver.base_lr, cfg.solver.momentum)
else:
raise ValueError('Not support `optimizer`: {}.'.format(
cfg.solver.optimizer))
return optimizer
classes = np.unique(t_train) # 定義されたクラスラベル
num_classes = len(classes) # クラス数
dim_features = x_train.shape[-1] # xの次元
# 超パラメータの定義
learning_rate = 0.5 # learning_rate(学習率)を定義する
max_iteration = 100 # 学習させる回数
batch_size = 200 # ミニバッチ1つあたりのサンプル数
dim_hidden = 200 # 隠れ層の次元数を定義する
linear_1 = F.Linear(dim_features, dim_hidden)
linear_2 = F.Linear(dim_hidden, num_classes)
model = FunctionSet(linear_1=linear_1,
linear_2=linear_2)
optimizer = SGD(learning_rate)
optimizer.setup(model)
loss_history = []
train_accuracy_history = []
loss_valid_history = []
valid_accuracy_history = []
valid_accuracy_best = 0
valid_loss_best = 10
num_batches = num_train / batch_size # ミニバッチの個数
num_valid_batches = num_valid / batch_size
# 学習させるループ
for epoch in range(max_iteration):
print "epoch:", epoch
y = self.h2y(h)
return y
max_epoch = 100
hidden_size = 512 #100
bptt_length = 30
batch_size = 100
lr = 0.05 #1e-4
indices, char_to_id, id_to_char = load_shakespear()
iterator = RnnIterator(indices, batch_size)
vocab_size = len(char_to_id)
rnn = SimpleRNN(vocab_size, hidden_size, vocab_size)
model = L.Classifier(rnn)
optimizer = SGD(lr=lr)
optimizer.setup(model)
def generate_sample(n=30, init_char=' '):
rnn.reset_state()
s = ''
x = np.array([char_to_id[init_char]])
for i in range(n):
y = rnn(x)
m = y.data.argmax()
c = id_to_char[m]
s += c
x = np.array([m])
def setUp(self):
self.optimizer = SGD(0.1)
self.model = LinearModel(self.optimizer)
