def fit(self,
x,
t,
max_epoch=10,
batch_size=32,
max_grad=None,
eval_interval=20):
data_size = len(x)
max_iters = data_size // batch_size
self.eval_interval = eval_interval
model, optimizer = self.model, self.optimizer
total_loss = 0
loss_count = 0
start_time = time.time()
for epoch in range(max_epoch):
# シャッフル
idx = numpy.random.permutation(numpy.arange(data_size))
x = x[idx]
t = t[idx]
for iters in range(max_iters):
batch_x = x[iters * batch_size:(iters + 1) * batch_size]
batch_t = t[iters * batch_size:(iters + 1) * batch_size]
# 勾配を求め、パラメータを更新
loss = model.forward(batch_x, batch_t)
model.backward()
params, grads = remove_duplicate(model.params,
model.grads) # 共有された重みを1つに集約
if max_grad is not None:
clip_grads(grads, max_grad)
optimizer.update(params, grads)
total_loss += loss
loss_count += 1
# 評価
if (eval_interval
is not None) and (iters % eval_interval) == 0:
avg_loss = total_loss / loss_count
elapsed_time = time.time() - start_time
print(
'| epoch %d | iter %d / %d | time %d[s] | loss %.2f' %
(self.current_epoch + 1, iters + 1, max_iters,
elapsed_time, avg_loss))
self.loss_list.append(float(avg_loss))
total_loss, loss_count = 0, 0
self.current_epoch += 1
def fit(self, x, t, max_epoch=10, batch_size=32, max_grad=None, eval_interval=20):
data_size = len(x)
max_iters = data_size // batch_size
self.eval_interval = eval_interval
total_loss = 0
loss_count = 0
model, optimizer = self.model, self.optimizer
start_time = time.time()
for epoch in range(max_epoch):
# shuffle data
idx = np.random.permutation(data_size)
x = x[idx]
t = t[idx]
for iters in range(max_iters):
batch_x = x[iters*batch_size:(iters+1)*batch_size]
batch_t = t[iters*batch_size:(iters+1)*batch_size]
# Find gradient and update parameters
loss = model.forward(batch_x, batch_t)
model.backward()
# 共有された重みを1つに集約
# これは、word2vecなど、重複の重みを取り除くことです
params, grads = remove_duplicate(model.params, model.grads)
# 勾配爆発を防止
if max_grad is not None:
clip_grads(grads, max_grad)
optimizer.update(params, grads)
total_loss += loss
loss_count += 1
# Learning Output
if (eval_interval is not None) and (iters % eval_interval) == 0:
avg_loss = total_loss / loss_count
elapsed_time = time.time() - start_time
print("| epoch %d | iter %d / %d | time %d[s] | loss %.2f" % (epoch+1, iters+1, max_iters, elapsed_time, avg_loss))
self.loss_list.append(float(avg_loss))
total_loss, loss_count = 0, 0
self.current_epoch += 1
def fit(self,
xs,
ts,
max_epoch=10,
batch_size=20,
time_size=35,
max_grad=None,
eval_interval=20):
data_size = len(xs)
max_iters = data_size // (batch_size * time_size)
self.time_idx = 0
self.ppl_list = []
self.eval_interval = eval_interval
model, optimizer = self.model, self.optimizer
total_loss = 0
loss_count = 0
start_time = time.time()
for epoch in range(max_epoch):
for iters in range(max_iters):
batch_x, batch_t = self.get_batch(xs, ts, batch_size,
time_size)
# 勾配を求め、パラメータを更新
loss = model.forward(batch_x, batch_t)
model.backward()
params, grads = remove_duplicate(model.params,
model.grads) # 共有された重みを1つに集約
if max_grad is not None:
clip_grads(grads, max_grad)
optimizer.update(params, grads)
total_loss += loss
loss_count += 1
# パープレキシティの評価
if (eval_interval
is not None) and (iters % eval_interval) == 0:
ppl = np.exp(total_loss / loss_count)
elapsed_time = time.time() - start_time
print(
'| epoch %d | iter %d / %d | time %d[s] | perplexity %.2f'
% (self.current_epoch + 1, iters + 1, max_iters,
elapsed_time, ppl))
self.ppl_list.append(float(ppl))
total_loss, loss_count = 0, 0
self.current_epoch += 1