def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = F.log1p(x)
testing.assert_allclose(numpy.log1p(self.x),
y.data,
atol=1e-7,
rtol=1e-7)
def lossfun_multi_leam(self, z, t):
with cuda.get_device(z):
z_class = z[-self.n_class:]
z = z[:-self.n_class]
t_class = Variable(cp.array(range(self.n_class)))
loss = - (F.mean(t * F.log(z) + (1-t)*F.log1p(-z))) / z.shape[0] + F.softmax_cross_entropy(z_class, t_class)
return loss
def update_core(self):
batch = self._iterators['main'].next()
l_batch_origin = [i[2] for i in batch]
changeflag = [i[3] for i in batch]
x = chainer.cuda.to_gpu(np.array([i[0] for i in batch]))
value = self.value
labels = [l[1] for l in batch]
row_idx, col_idx, val_idx = [], [], []
for i in range(len(labels)):
l_list = list(set(labels[i])) # remove duplicate cateories to avoid double count
for y in l_list:
row_idx.append(i)
col_idx.append(y)
val_idx.append(1)
m = len(labels)
n = self.class_dim
t = sp.csr_matrix((val_idx, (row_idx, col_idx)), shape=(m, n), dtype=np.int8).todense()
t = chainer.cuda.to_gpu(t)
optimizer = self._optimizers['main']
y = optimizer.target(x)
loss = F.sigmoid_cross_entropy(y, t)
flag=1
#それぞれのカテゴリーに対して誤差を計算する
for i1,i2,i3,i4 in zip_longest(t,y,l_batch_origin,changeflag):
if changeflag == '1':
for j1,j2 in zip_longest(i1,i2):
myloss = F.sigmoid_cross_entropy(j2,j1) + (j1*(F.log(F.sigmoid(j2))))#ラベルの変更がされている場合のlossの計算
if math.isnan(myloss.data):
myloss = F.sigmoid_cross_entropy(j2,j1)+(j1*(F.log1p(F.sigmoid(j2))))
if flag == 1:
all_loss = myloss
flag = 0
else:
all_loss += myloss
else:
for j1,j2 in zip_longest(i1,i2):
myloss = F.sigmoid_cross_entropy(j2,j1) #ラベルの変更がされていない場合のlossの計算
if flag == 1:
all_loss = myloss
flag = 0
else:
all_loss += myloss
all_loss = all_loss/t.size
optimizer.target.cleargrads()
all_loss.backward()#逆伝播
optimizer.update()#パラメータの更新
chainer.reporter.report({'main/loss':all_loss})#lossの書き出し
def _calculate_kl(W, log_sigma2):
log_alpha = F.clip(log_sigma2 - F.log(W ** 2 + 1e-8), -8., 8.)
clip_mask = (log_alpha.data > loga_threshold)
normalizer = 1. / W.size
reg = (0.63576 * F.sigmoid(1.87320 + 1.48695 * log_alpha)) + \
(- 0.5 * F.log1p(F.exp(- log_alpha))) - 0.63576
xp = cuda.get_array_module(reg)
reg = F.where(clip_mask, xp.zeros(reg.shape).astype('f'), reg)
return - F.sum(reg) * normalizer
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = F.log1p(x)
testing.assert_allclose(
numpy.log1p(self.x), y.data, atol=1e-7, rtol=1e-7)
def log1p(self, x):
return F.log1p(x)
def lossfun_multi_cnn(z, t):
with cuda.get_device(z):
z = z
loss = - (F.mean(t * F.log(z) + (1-t)*F.log1p(-z)))
return loss
def lossfun(self, y, t):
#return F.sigmoid_cross_entropy(y,t)
return (y * (t - (y >= 0)) - F.log1p(F.exp(-F.abs(y))))
def forward(self, inputs, device):
x, = inputs
return functions.log1p(x),
sum_entropy1 = 0.0
sum_entropy2 = 0.0
n = 0
for i in range(0, len(hcpes) - BATCH_SIZE, BATCH_SIZE):
x1, x2 = mini_batch(hcpes[i:i + BATCH_SIZE])
with chainer.no_backprop_mode():
with chainer.using_config('train', False):
y1, y2 = model(x1, x2)
p1 = F.softmax(y1)
#entropy1 = F.sum(- p1 * F.log(p1), axis=1)
y1_max = F.max(y1, axis=1, keepdims=True)
log_p1 = y1 - (
F.log(F.sum(F.exp(y1 - y1_max), axis=1, keepdims=True)) + y1_max)
entropy1 = F.sum(-p1 * log_p1, axis=1)
sum_entropy1 += F.mean(entropy1).data
p2 = F.sigmoid(y2)
#entropy2 = -(p2 * F.log(p2) + (1 - p2) * F.log(1 - p2))
log1p_ey2 = F.log1p(F.exp(y2))
entropy2 = -(p2 * (y2 - log1p_ey2) + (1 - p2) * -log1p_ey2)
sum_entropy2 += F.mean(entropy2).data
n += 1
avr_entropy1 = sum_entropy1 / n
avr_entropy2 = sum_entropy2 / n
print(avr_entropy1, avr_entropy2)