def nlogn_loss(prediction, label):
residual = prediction * 255 - label * 255
diff_abs = F.absolute(residual) + 1
loss = F.mean(diff_abs * F.log2(diff_abs) / 256)
return loss
def compute_entro_X(x):
#エントロピー的なものを計算
delta = chainer.Variable(1e-6 * cp.ones((x.shape[0], 1, x.shape[2], x.shape[3]), dtype = np.float32 ))
#ch平均
x_m = F.relu(F.mean(x, axis=1, keepdims = True))
#ノーマライズ
x_n = x_m / (F.sum(x_m) / x_m.shape[0] + delta) * 100
#0-1にクリップ
x_n = F.clip(x_n, 0, 1)
#ピクセルごとにエントロピー*を計算
entro = x_n * F.log2( 1/(x_n + delta)) + (1 - x_n) * F.log2( 1/(1 - x_n + delta))
#ピクセルのエントロピーを合計
entro = F.sum(entro)
return cp.asnumpy(entro.data) # ,cp.asnumpy(x_n.data)*255
def final(self, x):
xp = self.xp
x = x.astype(np.int32)
p = F.softmax(self.d_pred, axis=1)
h = xp.eye(self.q_num)[x]
d = xp.sum(h, axis=(0, 2, 3))
info = -F.sum(d * F.log2(p)) / x.shape[0]
return info
def forward(x_data, y_data, train=True):
x, t = Variable(x_data), Variable(y_data)
h1 = F.dropout(F.relu(model.l1(x)), train=train)
h2 = F.dropout(F.relu(model.l2(h1)), train=train)
y = model.l3(h2)
y_prob = F.softmax(y)
entropy = -F.sum((F.log2(y_prob) * y_prob))
if train:
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
if not train:
return entropy
def center_corner_nlogn_loss(prediction, label):
residual = prediction * 255 - label * 255
diff_abs = F.absolute(residual) + 1
diff_abs.array[:, :, 0:4, 0:4] += diff_abs.array[:, :, 0:4, 0:4]
diff_abs.array[:, :, 0:4, -4:] += diff_abs.array[:, :, 0:4, -4:]
diff_abs.array[:, :, -4:, 0:4] += diff_abs.array[:, :, -4:, 0:4]
diff_abs.array[:, :, -4:, -4:] += diff_abs.array[:, :, -4:, -4:]
diff_abs.array[:,:,diff_abs.shape[2] // 2 - 4: diff_abs.shape[2] // 2 + 4, diff_abs.shape[3] // 2 - 4: diff_abs.shape[3] // 2 + 4 ] += \
diff_abs.array[:,:,diff_abs.shape[2] // 2 - 4: diff_abs.shape[2] // 2 + 4, diff_abs.shape[3] // 2 - 4: diff_abs.shape[3] // 2 + 4 ]
loss = F.mean(diff_abs * F.log2(diff_abs) / 256)
return loss
def _predict(samples, mode='variance',
reduce_mean=None, reduce_var=None,
eps=1e-8):
mean = F.mean(samples, axis=0)
if mode == 'variance':
var = samples - mean
var = F.mean(F.square(var), axis=0)
elif mode == 'entropy':
var = - mean * F.log2(mean + eps)
else:
raise NotImplementedError('unsupported mode..')
if reduce_mean is not None:
mean = reduce_mean(mean)
if reduce_var is not None:
var = reduce_var(var)
return mean, var
def _get_entropy(self, values):
return F.average((-values * F.log2(F.sigmoid(values)
+ self.discriminator_value_offset)
- (1 - values) * F.log2(1
- F.sigmoid(values)
+ self.discriminator_value_offset))) # NOQA
def log2(self, x):
return F.log2(x)