def __call__(self, x, target, bf=False):
y = self.glimpse_forward(x)
if chainer.config.train:
loss = F.softmax_cross_entropy(y, target)
return loss
else:
y = F.softmax(y)
index = xp.array(range(y.data.shape[0]))
acc = y.data[index, target]
acc = chainer.cuda.to_cpu(acc)
return acc.sum()
def __call__(self, x, target):
h = F.relu(self.norm_1_1(self.cnn_1_1(x)))
h = F.relu(self.norm_1_2(F.max_pooling_2d(self.cnn_1_2(h), 2, stride=2)))
h = F.relu(self.norm_2_1(self.cnn_2_1(h)))
h = F.relu(self.norm_2_2(F.max_pooling_2d(self.cnn_2_2(h), 2, stride=2)))
h = F.relu(self.norm_3_1(self.cnn_3_1(h)))
h = F.relu(self.norm_3_2(self.cnn_3_2(h)))
h = F.relu(self.norm_3_3(F.max_pooling_2d(self.cnn_3_3(h), 2, stride=2)))
h = F.relu(self.norm_f1(self.full_1(h)))
y = self.full_2(h)
if chainer.config.train:
loss = F.softmax_cross_entropy(y, target)
return loss
else:
y = F.softmax(y)
index = xp.array(range(y.data.shape[0]))
acc = y.data[index, target]
acc = chainer.cuda.to_cpu(acc)
return acc.sum()