def __call__(self, x):
h = F.relu(self.l0(x))
h = F.prelu(self.l1(h))
h = F.relu(self.l2(h))
h = F.prelu(self.l3(h))
return h
def check_forward(self, x_data, W_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
masked = numpy.ma.masked_greater_equal(y_expect, 0, copy=False)
shape = (1, ) + W.shape + (1, ) * (x.ndim - W.ndim - 1)
masked *= self.W.reshape(shape)
testing.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, W_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
masked = numpy.ma.masked_greater_equal(y_expect, 0, copy=False)
shape = (1,) + W.shape + (1,) * (x.ndim - W.ndim - 1)
masked *= self.W.reshape(shape)
testing.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, W_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
for i in numpy.ndindex(self.x.shape):
if self.x[i] < 0:
y_expect[i] *= self.W
testing.assert_allclose(y_expect, y.data)
def check_forward(self, x_data, W_data):
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
for i in numpy.ndindex(self.x.shape):
if self.x[i] < 0:
y_expect[i] *= self.W
testing.assert_allclose(
y_expect, y.data)
def test_forward(self, backend_config):
# TODO(niboshi): Support it
if backend_config.use_chainerx and self.dtype == numpy.float16:
raise unittest.SkipTest('ChainerX does not support float16')
x_data, W_data = backend_config.get_array((self.x, self.W))
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
masked = numpy.ma.masked_greater_equal(y_expect, 0, copy=False)
shape = (1, ) + W.shape + (1, ) * (x.ndim - W.ndim - 1)
masked *= self.W.reshape(shape)
testing.assert_allclose(y_expect, y.data)
def test_forward(self, backend_config):
# TODO(niboshi): Support it
if backend_config.use_chainerx and self.dtype == numpy.float16:
raise unittest.SkipTest('ChainerX does not support float16')
x_data, W_data = backend_config.get_array((self.x, self.W))
x = chainer.Variable(x_data)
W = chainer.Variable(W_data)
y = functions.prelu(x, W)
self.assertEqual(y.data.dtype, self.dtype)
y_expect = self.x.copy()
masked = numpy.ma.masked_greater_equal(y_expect, 0, copy=False)
shape = (1,) + W.shape + (1,) * (x.ndim - W.ndim - 1)
masked *= self.W.reshape(shape)
testing.assert_allclose(y_expect, y.data)
def f(x, W):
y = functions.prelu(x, W)
return y * y
def forward(self, inputs, device):
x, W = inputs
y = functions.prelu(x, W)
return y,
def forward(self, inputs, device):
x, W = inputs
y = functions.prelu(x, W)
return y,
def __call__(self, x, t, train):
"""
h = F.max_pooling_2d(F.relu(self.mlpconv1(x)), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.mlpconv2(h)), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.mlpconv3(h)), 3, stride=2)
h = self.mlpconv4(F.dropout(h, train=self.train))
y = F.reshape(F.average_pooling_2d(h, 6), (x.data.shape[0], 1000))
"""
"""
ninのまね
h = F.max_pooling_2d(F.relu(self.conv1(x)), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.conv2(h)), 3, stride=2)
h = F.max_pooling_2d(F.relu(self.conv3(h)), 3, stride=2)
h = self.conv4(F.dropout(h, train=self.train))
y = F.reshape(F.average_pooling_2d(h, 6), (x.data.shape[0], 1000))
"""
#---------------------------------------------60model
#h = self.conv1(x)
#h = F.relu(h)
#h = F.max_pooling_2d(h, 3, stride=2)
#h = self.conv2(h)
#h = F.relu(h)
#h = F.average_pooling_2d(h, 3, stride=2)
#h = self.conv3(h)
#h = F.relu(h)
#h = F.average_pooling_2d(h, 3, stride=2)
#h = self.conv4(h)
#h = F.relu(h)
#h = F.average_pooling_2d(h, 3, stride=2)
#y = F.reshape(F.average_pooling_2d(h, 6), (x.data.shape[0],48))
#----------------------------------------vasily
#h = F.relu(self.bn1(self.conv1(x)))
#h = F.relu(self.bn2(self.conv2(h)))
#h = F.relu(self.bn3(self.conv3(h)))
#h = F.relu(self.bn4(self.conv4(h)))
#y = self.fl(h)
h = self.conv1(x)
#h = F.relu(h)
h = F.prelu(h, xp.ones((8, 221, 221), dtype=xp.float32) * 0.25)
h = F.max_pooling_2d(h, 3, stride=2)
h = self.conv2(h)
#h = F.relu(h)
h = F.prelu(h, xp.ones((16, 106, 106), dtype=xp.float32) * 0.15)
h = F.average_pooling_2d(h, 3, stride=2)
h = self.conv3(h)
#h = F.relu(h)
#h = F.relu(F.dropout(h, ratio=0.3,train=train))
h = F.prelu(h, xp.ones((32, 50, 50), dtype=xp.float32) * 0.05)
h = F.average_pooling_2d(h, 3, stride=2)
h = self.conv4(h)
h = F.prelu(F.dropout(h, ratio=0.4, train=train),
xp.ones((48, 22, 22), dtype=xp.float32) * 0.01)
#h = F.prelu(h,xp.ones((48,22,22),dtype=xp.float32)*0.01)
h = F.average_pooling_2d(h, 3, stride=2)
#y=self.mo(F.dropout(h,ratio=0.5,train=train))
y = F.reshape(F.average_pooling_2d(h, 6), (x.data.shape[0], 48))
"""
#print x.data.ndim
#convolution->bn->reluが安定するらしい・・・
h = self.conv1(x)
#h = self.bn1(h)
h = F.relu(h)
#h = self.bn1(h)
#n= normarize window size
#h = F.local_response_normalization(h,n=3)
#wmax_pooling_2s(x,indowsize,)
h = F.max_pooling_2d(h, 3, stride=3)
#print h.data.shape
h = self.conv2(h)
#h = self.bn2(h)
h = F.relu(h)
#h = F.max_pooling_2d(h, 3, stride=2)
h = F.average_pooling_2d(h, 3, stride=3)
h = self.conv3(h)
#h = self.bn3(h)
h = F.relu(h)
#h = self.bn3(h)
#x, poolingWindowSize, stride=, pad=
#h = F.max_pooling_2d(h, 3, stride=2)
h = F.average_pooling_2d(h, 3, stride=3)
h=self.l1(h)
h=F.relu(h)
h = F.dropout(h, ratio=0.5, train=train)#0.3
y = self.l2(h)
#h = self.bn4(h)
#h = F.relu(h)
##h=self.l2(h)
#h = self.bn4(h)
##h = F.relu(h)
#h=F.relu(self.l2(h))
#y = F.dropout(self.l3(h),ratio=0.3,train=train)#0.2
#h = self.bn5(h)
#y = self.l3(h)
#h = F.relu(self.conv1(x))
#h = self.bn4(h)
#h = F.average_pooling_2d(h, 2)
#h = F.relu(self.conv2(h))
#h = self.bn5(h)
#h = F.average_pooling_2d(h, 2)
#h = F.relu(self.conv3(h))
#h = self.bn6(h)
#h = F.average_pooling_2d(h, 2)
#h = F.relu(self.l1(x))
#h = F.dropout(F.relu(self.l1(h)), train=train)
#y = self.l2(h)
"""
#http://qiita.com/supersaiakujin/items/ccdb41c1f33ad5d27fdf
#活性化関数softmax=exp(a)/sum(exp(a))は、何でもありの入力の値を確率に直す
if train:
#cross_entropyは誤差関数。すべてのラベルについて-sum(log(softmax(y))*Label)する。これを最小化したい。
return F.softmax_cross_entropy(y, t), F.accuracy(y, t)
else:
return F.accuracy(y, t)