def __init__(self):
super(VGG, self).__init__()
links = [('conv1_1', L.Convolution2D(3, 64, 3, stride=1, pad=1))]
links += [('conv1_2', L.Convolution2D(64, 64, 3, stride=1, pad=1))]
links += [('_mpool1', F.MaxPooling2D(2, 2, 0, True, True))]
links += [('conv2_1', L.Convolution2D(64, 128, 3, stride=1, pad=1))]
links += [('conv2_2', L.Convolution2D(128, 128, 3, stride=1, pad=1))]
links += [('_mpool2', F.MaxPooling2D(2, 2, 0, True, True))]
links += [('conv3_1', L.Convolution2D(128, 256, 3, stride=1, pad=1))]
links += [('conv3_2', L.Convolution2D(256, 256, 3, stride=1, pad=1))]
links += [('conv3_3', L.Convolution2D(256, 256, 3, stride=1, pad=1))]
links += [('_mpool3', F.MaxPooling2D(2, 2, 0, True, True))]
links += [('conv4_1', L.Convolution2D(256, 512, 3, stride=1, pad=1))]
links += [('conv4_2', L.Convolution2D(512, 512, 3, stride=1, pad=1))]
links += [('conv4_3', L.Convolution2D(512, 512, 3, stride=1, pad=1))]
links += [('_mpool4', F.MaxPooling2D(2, 2, 0, True, True))]
links += [('conv5_1', L.Convolution2D(512, 512, 3, stride=1, pad=1))]
links += [('conv5_2', L.Convolution2D(512, 512, 3, stride=1, pad=1))]
links += [('conv5_3', L.Convolution2D(512, 512, 3, stride=1, pad=1))]
links += [('_mpool5', F.MaxPooling2D(2, 2, 0, True, True))]
links += [('fc6', L.Linear(25088, 4096))]
links += [('_dropout6', F.Dropout(0.5))]
links += [('fc7', L.Linear(4096, 4096))]
links += [('_dropout7', F.Dropout(0.5))]
links += [('fc8', L.Linear(4096, 1000))]
for link in links:
if not link[0].startswith('_'):
self.add_link(*link)
self.forward = links
def check_backward(self, x_data, y_grad):
dropout = functions.Dropout(self.ratio)
def f(x):
return dropout.apply((x, ))[0]
gradient_check.check_backward(f, x_data, y_grad,
**self.check_double_backward_options)
def check_immutable(self, inputs, backend_config):
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
with backend_config:
dropout = functions.Dropout(0.5)
y1, = dropout.apply(inputs)
y2, = dropout.apply(inputs)
testing.assert_allclose(y1.data, y2.data)
def step(self, hs, h, id):
attention_out = self.attention(hs, h)
embed_out = self.embed(id)
embed_out = self.embed_proj(embed_out)
next_h = attention_out + h + embed_out
for layer_id in range(self.layer_num):
next_h = getattr(self, "rnn_cell_" + str(layer_id))(next_h)
next_h = F.Dropout(0.75)(next_h)
out = self.out(next_h)
return next_h, out
def check_backward(self, inputs, grad_outputs, backend_config):
if backend_config.use_cuda:
inputs = cuda.to_gpu(inputs)
grad_outputs = cuda.to_gpu(grad_outputs)
# Instantiate the function class directly in order to reuse the mask,
# because f will be called repeatedly.
dropout = functions.Dropout(self.ratio)
def f(*inputs):
return dropout.apply(inputs)
with backend_config:
gradient_check.check_backward(f, inputs, grad_outputs,
**self.check_backward_options)
def check_immutable(self, x_data):
d = functions.Dropout(0.5)
y1, = d.apply((chainer.Variable(x_data), ))
y2, = d.apply((chainer.Variable(x_data), ))
testing.assert_allclose(y1.data, y2.data)
def __call__(self, x):
temp = F.relu(self.l1(x)+self.state)
self.state = F.Dropout(F.relu(self.l2(temp)))
return self.l3(self.state)
