def forward(self, x):
x = functional.reshape(x, shape=(x.shape[0], -1))
maximum = 2**self._num_bits - 1
w = self._w * self._factor
w = w.round()
w = functional.where(w < maximum, w, maximum)
w = functional.where(w > -maximum, w, -maximum)
w /= self._factor
return functional.dot(x, w) + self._b
def forward(self, x):
maximum = 2**self._weight_bits - 1
w = self._k / self._k_max
w *= maximum
w = w.round()
w = functional.where(w > maximum, maximum, w)
w = functional.where(w < -maximum, -maximum, w)
w /= maximum
w += functional.random_normal(size=w.shape, mean=0.0, std=self._noise)
w *= self._k_max
return functional.conv2d(x, w, self._stride,
self._padding) + functional.reshape(
self._b, (1, -1, 1, 1))
def forward(self, x):
x = functional.reshape(x, shape=(x.shape[0], -1))
maximum = 2**self._weight_bits - 1
w = self._w / self._w_max
w *= maximum
w = w.round()
w = functional.where(w > maximum, maximum, w)
w = functional.where(w < -maximum, -maximum, w)
w /= maximum
w += functional.random_normal(size=w.shape, mean=0.0, std=self._noise)
w *= self._w_max
return functional.dot(x, w) + self._b
def forward(self, x):
x = functional.reshape(x, shape=(x.shape[0], -1))
maximum = 2**self._num_bits - 1
w = self._w * self._factor
w = functional.where(w < maximum, w, maximum)
w = functional.where(w > -maximum, w, -maximum)
w = w.round()
w += functional.random_normal(w.shape, 0,
self._noise * self._one_cell_factor,
w.dtype)
w /= self._factor
return functional.dot(x, w) + self._b
MaxPool2d([2, 2], 2),
Conv2d([50, 20, 5, 5]),
ReLU(),
MaxPool2d([2, 2], 2),
Linear([800, 1250]),
ReLU(),
Linear([1250, 120]),
ReLU(),
Linear([120, 10]),
Softmax()
],
name='mnist-lenet')
x = functional.placeholder('x', dims=2)
y = functional.placeholder('y', dims=1, dtype='int32')
y_ = net.forward(functional.reshape(x, (-1, 1, 28, 28)))
loss = CrossEntropy().minimize(y_, OneHot(10).turn(y))
accuracy = Accuracy().measure(y_, y)
updates = SGD(learning_rate=0.05,
momentum=0.9).updates(net.parameters(),
net.differentiate(loss))
print('Begin compile')
train_op = compile(inputs=[x, y], outputs=[accuracy], updates=updates)
print('Compiled train_op')
test_op = compile(inputs=[x, y], outputs=[accuracy])
print('Compiled test_op')
batch_size = int(batch)
def forward(self, x):
k = functional.where(self._mask, self._k, 0)
return functional.conv2d(x, k, self._stride,
self._padding) + functional.reshape(
self._b, (1, -1, 1, 1))
def forward(self, x):
x = functional.reshape(x, shape=(x.shape[0], -1))
w = functional.where(self._mask, self._w, 0)
return functional.dot(x, w) + self._b
def forward(self, x):
return functional.conv2d(x, self._k, self._stride, self._padding) + functional.reshape(self._b, (1, -1, 1, 1))
def forward(self, x):
x = functional.reshape(x, shape=(x.shape[0], -1))
return functional.dot(x, self._w) + self._b