当前位置: 首页>>代码示例>>Python>>正文


Python ZoomableAttentionWindow.write方法代码示例

本文整理汇总了Python中attention.ZoomableAttentionWindow.write方法的典型用法代码示例。如果您正苦于以下问题:Python ZoomableAttentionWindow.write方法的具体用法?Python ZoomableAttentionWindow.write怎么用?Python ZoomableAttentionWindow.write使用的例子?那么, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在attention.ZoomableAttentionWindow的用法示例。


在下文中一共展示了ZoomableAttentionWindow.write方法的3个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。

示例1: AttentionWriter

# 需要导入模块: from attention import ZoomableAttentionWindow [as 别名]
# 或者: from attention.ZoomableAttentionWindow import write [as 别名]
class AttentionWriter(Initializable):
    def __init__(self, input_dim, output_dim, channels, width, height, N, **kwargs):
        super(AttentionWriter, self).__init__(name="writer", **kwargs)

        self.channels = channels
        self.img_width = width
        self.img_height = height
        self.N = N
        self.input_dim = input_dim
        self.output_dim = output_dim

        assert output_dim == channels*width*height

        self.zoomer = ZoomableAttentionWindow(channels, height, width, N)
        self.z_trafo = Linear(
                name=self.name+'_ztrafo',
                input_dim=input_dim, output_dim=5, 
                weights_init=self.weights_init, biases_init=self.biases_init,
                use_bias=True)

        self.w_trafo = Linear(
                name=self.name+'_wtrafo',
                input_dim=input_dim, output_dim=channels*N*N, 
                weights_init=self.weights_init, biases_init=self.biases_init,
                use_bias=True)

        self.children = [self.z_trafo, self.w_trafo]

    @application(inputs=['h'], outputs=['c_update'])
    def apply(self, h):
        w = self.w_trafo.apply(h)
        l = self.z_trafo.apply(h)

        center_y, center_x, delta, sigma, gamma = self.zoomer.nn2att(l)

        c_update = 1./gamma * self.zoomer.write(w, center_y, center_x, delta, sigma)

        return c_update

    @application(inputs=['h'], outputs=['c_update', 'center_y', 'center_x', 'delta'])
    def apply_detailed(self, h):
        w = self.w_trafo.apply(h)
        l = self.z_trafo.apply(h)

        center_y, center_x, delta, sigma, gamma = self.zoomer.nn2att(l)

        c_update = 1./gamma * self.zoomer.write(w, center_y, center_x, delta, sigma)

        return c_update, center_y, center_x, delta

    @application(inputs=['x','h'], outputs=['c_update', 'center_y', 'center_x', 'delta'])
    def apply_circular(self,x,h):
        #w = self.w_trafo.apply(h)
        l = self.z_trafo.apply(h)

        center_y, center_x, delta, sigma, gamma = self.zoomer.nn2att(l)

        c_update = 1./gamma * self.zoomer.write(x, center_y, center_x, delta, sigma)

        return c_update, center_y, center_x, delta
开发者ID:drewlinsley,项目名称:draw_classify,代码行数:62,代码来源:draw.py

示例2: AttentionWriter

# 需要导入模块: from attention import ZoomableAttentionWindow [as 别名]
# 或者: from attention.ZoomableAttentionWindow import write [as 别名]
class AttentionWriter(Initializable):
    def __init__(self, input_dim, output_dim, width, height, N, **kwargs):
        super(AttentionWriter, self).__init__(name="writer", **kwargs)

        self.img_width = width
        self.img_height = height
        self.N = N
        self.input_dim = input_dim
        self.output_dim = output_dim

        assert output_dim == width * height

        self.zoomer = ZoomableAttentionWindow(height, width, N)
        self.z_trafo = Linear(
            name=self.name + "_ztrafo",
            input_dim=input_dim,
            output_dim=5,
            weights_init=self.weights_init,
            biases_init=self.biases_init,
            use_bias=True,
        )

        self.w_trafo = Linear(
            name=self.name + "_wtrafo",
            input_dim=input_dim,
            output_dim=N * N,
            weights_init=self.weights_init,
            biases_init=self.biases_init,
            use_bias=True,
        )

        self.children = [self.z_trafo, self.w_trafo]

    @application(inputs=["h"], outputs=["c_update"])
    def apply(self, h):
        w = self.w_trafo.apply(h)
        l = self.z_trafo.apply(h)

        center_y, center_x, delta, sigma, gamma = self.zoomer.nn2att(l)

        c_update = 1.0 / gamma * self.zoomer.write(w, center_y, center_x, delta, sigma)

        return c_update

    @application(inputs=["h"], outputs=["c_update", "center_y", "center_x", "delta"])
    def apply_detailed(self, h):
        w = self.w_trafo.apply(h)
        l = self.z_trafo.apply(h)

        center_y, center_x, delta, sigma, gamma = self.zoomer.nn2att(l)

        c_update = 1.0 / gamma * self.zoomer.write(w, center_y, center_x, delta, sigma)

        return c_update, center_y, center_x, delta
开发者ID:zan2434,项目名称:draw,代码行数:56,代码来源:draw.py

示例3: main

# 需要导入模块: from attention import ZoomableAttentionWindow [as 别名]
# 或者: from attention.ZoomableAttentionWindow import write [as 别名]
def main(name, epochs, batch_size, learning_rate):
    if name is None:
        name = "att-rw" 

    print("\nRunning experiment %s" % name)
    print("         learning rate: %5.3f" % learning_rate) 
    print()


    #------------------------------------------------------------------------

    img_height, img_width = 28, 28
    
    read_N = 12
    write_N = 14

    inits = {
        #'weights_init': Orthogonal(),
        'weights_init': IsotropicGaussian(0.001),
        'biases_init': Constant(0.),
    }
    
    x_dim = img_height * img_width

    reader = ZoomableAttentionWindow(img_height, img_width,  read_N)
    writer = ZoomableAttentionWindow(img_height, img_width, write_N)

    # Parameterize the attention reader and writer
    mlpr = MLP(activations=[Tanh(), Identity()], 
                dims=[x_dim, 50, 5], 
                name="RMLP",
                **inits)
    mlpw = MLP(activations=[Tanh(), Identity()],
                dims=[x_dim, 50, 5],
                name="WMLP",
                **inits)

    # MLP between the reader and writer
    mlp = MLP(activations=[Tanh(), Identity()],
                dims=[read_N**2, 300, write_N**2],
                name="MLP",
                **inits)

    for brick in [mlpr, mlpw, mlp]:
        brick.allocate()
        brick.initialize()

    #------------------------------------------------------------------------
    x = tensor.matrix('features')

    hr = mlpr.apply(x)
    hw = mlpw.apply(x)

    center_y, center_x, delta, sigma, gamma = reader.nn2att(hr)
    r = reader.read(x, center_y, center_x, delta, sigma)

    h = mlp.apply(r)

    center_y, center_x, delta, sigma, gamma = writer.nn2att(hw)
    c = writer.write(h, center_y, center_x, delta, sigma) / gamma
    x_recons = T.nnet.sigmoid(c)

    cost = BinaryCrossEntropy().apply(x, x_recons)
    cost.name = "cost"

    #------------------------------------------------------------
    cg = ComputationGraph([cost])
    params = VariableFilter(roles=[PARAMETER])(cg.variables)

    algorithm = GradientDescent(
        cost=cost, 
        params=params,
        step_rule=CompositeRule([
            RemoveNotFinite(),
            Adam(learning_rate),
            StepClipping(3.), 
        ])
        #step_rule=RMSProp(learning_rate),
        #step_rule=Momentum(learning_rate=learning_rate, momentum=0.95)
    )


    #------------------------------------------------------------------------
    # Setup monitors
    monitors = [cost]
    #for v in [center_y, center_x, log_delta, log_sigma, log_gamma]:
    #    v_mean = v.mean()
    #    v_mean.name = v.name
    #    monitors += [v_mean]
    #    monitors += [aggregation.mean(v)]

    train_monitors = monitors[:]
    train_monitors += [aggregation.mean(algorithm.total_gradient_norm)]
    train_monitors += [aggregation.mean(algorithm.total_step_norm)]

    # Live plotting...
    plot_channels = [
        ["cost"],
    ]

#.........这里部分代码省略.........
开发者ID:Philip-Bachman,项目名称:NN-Python,代码行数:103,代码来源:run-att-rw.py


注:本文中的attention.ZoomableAttentionWindow.write方法示例由纯净天空整理自Github/MSDocs等开源代码及文档管理平台,相关代码片段筛选自各路编程大神贡献的开源项目,源码版权归原作者所有,传播和使用请参考对应项目的License;未经允许,请勿转载。