自定义Lasagne逐元素乘积层遇维度不匹配问题求助

我有一个形状为**(序列长度19 × 特征数21)的输入序列，输入神经网络后需要一个能对输入与权重执行逐元素乘法（而非点积）**的层，输出形状规则如下：

• 若单元数为1，输出形状保持(19×21)
• 若单元数为8，输出形状应为(8,19,21)（批量维度在前时为(None,8,19,21)，对应扁平化后的(None,8,399)）

我尝试基于Lasagne自定义层，但运行时出现维度不匹配错误，以下是详细代码与错误信息：

class ElementwiseMulLayer(lasagne.layers.Layer):
    def __init__(self, incoming, num_units, W=lasagne.init.Normal(0.01),**kwargs):
        super(ElementwiseMulLayer, self).__init__(incoming, **kwargs)
        self.num_inputs = self.input_shape[1]
        self.num_units = num_units
        self.W = self.add_param(W, (self.num_inputs,num_units), name='W')

    def get_output_for(self, input, **kwargs):
        #return T.dot(input, self.W)
        result=input*self.W
        return result

    def get_output_shape_for(self, input_shape):
        return (input_shape[0], self.num_units,self.num_inputs)  

l_in_2 = lasagne.layers.InputLayer(shape=(None, 9*19*21))
l_reshape_l_in_2 = lasagne.layers.ReshapeLayer(l_in_2, (-1, 9,19,21))
l_reshape_l_in_2_EL = lasagne.layers.ExpressionLayer(l_reshape_l_in_2, lambda X: X[:,0,:,:], output_shape='auto') 
l_reshape_l_in_2_EL = lasagne.layers.ReshapeLayer(l_reshape_l_in_2_EL, (-1, 19*21))
l_out1 = ElementwiseMulLayer(l_reshape_l_in_2_EL, num_units=8, name='my_EW_layer')
l_out1 = lasagne.layers.ReshapeLayer(l_out1, (-1, 8*399))
l_out = lasagne.layers.DenseLayer(l_out1,
                                num_units = 19*21,
                                W = lasagne.init.Normal(),
                                nonlinearity = lasagne.nonlinearities.rectify)   

ValueError: GpuElemwise. Input dimension mis-match. Input 1 (indices start at 0) has shape[0] == 399, but the output's size on that axis is 64.
Apply node that caused the error: GpuElemwise{mul,no_inplace}(GpuReshape{2}.0, my_dot_layer.W)
Toposort index: 23
Inputs types: [GpuArrayType<None>(float32, matrix), GpuArrayType<None>(float32, matrix)]
Inputs shapes: [(64, 399), (399, 8)]
Inputs strides: [(14364, 4), (32, 4)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[GpuReshape{2}(GpuElemwise{mul,no_inplace}.0, TensorConstant{[  -1 3192]})]]

我尝试将W的形状改为(self.num_inputs,num_units, self.num_inputs)，仍出现维度不匹配：

ValueError: GpuElemwise. Input dimension mis-match. Input 1 (indices start at 0) has shape[1] == 8, but the output's size on that axis is 64.
Apply node that caused the error: GpuElemwise{mul,no_inplace}(InplaceGpuDimShuffle{x,0,1}.0, my_EW_layer.W)
Toposort index: 26
Inputs types: [GpuArrayType<None>(float32, (True, False, False)), GpuArrayType<None>(float32, 3D)]
Inputs shapes: [(1, 64, 399), (399, 8, 399)]
Inputs strides: [(919296, 14364, 4), (12768, 1596, 4)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[GpuReshape{2}(GpuElemwise{mul,no_inplace}.0, TensorConstant{[  -1 3192]})]]

问题出在维度广播不匹配：输入是(batch_size, num_inputs)即(64,399)，原W是(399,8)，直接逐元素乘时无法对齐维度。需要调整W的形状并对输入做维度扩展，让广播机制正确工作：

修改自定义层的代码如下：

class ElementwiseMulLayer(lasagne.layers.Layer):
    def __init__(self, incoming, num_units, W=lasagne.init.Normal(0.01),**kwargs):
        super(ElementwiseMulLayer, self).__init__(incoming, **kwargs)
        self.num_inputs = self.input_shape[1]
        self.num_units = num_units
        # 调整W的形状为(num_units, num_inputs)，对应每个单元一组输入维度的权重
        self.W = self.add_param(W, (self.num_units, self.num_inputs), name='W')

    def get_output_for(self, input, **kwargs):
        # 把输入从(batch, num_inputs)扩展为(batch, 1, num_inputs)
        input_expanded = input.dimshuffle(0, 'x', 1)
        # W的形状是(num_units, num_inputs)，广播后变为(1, num_units, num_inputs)
        # 相乘后得到(batch, num_units, num_inputs)，符合预期输出形状
        result = input_expanded * self.W
        return result

    def get_output_shape_for(self, input_shape):
        return (input_shape[0], self.num_units, self.num_inputs)  

原理说明 #

输入扩展维度后形状变为(64,1,399)，W的形状是(8,399)，Lasagne的广播机制会自动将W扩展为(1,8,399)，和输入逐元素相乘后得到(64,8,399)的输出，与定义的输出形状一致，后续Reshape操作也能正常执行。

内容的提问来源于stack exchange，提问作者HATEM EL-AZAB