paddle训练时，为什么会出现参数的数量不一致的错误？

[joseph-chan]

在训练一个简单的三层结构的autoencoder时，出现网络的参数size的问题。整个网络结构的代码如下，为什么会出现参数数量不一致错误？ input_layer_dim = 100 hidden_layer_dim = 1024 The network structure is as follow:

      input_attr = ParamAttr(name = 'input.w', initial_mean = 0.,  sparse_update=True,
                l2_rate= 1e-3, initial_std = 1. / np.sqrt(self.input_layer_dim / 2.))

        input_bias = ParamAttr(name='input.bias', initial_mean=0., initial_std=0.)
        hidden_attr = ParamAttr( name = 'hidden.w',
                l2_rate= 1e-3, initial_mean = 0., initial_std = 1. / np.sqrt(self.hidden_layer_dim / 2.))
        hidden_bias = ParamAttr(name='input.bias', initial_mean=0., initial_std=0.)
       self.aeinput = paddle.layer.data(
                name='ae_input',
                type=paddle.data_type.sparse_vector(self.input_layer_dim))
        #hidden layer

        hiddenlayer = paddle.layer.fc(input = self.aeinput, size = self.hidden_layer_dim, param_attr = input_attr,bias_attr = input_bias, act = paddle.activation.Softmax())

        self.predict = paddle.layer.fc(input = hiddenlayer, size = self.input_layer_dim, param_attr = hidden_attr,bias_attr = hidden_bias, act = paddle.activation.Softmax())

出现以下错误：

351 Weight.cpp:28] Check failed: param->getSize() == width * height (1024 vs. 100) 
*** Check failure stack trace: ***
    @     0x7f4d29ca2d7d  google::LogMessage::Fail()
    @     0x7f4d29ca682c  google::LogMessage::SendToLog()
    @     0x7f4d29ca2873  google::LogMessage::Flush()
    @     0x7f4d29ca7d3e  google::LogMessageFatal::~LogMessageFatal()
    @     0x7f4d29b27b32  paddle::Weight::Weight()
    @     0x7f4d299fcd0a  paddle::FullyConnectedLayer::init()
    @     0x7f4d29a60f8f  paddle::NeuralNetwork::init()
    @     0x7f4d29a896d1  paddle::GradientMachine::create()
    @     0x7f4d29c7fca3  GradientMachine::createFromPaddleModelPtr()
    @     0x7f4d29c7fe7f  GradientMachine::createByConfigProtoStr()
    @     0x7f4d2992fd3d  _wrap_GradientMachine_createByConfigProtoStr

[NHZlX]

input_bias = ParamAttr(name='input.bias', initial_mean=0., initial_std=0.)
        hidden_attr = ParamAttr( name = 'hidden.w',
                l2_rate= 1e-3, initial_mean = 0., initial_std = 1. / np.sqrt(self.hidden_layer_dim / 2.))
        hidden_bias = ParamAttr(name='input.bias', initial_mean=0., initial_std=0.)

bias 的问题， 这里相同的名字制定了两次... 一般情况，bias 默认初始值都是0，直接 paddle.layer.fc(input = hiddenlayer, size = self.input_layer_dim,bias_attr = True, act = paddle.activation.Softmax()) 就行

[joseph-chan]

确实是粗心，名字”input.bias“重复了，感谢帮忙指出问题