如题，我用lstm做一个时序数据分类预测，损失函数是交叉熵损失，结果train loss很快降到0.69，之后也缓慢下降；但val loss从一开始训练就在0.69左右波动，没有下降和收敛的兆头，一些参数设置如下，百度了试过改参数初始化、加bn、改lstm层数、还把adam换了都不行，loss根本不下降，求求救救我！！（相同的数据用SVM做分类，准确率还可，但到lstm不知道咋了

class NewModel(paddle.nn.Layer):
    def __init__(self,input_size,hidden_size,num_layers,drop_out,num_classes):
        super(NewModel,self).__init__()
        self.lstm=paddle.nn.LSTM(input_size=input_size,hidden_size=hidden_size,num_layers=num_layers,dropout=drop_out)
        self.bn=paddle.nn.BatchNorm1D(hidden_size）
        self.fc=paddle.nn.Linear(hidden_size,out_features=num_classes,weight_attr=paddle.ParamAttr(initializer=paddle.nn.initializer.XavierNormal()),bias_attr=paddle.ParamAttr(initializer=paddle.nn.initializer.Constant(0.2)))
    def forward(self,inputs,sequence_len):
        y,(h,c)=self.lstm(inputs,sequence_length=sequence_len) # batch_size*hiddensize
        outputs=paddle.squeeze(y[:,-1,:],axis=1) # batch_size*1*hidden_size
        outputs=self.bn(outputs)
        outputs=self.fc(outputs)
        return outputs

network=NewModel(input_size=feauture_n,hidden_size=128,num_layers=2,drop_out=0,num_classes=2)

LR=1e-5
epoches=200
batch_size=128*3
model=paddle.Model(network)
optimizer=paddle.optimizer.Adam(learning_rate=LR,parameters=model.parameters(),weight_decay=5e-4)
loss = paddle.nn.CrossEntropyLoss()

metric = paddle.metric.Accuracy()