药物副作用预测网络-尝试处理降低损失值

参考资料了解了一些损失值居高不下的原因，

一一进行尝试

1.调整batchsize，初始设置了512，疑似过大，调整为32与16进行尝试

依然稳定在0.32左右，没有较大改变

2.尝试变更初始化矩阵

选择初始化0、均值、正态进行尝试

class MLP(torch.nn.Module):
    def __init__(self,n_feature,n_hidden,num_labels,dropout = 0.5):
        super(MLP, self).__init__()
        self.dropout = torch.nn.Dropout(dropout)
        # 定义隐藏层和输出层
        self.hidden_1 = torch.nn.Linear(n_feature, n_hidden)
        init.normal_(self.hidden_1.weight, mean=0, std=0.01)  # 正态分布初始化隐藏层权重
        self.bn1 = torch.nn.BatchNorm1d(n_hidden)

        # 逐层减少参数数量，有助于防止过拟合
        self.hidden_2 = torch.nn.Linear(n_hidden, n_hidden // 2)
        init.normal_(self.hidden_2.weight, mean=0, std=0.01)  # 正态分布初始化隐藏层权重
        self.bn2 = torch.nn.BatchNorm1d(n_hidden // 2)

        self.hidden_3 = torch.nn.Linear(n_hidden // 2, n_hidden // 4)
        init.normal_(self.hidden_3.weight, mean=0, std=0.01)  # 正态分布初始化隐藏层权重
        self.bn3 = torch.nn.BatchNorm1d(n_hidden // 4)

        self.hidden_4 = torch.nn.Linear(n_hidden // 4, n_hidden // 8)
        init.normal_(self.hidden_4.weight, mean=0, std=0.01)  # 正态分布初始化隐藏层权重
        self.bn4 = torch.nn.BatchNorm1d(n_hidden // 8)

        self.out = torch.nn.Linear(n_hidden // 8, num_labels)
    #前向
    # 传播
    def forward(self, x):
        x = F.relu(self.hidden_1(x))
        x = self.dropout(self.bn1(x))
        x = F.relu(self.hidden_2(x))
        x = self.dropout(self.bn2(x))
        x = F.relu(self.hidden_3(x))
        x = self.dropout(self.bn3(x))
        x = F.relu(self.hidden_4(x))
        x = self.dropout(self.bn4(x))
        x = self.out(x)
        return x

基本没有改善

疑似模型结构或特征提取有异常