DeligientSloth commented 5 years ago

It seems that the convolution kernel parameters of different convolutional layers are the same（self.conv3 = nn.Conv2d(self.channel_size, self.channel_size, (3, 1), stride=1)）， what is purpose？

DeligientSloth commented 5 years ago

老哥应该是个中国人吧，你写的网络怎么看都感觉有点问题，实际效果也不是很好，这么深层的网络竟然欠拟合，我改了一下，增加了一个函数，每次调用都会返回一个卷积核回来，相当于卷积核参数不共享的，你网络里面conv3所有参数都共享，论文里面好像没提到要参数共享： def conv3x3(self): return nn.Conv2d(self.channel_size, self.channel_size, (3, 1), stride=1)

Convolution

    x = self.padding_conv(px)
    x = F.relu(x)
    x = self.conv3x3()(x)

DeligientSloth commented 5 years ago

应该是这样的 `class ResnetBlock(nn.Module): def init(self, channel_size): super(ResnetBlock, self).init()

    self.channel_size = channel_size
    self.maxpool = nn.Sequential(
        nn.ConstantPad1d(padding=(0, 1), value=0),
        nn.MaxPool1d(kernel_size=3, stride=2)
    )
    self.conv = nn.Sequential(
        nn.BatchNorm1d(num_features=self.channel_size),
        nn.ReLU(),
        nn.Conv1d(self.channel_size, self.channel_size, kernel_size=3, padding=1),

        nn.BatchNorm1d(num_features=self.channel_size),
        nn.ReLU(),
        nn.Conv1d(self.channel_size, self.channel_size, kernel_size=3, padding=1),
    )

def forward(self, x):
    x_shortcut = self.maxpool(x)
    x = self.conv(x_shortcut)
    x = x + x_shortcut
    #         print('resnet:{}'.format(x.shape))
    return x

class NeuralNet(nn.Module): """ DPCNN model, 3

region embedding: using TetxCNN to generte
two 3 conv(padding) block

maxpool->3 conv->3 conv with resnet block(padding) feature map: len/2 """

def init(self): super(NeuralNet, self).init() self.channel_size = 250

self.embedding = nn.Embedding(max_features, embed_size)
self.embedding.weight = nn.Parameter(torch.tensor(embedding_matrix, dtype=torch.float32))
self.embedding.weight.requires_grad = False

# region embedding
self.region_embedding = nn.Sequential(
    nn.Conv1d(embed_size, self.channel_size, kernel_size=3, padding=1),
    nn.BatchNorm1d(num_features=self.channel_size),
    nn.ReLU(),
    nn.Dropout(0.2)
)
self.conv_block = nn.Sequential(
    nn.BatchNorm1d(num_features=self.channel_size),
    nn.ReLU(),
    nn.Conv1d(self.channel_size, self.channel_size, kernel_size=3, padding=1),
    nn.BatchNorm1d(num_features=self.channel_size),
    nn.ReLU(),
    nn.Conv1d(self.channel_size, self.channel_size, kernel_size=3, padding=1),
)

self.seq_len = maxlen
resnet_block_list = []
while (self.seq_len > 2):
    resnet_block_list.append(ResnetBlock(self.channel_size))
    self.seq_len = self.seq_len // 2
#         print('seqlen{}'.format(self.seq_len))
self.resnet_layer = nn.Sequential(*resnet_block_list)

self.linear_out = nn.Linear(self.seq_len * self.channel_size, 1)  # 改成输出一个值

def forward(self, x): batch = x.shape[0] print("===========start==============")

x = self.embedding(x)
x = x.permute(0, 2, 1)
x = self.region_embedding(x)
#         print(x.shape)

x = self.conv_block(x)
#         print(x.shape)

x = self.resnet_layer(x)

x = x.permute(0, 2, 1)
#         print(x.shape)
#         print("===========end==============")
x = x.contiguous().view(x.size(0), -1)
out = self.linear_out(x)
return out

`