Simple resnet implementation

vision
1

Hey there,
I am trying to create a simple resnet, possibly with 2-3 resblocks , but having trouble implementing it, can anyone help me here? Thank you

2

hey, just a comment, if you post your code and point out where you are having trouble , you might find it easier to get help on this.

3

Sure, I will do that. Here is what I have so far.

class toy_resnet(nn.Module):

def __init__(self):
    super(toy_resnet,self).__init__()

    self.conv1 = nn.Sequential(
        Conv2d(3, 8, kernel_size=3, stride=1, padding='same'),
        BatchNorm2d(8),
        ReLU(inplace=True),
        Conv2d(8, 8, kernel_size=3, stride=1, padding='same'),
        BatchNorm2d(8),
        ReLU(inplace=True)
         )

    self.conv2 = nn.Sequential(
        Conv2d(8, 8, kernel_size=3, stride=1, padding='same'),
        BatchNorm2d(8),
        ReLU(inplace=True),
        Conv2d(8, 8, kernel_size=3, stride=1, padding='same'),
        BatchNorm2d(8),
        ReLU(inplace=True)
           )

    self.classifier = nn.Sequential(
        nn.Flatten(),
        nn.Linear(24200,37*10),
        nn.Dropout(0.5)
    )
    self.max_pool = MaxPool2d(kernel_size=3, stride=2)
def forward(self,input):
    # x = self.conv1(input.float().to('cuda'))
    # #print(x.shape)
    # x = self.conv2(x.float().to('cuda'))
    # #print(x.shape)
    # x = self.conv3(x.float().to('cuda'))
    #res = input
    x = self.conv1(input.float())
    x = self.max_pool(x)
    res= x

    #x+=input
    #print(x.shape)
    x = self.conv2(x.float())
    out = x+res
    
    x = self.conv2(out.float())
    res = x+out

    out = self.max_pool(res)
    #print(x.shape)
    #x = self.conv3(x.float())
    #print(x.shape)
    x = self.classifier(out)
    return x
4

thanks, so your model is supposed to have 370 output classes with a dropout layer as the last layer ( before softmax presumably) . where are you facing issues here?

5

Well, I just want to know if i am doing it correctly, and my second concern is that it’s not doing so well with comparison to just a simple network with just few conv layers and filters.

6

i guess from the the looks of it the implementation seems fine to me, maybe someone else can comment. the network “not doing so well” as a description is a very sparse to make a comment, i’ll make a general comment. networks with more parameters are prone to overfitting, where their train error goes down but the validation error does not do as well as compared to a network with less parameters. if this is the case , you can search for regularization as a keyword.
good luck

1 Like
7

Thank you for your help.