Help with model

1

Hello, i am using dataset for detection of road that consist of 500 training images and masks. When other guys used this dataset they had PERFECT results and let me show you my result.
Result of training - output - YouTube

I want to ask if you can help with this model, what to improve on this model to get better results? I tried all possible learning rates and epochs. So this is problem in model not in training. Thank you.

I am using this model for it.

def double_conv(in_channels, out_channels):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, 3, padding=1),
        torch.nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True),
        nn.Conv2d(out_channels, out_channels, 3, padding=1),
        torch.nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True)
    )

class Model(nn.Module):
    def __init__(self, n_class=5):
        super(Model).__init__()

        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

        self.dconv_down1 = double_conv(3, 64)
        self.dconv_down2 = double_conv(64, 128)
        self.dconv_down3 = double_conv(128, 256)
        self.dconv_down4 = double_conv(256, 512)

        self.maxool = nn.MaxPool2d(2)
        self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)

        self.dconv_up3 = double_conv(256 + 512, 256)
        self.dconv_up2 = double_conv(128 + 256, 128)
        self.dconv_up1 = double_conv(128 + 64, 64)

        self.conv_last = nn.Conv2d(64, n_class, 1)

    def forward(self, x):
        conv1 = self.dconv_down1(x)
        x = self.maxpool(conv1)

        conv2 = self.dconv_down2(x)
        x = self.maxpool(conv2)

        conv3 = self.dconv_down3(x)
        x = self.maxpool(conv3)

        x = self.dconv_down4(x)

        x = self.upsample(x)
        x = torch.cat([x, conv3], dim=1)

        x = self.dconv_up3(x)
        x = self.upsample(x)
        x = torch.cat([x, conv2], dim=1)

        x = self.dconv_up2(x)
        x = self.upsample(x)
        x = torch.cat([x, conv1], dim=1)

        x = self.dconv_up1(x)

        out = self.conv_last(x)

        return out
2

I don’t know which criterion you are using, but adding the dice loss might help for segmentation use cases.
Also, you could try to add skip connections into your conv blocks, which seemed to work out pretty well in the past.

1 Like
3

Hello, thank you for reply, iam using this type of loss- MSE

loss = ((y - y_pred) ** 2).mean()

Please can you give me example how to use “dice loss” ? I was searching on internet but I didn’t get any solution for this model.

4

Assuming you are working on a multi-class segmentation use case, I would use nn.CrossEntropyLoss as the “base” loss instead of MSELoss.
For the dice loss implementation, you could have a look at this issue or the implementation from torchgeometry.