About torchvision.models

I use pytorch to create model ,this model has dilation >2 layer ?

resnet101 = torchvision.models.resnet101(pretrained=True)

All conv layers should have a dilation of (1, 1):

resnet101 = models.resnet101(pretrained=False)

for name, module in resnet101.named_modules():
    if isinstance(module, nn.Conv2d):
        print(name, module.dilation)

If I want to set a dilation convolution of more than 2 in the last 2 blocks, and I want them to have a step size of 1, how can I change it?

You could change the dilation by simply assigning a new value to the module:

resnet101.layer4[2].conv1.dilation = (2, 2)

However, since the outputs might be used on a residual connection etc. you would have to make sure to provide the desired spatial size of the new output.