Issue with output size

vision
1

I’ve been trying to make use of pretrained models, their seems to be some issue along the way:
When input is passed to conv layers, the problem persists.

class Model(nn.Module):
    def __init__(self, pretrained_model):
        super(Model,self).__init__()
        self.pretrained_model = pretrained_model
        self.base = nn.Sequential(*list(pretrained_model.children())[:-1])
        self.linear = nn.Linear(512,340)

    def forward(self, x):
        x = self.base(x)
        x = F.avg_pool2d(x, x.size()[2:])
        f = x.view(x.size(0), -1)
        y = self.linear(f)
        return 

ft_model = torchvision.models.resnet18(pretrained=True)
model = Model(ft_model)

k = model(x)

The error occurs at forward module x=self.base(x),
Given input size: (512x3x3). Calculated output size: (512x-3x-3).

Also this representation of output size is not consistent .What’s going wrong ? I’m on Pytorch master.

2

Yes, It expects channel first like [bs,C,H,W].

You can use permute in this case.

img = img.permute(0,3,1,2)
# this will change [bs,H,W,C] => [bs,C,H,W]
3

I did this already as i had mentioned.

4

what’s the size of input image that you are passing through the network? It seems the input you are passing is having less size than required by pretrained model.