Converting a VGG model into sequential, but getting different outputs

vision
1

Background:
I’m working on an adversarial detector method which requires to access the outputs from each hidden layer (Local Intrinsic Dimensionality Detector - a.k.a.: LID).
I loaded a pretrained VGG16 from torchvision.models.

To access the output from each hidden layer, I put it into a sequential model:

vgg16 = models.vgg16(pretrained=True)

vgg16_seq = nn.Sequential(*(
    list(list(vgg16.children())[0]) + 
    [nn.AdaptiveAvgPool2d((7, 7)), nn.Flatten()] + 
    list(list(vgg16.children())[2])))

I looked into the torchvision VGG implementation, it uses the [feature..., AvgPool, flatten, classifier...] structure.
Since AdaptiveAvgPool2d layer and Flatten layer have no parameters, I assume this should work, but I have different outputs.

output1 = vgg16(X_small)
print(output1.size())
output2 = vgg16_seq(X_small)
print(output2.size())
torch.equal(output1, output2)

They are in the same dimension but different outputs.

torch.Size([32, 1000])
torch.Size([32, 1000])
False

I tested the outputs right after the AdaptiveAvgPool2d layer, the outputs are equal:

output1 = nn.Sequential(*list(vgg16.children())[:2])(X_small)
print(output1.size())
output2 = nn.Sequential(*list(vgg16_seq)[:32])(X_small)
print(output2.size())
torch.equal(output1, output2)

torch.Size([32, 512, 7, 7])
torch.Size([32, 512, 7, 7])
True

Can someone point out what went wrong?
Thank you

2

By adding .eval(), both models get the same result.

vgg16.eval()
output1 = vgg16(X_small)
print(output1.size())

vgg16_seq.eval()
output2 = vgg16_seq(X_small)
print(output2.size())
torch.equal(output1, output2)

But I’m not sure why.

3

VGG16 uses dropout layers, which will be disabled when model.eval() is used and which explains the difference during the model.train() comparison.

1 Like