0

Supposing we want to add a new layer, say a linear layer, to the end of the classifier of another model, such as VGG16, why exactly do these two implementations lead to different results? More specifically, I don't understand why the first implementation produces 2 classfiers:

vgg = torchvision.models.vgg16(pretrained=True) vgg.classifer=nn.Sequential(vgg.classifier, nn.Linear(4096,300)) print(vgg)

output:

 VGG( (features): Sequential( (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU(inplace) (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (3): ReLU(inplace) (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (6): ReLU(inplace) (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (8): ReLU(inplace) (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (11): ReLU(inplace) (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (13): ReLU(inplace) (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (15): ReLU(inplace) (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (18): ReLU(inplace) (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (20): ReLU(inplace) (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (22): ReLU(inplace) (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (25): ReLU(inplace) (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (27): ReLU(inplace) (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (29): ReLU(inplace) (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (classifier): Sequential( (0): Linear(in_features=25088, out_features=4096, bias=True) (1): ReLU(inplace) (2): Dropout(p=0.5) (3): Linear(in_features=4096, out_features=4096, bias=True) (4): ReLU(inplace) (5): Dropout(p=0.5) (6): Linear(in_features=4096, out_features=1000, bias=True) ) (classifer): Sequential( (0): Sequential( (0): Linear(in_features=25088, out_features=4096, bias=True) (1): ReLU(inplace) (2): Dropout(p=0.5) (3): Linear(in_features=4096, out_features=4096, bias=True) (4): ReLU(inplace) (5): Dropout(p=0.5) (6): Linear(in_features=4096, out_features=1000, bias=True) ) (1): Linear(in_features=4096, out_features=300, bias=True) ) )

implementation2:

vgg = models.vgg16(pretrained=True) vgg=nn.Sequential(vgg, nn.Linear(4096,300)) print(vgg)

output:

Sequential( (0): VGG( (features): Sequential( (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (1): ReLU(inplace) (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (3): ReLU(inplace) (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (6): ReLU(inplace) (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (8): ReLU(inplace) (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (11): ReLU(inplace) (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (13): ReLU(inplace) (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (15): ReLU(inplace) (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (18): ReLU(inplace) (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (20): ReLU(inplace) (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (22): ReLU(inplace) (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (25): ReLU(inplace) (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (27): ReLU(inplace) (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (29): ReLU(inplace) (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False) ) (classifier): Sequential( (0): Linear(in_features=25088, out_features=4096, bias=True) (1): ReLU(inplace) (2): Dropout(p=0.5) (3): Linear(in_features=4096, out_features=4096, bias=True) (4): ReLU(inplace) (5): Dropout(p=0.5) (6): Linear(in_features=4096, out_features=1000, bias=True) ) ) (1): Linear(in_features=4096, out_features=300, bias=True) )
Improve this question

1 Answer 1

Reset to default
2

It's because you have a syntax error in the spelling of classifier. You have written it as

vgg.classifer=nn.Sequential(vgg.classifier, nn.Linear(4096,300))

Note the missing i after f in classifier on LHS. So, you're inadvertently creating a new group of layers named classifer by this line.

After correction:

vgg.classifier=nn.Sequential(vgg.classifier, nn.Linear(4096,300))

Moreover, in the first example, you're replacing the existing classifier layer with a sequential network having classifier portion of original vgg and Linear layer as the last layer.

In the second example, you're recreating the variable vgg with a new sequential network which contains the original vgg network with the addition of Linear layer as the last layer.

vgg=nn.Sequential(vgg, nn.Linear(4096,300))

Note the difference between the above two.

Improve this answer
Sign up to request clarification or add additional context in comments.

6 Comments

Lol thanks a lot, can't believe I was struggling so hard all for a syntax error :D Also, I don't quite understand the difference between putting the last layer in the classifier or at the top of the whole network...both produce the same results after running the forward function right?
@GhazalSahebzamani Yes, both yield the same result. I wanted to clarify the difference between the code. But note that the last layer of original vgg has out_features 1000 but the Linear layer you added has 4096 in_features. You'd want to change that.
I see. Thanks again for your help.
Glad to help @GhazalSahebzamani. You can accept the answer if your problem is solved :)
Sorry, I'm not really familiar with this place... how can I do that?
|

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.