Adding a Softmax Layer to Alexnet's Classifier

vision
1

Hi All,
I’m trying to remodel alexnet to a binary classifier. I wanted to add a Softmax layer to the classifier of the pretrained AlexNet to interpret the output of the last layer as probabilities. Till now the code I have written is -

model_ft = models.alexnet(pretrained=True)
# Frozen the weights of the cnn layers towards the beginning 
layers_to_freeze = [model_ft.features[0],model_ft.features[3],model_ft.features[6],model_ft.features[8]]
for layer in layers_to_freeze:
    for params in layer.parameters():
        params.requires_grad = False 

model_ft.classifier[6] = nn.Linear(in_features=4096, out_features=2, bias=True)
model_ft = nn.Sequential(model_ft,nn.Softmax(1))

And after this the map of the model shows -

Sequential(
  (0): AlexNet(
    (features): Sequential(
      (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
      (1): ReLU(inplace)
      (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
      (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
      (4): ReLU(inplace)
      (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
      (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (7): ReLU(inplace)
      (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (9): ReLU(inplace)
      (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (11): ReLU(inplace)
      (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
    )
    (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
    (classifier): Sequential(
      (0): Dropout(p=0.5)
      (1): Linear(in_features=9216, out_features=4096, bias=True)
      (2): ReLU(inplace)
      (3): Dropout(p=0.5)
      (4): Linear(in_features=4096, out_features=4096, bias=True)
      (5): ReLU(inplace)
      (6): Linear(in_features=4096, out_features=2, bias=True)
    )
  )
  (1): Softmax()
)

Even after adding this layer, I’m getting output tensors of dimension (4,2)

output = model(inputs)
print(output.shape)

This outputs - torch.Size([4, 2])
Why is the output of shape(4,2) ? Shouldn’t it be of shape (1,2), shown in this answer ?

2

The first dimension is the batch_size, so if you are passing 4 samples, the output will have the shape [4, 2].

3

Thanks for the explanation! Is there any way I could add the Softmax layer to the classifier as opposed to an isolated layer at the end ?

4

This should work:

model.classifier[6] = nn.Linear(in_features=4096, out_features=2, bias=True)
model.classifier = nn.Sequential(*list(model.classifier) + [nn.Softmax(1)])

Note however, that you cannot pass the output to nn.CrossEntropyLoss.
You could apply torch.log on it and pass it to nn.NLLLoss, which might be numerically unstable compared to F.log_softmax.

2 Likes
5

Sorry to bump this post - but why can’t the OP use CE Loss with a Softmax layer? For example, I’m trying to do the same thing here (but convert it to a probability across 1000 classes (Imagenet), but it’s not training with CE Loss.

6

nn.CrossEntropyLoss applies F.log_softmax internally, which would yield a different loss, if you already apply F.softmax manually on the logits (note that F.log_softmax would yield the same results) as seen here:

batch_size, nb_classes = 64, 10
logits = torch.randn(batch_size, nb_classes)
target = torch.randint(0, nb_classes, (batch_size,))
criterion = nn.CrossEntropyLoss()

loss = criterion(logits, target)
loss_sm = criterion(F.softmax(logits, dim=1), target)
loss_log_sm = criterion(F.log_softmax(logits, dim=1), target)

print(loss, loss_sm, loss_log_sm)
> tensor(2.7899) tensor(2.3108) tensor(2.7898)
2 Likes