Sir,
I want to confirm like If I want to apply LeNet or AlexNet in digit recognition then what will be the output dimensions of the two DENSE LAYERS AT THE END??
WILL IT BE 101 AND 101 FOR BOTH or 120 and 84 is fixed that we have the take that no. of outputs for whatever dataset it is??
The output will be 84 only but you need to reduce the size and output to be 10 then you need to add some more max pooling layer and FC layers so that the size will be 10x1.
By the way VGG net is trained in Imagenet datasets, so the inputs and output will be different.
But you can use the same architecture of it and reduce the size to 10x1 do the transfer learning by training with MNIST in between as get the output.