Some problem about fcrn-depthprediction HOT 1 CLOSED

Hi,

The ResNet-50 architecture is slightly modified, only the last layers have to be removed (i.e. average pooling, fully-connected layer and softmax), before plugging in the upsampling part.

net = load(imagenet-resnet-50-dag.mat');
net = dagnn.DagNN.loadobj(net) ;
net.removeLayer('pool5');   
net.removeLayer('prob');   
net.removeLayer('fc1000');

But please note that the architecture in NYU_ResNet-UpProj.mat already includes these changes.
Similarly for AlexNet or VGG-16 one would have to remove all fully connected layers.

from fcrn-depthprediction.