conv neural network - Does the dropout layer need to be defined in deploy.prototxt in caffe? -
in alexnet implementation in caffe, saw following layer in deploy.prototxt file:
layer { name: "drop7" type: "dropout" bottom: "fc7" top: "fc7" dropout_param { dropout_ratio: 0.5 } }
now key idea of dropout randomly drop units (along connections) neural network during training.
does mean can delete layer deploy.prototxt, file meant used during testing only?
yes. dropout not required during testing.
even if include dropout layer, nothing special happens during testing. see source code of dropout forward pass:
if (this->phase_ == train) { // code } else { caffe_copy(bottom[0]->count(), bottom_data, top_data); //code copy bottom blob top blob }
as seen in source code, bottom blob data copied top blob data memory if not on training phase.
Comments
Post a Comment