caffe: probability distribution for regression / expanding classification (softmax layer) to allow 3D output -
i have working network pixel-wise image segmentation, i.e. 2 images input 1 data 1 label (ground_truth). therefore, using softmaxwithloss
layer shown below:
layer { name: "conv" type: "convolution" bottom: "bottom" top: "conv" convolution_param { num_output: 256 # <-- "256 classes" ... } } layer { name: "loss" type: "softmaxwithloss" bottom: "data" bottom: "label" top: "loss" }
my input image values range [0-255], why have 256 classes. want transform segmentation / classification task regression task. assumed things have change loss layer
, num_output
in convolution layer
this:
layer { name: "conv" type: "convolution" bottom: "bottom" top: "conv" convolution_param { num_output: 1 # <-- "regression" ... } } layer { name: "loss" type: "euclideanloss" bottom: "data" bottom: "label" top: "loss" }
my regression task not produce satisfying results @ all.
i think problem regression task compared classification task classification task have distribution on depth values --> num_output = 256
--> in end have result of 256 x 128 x x128
means 1 probability each depth class
. regression task output num_output = 1
returns one depth value
per pixel rather 256
classification task.
- classification: know how adjust
softmax layer
in order produce results like:256 x a_number_greater_1 x width x depth
rather256 x width x depth
.
or
- regression: there other approach achieve results distribution of probabilities
softmax layer
does?
Comments
Post a Comment