experiment.txt

100 plane images: 25% pixel sampling 

Dataset-1: Smooth Gaussian No Noise
Dataset-2: Noise ~ guassian var = 1 * length^2/5
	- implicitly two controls here if we remove bounding box features

Dataset 3- ditto 1, hard feature
Dataset 4- ditto 2, hard features

Dataset 5, mask feature 
Dateset 6, mask feature 

cluster 3: mask, hard, 
cluster 4: smooth, control

Commands

control <-- used noise-smooth-train.mat

train 75 units 85 units

save in matlab_segmentation_models


diary('/h/53/xukelvin/matlab_segmentation/models/may8/control/85.txt');
model = mlp2layer_train(tdata{1}(:,[1:107]),tlabels, 85, 0.5, 0.9, 0, 4000, 20, vdata{1}(:,[1:107]), vlabels, 200, '/h/53/xukelvin/matlab_segmentation/models/may8/control'); diary off;

-------------

generate a test set for every case:

Choose best model for each case and run tuning code in parallel. 

-> IOU and pixel accuracy. 

1: smooth
2: smooth noise. 
3: hard
4: hard noise
5: mask
6: mask noise

after matlab reshape it will all be in order

Best models

1: smooth               800.mat
2: smooth noise.        600.mat
3: hard                 1600.mat
4: hard noise           200.mat
5: mask                 200.mat
6: mask noise           1000.mat
7. control              m4000.

Rerunning experiment standardizing the set

diary('/u/xukelvin/matlab_segmentation/models/apr7/smooth/85.txt'); 


Rerunning experiments <--- keeping track of the mean and standard deviation of the training set 

Best Models:

Round 2:

1: smooth   m800.mat
2: smooth noise  m400.mat
3: hard         m1200.mat 
4: hard noise   m1400.mat
5: mask         m3600.mat
6: mask noise   m1400.mat