forked from pdollar/edges
-
Notifications
You must be signed in to change notification settings - Fork 0
/
edgesSweeps.m
211 lines (198 loc) · 8.62 KB
/
edgesSweeps.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
function edgesSweeps()
% Parameter sweeps for structured edge detector.
%
% Running the parameter sweeps requires altering internal flags.
% The sweeps are not well documented, use at your own discretion.
%
% Structured Edge Detection Toolbox Version 3.01
% Code written by Piotr Dollar, 2014.
% Licensed under the MSR-LA Full Rights License [see license.txt]
% select type and location of cluster (see fevalDistr.m)
rtDir = 'D:\code\research\edges\';
pDistrTrn={'type','local'}; pDistrTst=pDistrTrn;
% define parameter sweeps
expNms= {'MD-imWidth','MD-gtWidth','TR-nData','TR-nPos','TR-nImgs',...
'TR-sharpen','TR-nTrees','TR-split','TR-minChild','TR-maxDepth',...
'TR-fracFtrs','TR-nSamples','TR-nClasses','TR-discretize',...
'FT-nCells','FT-nOrients','FT-normRad','FT-shrink',...
'FT-grdSmooth','FT-chnSmooth','FT-simSmooth','final'};
expNms=expNms(1:end); T=5; full=3; dataset='BSDS';
[opts,lgd,lbl]=createExp(rtDir,expNms,dataset,full);
% run training and testing jobs
[jobsTrn,jobsTst] = createJobs(rtDir,opts,dataset,T); N=length(expNms);
fprintf('nTrain = %i; nTest = %i\n',length(jobsTrn),length(jobsTst));
tic, s=fevalDistr('edgesTrain',jobsTrn,pDistrTrn); assert(s==1); toc
tic, s=fevalDistr('edgesEval',jobsTst,pDistrTst); assert(s==1); toc
% plot results
for e=1:N, plotExps(expNms{e},opts{e},lgd{e},lbl{e},T,1); end
end
function plotExps( expNm, opts, lgd, lbl, T, type )
% get all results and display error
disp([expNm ' [' lbl ']']); N=length(lgd);
res=zeros(N,T); mNms=cell(1,N);
for e=1:N, mNms{e}=[opts(e).modelDir 'val/' opts(e).modelFnm]; end
for e=1:N, for t=1:T, r=dlmread([mNms{e} 'T' int2str2(t,2) ...
'-eval/eval_bdry.txt']); r=r([4 7 8]); res(e,t)=r(type); end; end
stds=std(res,0,2)*100; R=mean(res,2)*100; msg=' %.2f +/- %.2f [%s]\n';
for e=1:N, fprintf(msg,R(e),stds(e),lgd{e}); end
if(0), disp(res); disp(max(res,[],2)); end
types={'ODS','OIS','AP'}; type=types{type};
% plot sweeps (two cases for format of x labels)
figPrp = {'Units','Pixels','Position',[800 600 640 220]};
figure(1); clf; set(1,figPrp{:}); set(gca,'FontSize',24); clr=[0 .69 .94];
pPl1={'LineWidth',3,'MarkerSize',15,'Color',clr,'MarkerFaceColor',clr};
pPl2=pPl1; clr=[1 .75 0]; pPl2{6}=clr; pPl2{8}=clr; d=0;
for e=1:N, if(lgd{e}(end)=='*'), d=e; end; end; if(d), lgd{d}(end)=[]; end
plot(R,'-d',pPl1{:}); hold on; if(d),plot(d,R(d),'d',pPl2{:}); end; e=.001;
ylabel([type ' \times 100']); axis([.5 N+.5 min([R; 66])-e max([R; 72])+e])
if(isempty(lbl)), imLabel(lgd,'bottom',30,{'FontSize',24}); lgd=[]; end
if(0); xlabel(lbl); end; set(gca,'XTick',1:1:N,'XTickLabel',lgd(1:1:N));
% save plot
plDir=[opts(1).modelDir 'plots' type '/']; fFig=[plDir expNm];
if(~exist(plDir,'dir')), mkdir(plDir); end %#ok<*CTCH>
for t=1:25, try savefig(fFig,1,'png'); break; catch, pause(1), end; end
end
function [jobsTrn,jobsTst] = createJobs( rtDir, opts, dataset, T )
% Prepare all jobs (one train and one test job per set of opts).
opts=[opts{:}]; N=length(opts); NT=N*T;
opts=repmat(opts,1,T); nms=cell(1,NT);
jobsTrn=cell(1,NT); doneTrn=zeros(1,NT);
jobsTst=cell(1,NT); doneTst=zeros(1,NT);
thrs = logspace(-2,log10(.99),25)'; thrs(1)=1e-5;
if( strcmpi(dataset,'bsds') )
pTest={'dataType','val', 'thrs',thrs, 'cleanup',1,...
'opts',{'modelDir',[rtDir '/sweepsBSDS/'],...
'bsdsDir',[rtDir '/BSR/BSDS500/data/']} };
else
pTest={'dataType','val', 'thrs',thrs, 'cleanup',1, 'maxDist',.011,...
'opts',{'modelDir',[rtDir '/sweepsNYUD/'],...
'bsdsDir',[rtDir '/BSR/NYUD/data/']} };
end
for e=1:NT
t=ceil(e/N); opts(e).seed=(t-1)*100000+1;
nm=[opts(e).modelFnm 'T' int2str2(t,2)]; opts(e).modelFnm=nm;
mFnm=[opts(e).modelDir 'forest/' nm '.mat']; nms{e}=nm;
eFnm=[opts(e).modelDir 'val/' nm '-eval/eval_bdry.txt'];
doneTrn(e)=exist(mFnm,'file')==2; jobsTrn{e}={opts(e)};
doneTst(e)=exist(eFnm,'file')==2; jobsTst{e}=[mFnm pTest];
end
[~,kp]=unique(nms,'stable');
doneTrn=doneTrn(kp); jobsTrn=jobsTrn(kp); jobsTrn=jobsTrn(~doneTrn);
doneTst=doneTst(kp); jobsTst=jobsTst(kp); jobsTst=jobsTst(~doneTst);
end
function [opts,lgd,lbl] = createExp( rtDir, expNm, dataset, full )
% if expNm is a cell, call recursively and return
if( iscell(expNm) )
N=length(expNm); opts=cell(1,N); lgd=opts; lbl=opts;
for e=1:N, [opts{e},lgd{e},lbl{e}]=...
createExp(rtDir,expNm{e},dataset,full); end; return;
end
% default params for edgesTrain.m
opts=edgesTrain(); opts.nThreads=1; nData=2e5; nPos=.5;
opts.nPos=round(nData*nPos); opts.nNeg=round(nData*(1-nPos));
assert(any(strcmpi(dataset,{'bsds','nyud'})));
if( strcmpi(dataset,'bsds') )
opts.modelDir=[rtDir '/sweepsBSDS/']; opts.nImgs=200;
opts.bsdsDir=[rtDir '/BSR/BSDS500/data/'];
else
opts.modelDir=[rtDir '/sweepsNYUD/']; opts.nImgs=381;
opts.bsdsDir=[rtDir '/BSR/NYUD/data/'];
opts.rgbd=2; opts.fracFtrs=1/8;
end
% setup opts
optsDefault=opts; N=100; lgd=cell(1,N); ss=lgd;
opts=opts(ones(1,N)); hasDefault=1;
switch expNm
case 'MD-imWidth'
lbl='window size for x'; vs=[1:4 6 8]*8; N=length(vs);
for e=1:N, opts(e).imWidth=vs(e); end
for e=1:N, opts(e).gtWidth=min(vs(e),opts(e).gtWidth); end
case 'MD-gtWidth'
lbl='window size for y'; vs=[1:4 6 8]*4; N=length(vs);
for e=1:N, opts(e).gtWidth=vs(e); end
case 'TR-nData'
lbl='# train patches x10^4'; vs=[1 2 5 10 20 50 100 200]; N=length(vs);
for e=1:N, opts(e).nPos=round(vs(e)*1e4*nPos); end
for e=1:N, opts(e).nNeg=round(vs(e)*1e4*(1-nPos)); end
if(full<2), N=5; elseif(full<3), N=6; end
case 'TR-nPos'
lbl = 'fraction positive data'; vs=20:10:80; N=length(vs);
for e=1:N, opts(e).nPos=round(nData*vs(e)/100); end
for e=1:N, opts(e).nNeg=round(nData*(1-vs(e)/100)); end
for e=1:N, lgd{e}=sprintf('.%i',vs(e)/10); end
case 'TR-nImgs'
lbl='# train images'; vs=[10 20 50 100 opts(1).nImgs]; N=length(vs);
for e=1:N, opts(e).nImgs=vs(e); end
case 'TR-sharpen'
lbl='sharpening radius'; vs=0:4; N=length(vs);
for e=1:N, opts(e).sharpen=vs(e); end
case 'TR-nTrees'
lbl='# decision trees'; vs=2.^(0:4); N=length(vs);
for e=1:N, opts(e).nTrees=vs(e); end;
for e=1:N, opts(e).nTreesEval=max(1,vs(e)/2); end
if(full<2), N=4; end
case 'TR-split'
lbl='information gain';
ss={'gini','entropy','twoing'}; N=length(ss); lgd=ss;
for e=1:N, opts(e).split=ss{e}; end
case 'TR-minChild'
lbl='min samples per node'; vs=2:2:16; N=length(vs);
for e=1:N, opts(e).minChild=vs(e); end
case 'TR-maxDepth'
lbl='max tree depth'; vs=2.^(2:6); N=length(vs);
for e=1:N, opts(e).maxDepth=vs(e); end
case 'TR-fracFtrs'
lbl='fraction features'; vs=2.^(2:6); N=length(vs);
for e=1:N, opts(e).fracFtrs=1/vs(e); end
for e=1:N, lgd{e}=sprintf('1/%i',vs(e)); end
case 'TR-nSamples'
lbl='m (size of Z)'; vs=2.^(0:2:8); N=length(vs);
for e=1:N, opts(e).nSamples=vs(e); end
case 'TR-nClasses'
lbl='k (size of C)'; vs=2.^(1:5); N=length(vs);
for e=1:N, opts(e).nClasses=vs(e); end
case 'TR-discretize'
lbl='discretization type';
ss={'pca','kmeans'}; N=length(ss); lgd=ss;
for e=1:N, opts(e).discretize=ss{e}; end
case 'FT-nCells'
lbl='# grid cells'; vs=1:2:7; N=length(vs);
for e=1:N, opts(e).nCells=vs(e); end
if(full<2), N=3; end
case 'FT-nOrients'
lbl='# gradient orients'; vs=0:2:8; N=length(vs);
for e=1:N, opts(e).nOrients=vs(e); end
case 'FT-normRad'
lbl='normalization radius'; vs=[0 2.^(0:3)]; N=length(vs);
for e=1:N, opts(e).normRad=vs(e); end
case 'FT-shrink'
lbl='channel downsample'; vs=2.^(0:2); N=length(vs);
for e=1:N, opts(e).shrink=vs(e); end
case 'FT-grdSmooth'
lbl = 'gradient blur'; vs=[0 2.^(0:4)]; N=length(vs);
for e=1:N, opts(e).grdSmooth=vs(e); end
case 'FT-chnSmooth'
lbl='channel blur'; vs=[0 2.^(0:4)]; N=length(vs);
for e=1:N, opts(e).chnSmooth=vs(e); end
case 'FT-simSmooth'
lbl='self-similarity blur'; vs=[0 2.^(0:4)]; N=length(vs);
for e=1:N, opts(e).simSmooth=vs(e); end
case 'final'
if(strcmpi(dataset,'bsds')), vs=[0 2]; else
vs=0:3; rgbd=[2 2 1 0]; for e=2:4, opts(e).rgbd=rgbd(e); end; end
lbl='final'; N=length(vs); nData=2e6;
for e=2:N, opts(e).nPos=round(nData*nPos); end
for e=2:N, opts(e).nNeg=round(nData*(1-nPos)); end
if(full<3), N=1; end
otherwise, error('invalid exp: %s',expNm);
end
% produce final set of opts and find default opts
for e=1:N, if(isempty(lgd{e})), lgd{e}=int2str(vs(e)); end; end
for e=1:N, if(isempty(ss{e})), ss{e}=int2str2(vs(e),5); end; end
O=1:N; opts=opts(O); lgd=lgd(O); ss=ss(O); d=0;
for e=1:N, if(isequal(optsDefault,opts(e))), d=e; break; end; end
if(hasDefault && d==0), disp(expNm); assert(false); end
for e=1:N, opts(e).modelFnm=[expNm ss{e}]; end
if(hasDefault), lgd{d}=[lgd{d} '*']; opts(d).modelFnm='Default'; end
if(0), disp([ss' lgd']'); end
end