moved structure-learning back to main branch

run-all.sh

@@ -1,10 +1,9 @@
 
-ITERATIONS=100 #200 #50000 # K20: 10000 #200000   ## About 1min for each 1k, for Zipf
-OUTER_BLOCKS=100 #100 #100 outer / 100 burn works really well!  
-N=50000
+ITERATIONS=250 #50000 # K20: 10000 #200000   ## About 1min for each 1k, for Zipf
+OUTER_BLOCKS=200 #20  
+N=100000 #50000 # 131072 #131072 # 131072 # 128*1024
 OUTROOT=run/
-BURN=100 #100
-ENUMERATION=9
+BURN=0
 
 # The executable
 EXEC=./gpumcmc
@@ -15,16 +14,14 @@ shopt -s globstar
 OUTROOT=./run
 
 WHICHHALF=all
-
 # for DATA in $(ls -d data-sources/Regression/1); do
+
 # DATA=data-sources/Regression/-10_20
 # for DATA in $(ls -d data-sources/Regression/*); do
-# DATA=data-sources/Science/COBE
-# DATA=data-sources/Stats/Logistic
-# DATA=data-sources/NIST/MGH09
-# for WHICHHALF in 'first-half' 'even-half' 'all' ; do
-for DATA in data-sources/Science/COBE data-sources/Science/Hubble data-sources/Science/Galileo ; do
-# for DATA in $(ls -d data-sources/NIST/*) ; do
+# DATA=data-sources/Science/BalmerSeries
+for WHICHHALF in 'first-half' 'even-half' 'all' ; do
+for DATA in $(ls -d data-sources/Science/*) ; do
+# for DATA in $(ls -d data-sources/Science/*)  $(ls -d data-sources/Stats/*) $(ls -d data-sources/NIST/*) ; do
 # for DATA in $(ls -d data-sources/Science/*)  $(ls -d data-sources/Stats/*) $(ls -d data-sources/NIST/*) ; do
 
 	echo Running $DATA 
@@ -39,9 +36,13 @@ for DATA in data-sources/Science/COBE data-sources/Science/Hubble data-sources/S
 
 	# run the CUDA MCMC; must use gnu time in order to output
 	# But here we also run time so it prints on command line too
-	time /usr/bin/time --output=$OUT/time.txt $EXEC --burn=$BURN --enumeration=$ENUMERATION --iterations=$ITERATIONS --in=$DATA/data.txt --outer=$OUTER_BLOCKS --burn=$BURN_BLOCKS --N=$N --out=$OUT --$WHICHHALF
+	time /usr/bin/time --output=$OUT/time.txt $EXEC --burn=$BURN --iterations=$ITERATIONS --in=$DATA/data.txt --outer=$OUTER_BLOCKS --burn=$BURN_BLOCKS --N=$N --out=$OUT --$WHICHHALF
 
-	sort -g -k 3 --parallel=4 $OUT/samples.txt | tail -n 10000 > $OUT/tops.txt &
+	sort -g -k 5 --parallel=4 $OUT/samples.txt | tail -n 10000 > $OUT/tops.txt &
 
+	# And a python post-processing script to do plotting.
+	# Run in the background so we can move to the next plot
+	nice -n 19 python postprocess.py --directory=$OUT &
+
+done
 done
-# done

src/__PRIMITIVES.cu

@@ -1,18 +1,17 @@
 /* GENERATED BY make_primitives_header.py in Makefile. DO NOT EDIT */
 
 
-const int NUM_OPS=12;
+const int NUM_OPS=13;
 
 
 
-enum OPS { NOOP_=0,X_=1,ZERO_=2,ONE_=3,EXP_=4,LOG_=5,NEG_=6,MUL_=7,POW_=8,ADD_=9,DIV_=10,SUB_=11};
+enum OPS { NOOP_=0,CONSTANT_=1,X_=2,ZERO_=3,ONE_=4,EXP_=5,LOG_=6,NEG_=7,MUL_=8,POW_=9,ADD_=10,DIV_=11,SUB_=12};
 
 
 
-__device__ const int NARGS[]  = {0, 0,0,0,1,1,1,2,2,2,2,2 };
- const int hNARGS[] = {0, 0,0,0,1,1,1,2,2,2,2,2 };
- const char* NAMES[] = { "<NA>", "x","zero","one","exp","log","neg","mul","pow","add","div","sub" };
+__device__ const int NARGS[]  = {0, 0,0,0,0,1,1,1,2,2,2,2,2 };
+ const int hNARGS[] = {0, 0,0,0,0,1,1,1,2,2,2,2,2 };
+ const char* NAMES[] = { "<NA>", "constant","x","zero","one","exp","log","neg","mul","pow","add","div","sub" };
 
  // Non-defined primitives, used potentially by print_program_as_expression
 
-#define CONSTANT_ -991

src/__VM_INCLUDE.cu

@@ -1,6 +1,7 @@
 /* GENERATED BY make_primitives_header.py in Makefile. DO NOT EDIT */
 
 
+	case CONSTANT_: stack[newtop] = (constant_i<MAX_CONSTANTS) * h->constants[constant_i]; constant_i += (constant_i<MAX_CONSTANTS); break;
 	case X_: stack[newtop] = X; break;
 	case ZERO_: stack[newtop] = 0.0f; break;
 	case ONE_: stack[newtop] = 1.0f; break;

src/data.cu

@@ -22,45 +22,31 @@ typedef struct datum {
 	data_t sd; // stdev of the output|input. 
 } datum;
 
-// the range of data, including the SDs, NOT counting any exclusions for even/odd, etc.
-data_t MIN_X, MAX_X, MIN_Y, MAX_Y;
-
 
 // Load data froma  file, putting it into our structs.
 // This allows us to trim our data if we want
 vector<datum>* load_data_file(const char* datapath, int FIRST_HALF_DATA, int EVEN_HALF_DATA) {
 
-	MIN_X = INFINITY;
-	MIN_Y = INFINITY;
-	MAX_X = -INFINITY;
-	MAX_Y = -INFINITY;
-
 	FILE* fp = fopen(datapath, "r");
 	if(fp==NULL) { cerr << "*** ERROR: Cannot open file:\t" << datapath <<"\n"; exit(1);}
 
 	vector<datum>* d = new vector<datum>();
-	char* line = NULL; size_t len=0; 
+	char* line = NULL; size_t len=0; float x,y,sd; 
 	while( getline(&line, &len, fp) != -1) {
-
-		if( line[0] == '#' || strspn(line, " \r\n\t") == strlen(line)) continue;  // skip comments and whitespace
-
-		float x,y,sd; // these are set by sscanf
-		if (sscanf(line, "%f\t%f\t%f\n", &x, &y, &sd) == 3) { } // floats
-		else if (sscanf(line, "%e\t%e\t%e\n", &x, &y, &sd) == 3) { } // scientific notation
+		if( line[0] == '#' ) continue;  // skip comments
+		else if (sscanf(line, "%f\t%f\t%f\n", &x, &y, &sd) == 3) { // floats
+			d->push_back( (datum){.input=(data_t)x, .output=(data_t)y, .sd=(data_t)sd} );
+		}
+		else if (sscanf(line, "%e\t%e\t%e\n", &x, &y, &sd) == 3) { // scientific notation
+			d->push_back( (datum){.input=(data_t)x, .output=(data_t)y, .sd=(data_t)sd} );
+		}
+		else if ( strspn(line, " \r\n\t") == strlen(line) ) { // skip whitespace
+			continue;
+		}
 		else {
 			cerr << "*** ERROR IN PARSING INPUT\t" << line << endl;
 			exit(1);
 		}
-
-		// add these x,y,sd
-		d->push_back( (datum){.input=(data_t)x, .output=(data_t)y, .sd=(data_t)sd} );
-
-		// our min/max includes the SDs
-		if(x > MAX_X) MAX_X = x;
-		if(x < MIN_X) MIN_X = x;
-
-		if(y+sd > MAX_Y) MAX_Y = y+sd;
-		if(y-sd < MIN_Y) MIN_Y = y-sd;
 	}
 	fclose(fp);