rejection sampling to ensure identifiability

test_synthetic.py

@@ -203,7 +203,7 @@ def test_accuracy():
     size = 1000
     level = 0.01
     suppfX = range(-7, 8)
-    srcsX = ["uniform", "binomial", "negativeBinomial",
+    srcsX = ["negativeBinomial", "uniform", "binomial", "negativeBinomial",
              "geometric", "hypergeometric", "poisson", "multinomial"]
     print "-" * 70
     print "%18s%10s%10s%10s%10s%10s" % ("TYPE_X", "DC", "DR", "CISC", "CRISPE", "CRISP")
@@ -489,20 +489,35 @@ def test_significance():
 
 def test_hypercompression():
     m = 100
-    size = 100
+    size = 500
     alpha = 0.01
     suppfX = range(-7, 8)
     srcX = "geometric"
 
     fp = open("results/no-hypercompression.dat", "w")
     diffs = []
     decisions = []  # 1=correct, -1=incorrect, 0=wrong
-    for i in xrange(m):
+    nsamples = 0
+    while nsamples < m:
         X = generate_X(srcX, size)
         suppX = list(set(X))
         f = map_randomly(suppX, suppfX)
         N = generate_additive_N(size)
         Y = [f[X[i]] + N[i] for i in xrange(size)]
+
+        # check if f(x) + supp N are disjoint for x in domx
+        suppN = set(N)
+        decomps = []
+        for x in suppX:
+            fx = f[x]
+            sum_fx_suppN = set([fx + n for n in suppN])
+            decomps.append(sum_fx_suppN)
+
+        non_overlapping_noise = are_disjoint(decomps)
+        if non_overlapping_noise:
+            continue
+
+        nsamples += 1
         crisp_score = crisp(X, Y)
         diff = abs(crisp_score[0] - crisp_score[1])
 
@@ -562,7 +577,7 @@ def test_hypercompression():
 def test_sample_size():
     nsim = 1000
     level = 0.05
-    sizes = [100, 400, 700, 1000, 1300, 1600, 2000]
+    sizes = [200, 100, 400, 700, 1000, 1300, 1600, 2000]
     suppfX = range(-7, 8)
     srcX = "geometric"
 
@@ -621,8 +636,9 @@ def test_sample_size():
     fp.close()
 
 
-def test_accuracy_pvalue():
+def test_accuracy_alpha():
     print "testing various critical thresholds for DR"
+    sys.stdout.flus()
     nsim = 1000
     size = 1000
     level = 0.01
@@ -658,9 +674,11 @@ def test_accuracy_pvalue():
         assert nsamples == nsim
         acc = ncorrect * 100 / nsim
         print level, acc
+        sys.stdout.flus()
 
 
 if __name__ == "__main__":
-    # test_sample_size()
+    test_sample_size()
     # test_accuracy()
-    test_accuracy_pvalue()
+    # test_accuracy_alpha()
+    # test_hypercompression()