Update

examples/classification/SequenceCountInterpretabilityDemo.py

@@ -25,7 +25,7 @@ def main(args):
     experiment_results = metrics(args)
 
     X_train = np.zeros((args.examples, 1, args.sequence_length, 2), dtype=np.uint32)
-    Y_train = np.random.randint(0, 3, size=(args.examples), dtype=np.uint32)
+    Y_train = np.random.randint(0, 4, size=(args.examples), dtype=np.uint32)
     for i in range(args.examples):
         position_1 = np.random.randint(0, args.sequence_length-4)
         position_2 = position_1+1
@@ -67,7 +67,7 @@ def main(args):
 
             X_train[i,0,position_5,0] = 0
             X_train[i,0,position_5,1] = 0
-        else:
+        elif Y_train[i] == 2:
             X_train[i,0,position_1,0] = 1
             X_train[i,0,position_1,1] = 0
 
@@ -82,13 +82,27 @@ def main(args):
 
             X_train[i,0,position_5,0] = 0
             X_train[i,0,position_5,1] = 0
+        else:
+            X_train[i,0,position_1,0] = 1
+            X_train[i,0,position_1,1] = 0
 
+            X_train[i,0,position_2,0] = 0
+            X_train[i,0,position_2,1] = 0
+
+            X_train[i,0,position_3,0] = 0
+            X_train[i,0,position_3,1] = 0
+
+            X_train[i,0,position_4,0] = 0
+            X_train[i,0,position_4,1] = 0
+
+            X_train[i,0,position_5,0] = 0
+            X_train[i,0,position_5,1] = 0
 
         if np.random.rand() <= args.noise:
-            Y_train[i] = np.random.choice(np.setdiff1d([0,1,2], [Y_train[i]]))
+            Y_train[i] = np.random.choice(np.setdiff1d([0,1,2,3], [Y_train[i]]))
 
     X_test = np.zeros((args.examples//10, 1, args.sequence_length, 2), dtype=np.uint32)
-    Y_test = np.random.randint(0, 3, size=(args.examples//10), dtype=np.uint32)
+    Y_test = np.random.randint(0, 4, size=(args.examples//10), dtype=np.uint32)
     for i in range(args.examples//10):
         position_1 = np.random.randint(0, args.sequence_length-4)
         position_2 = position_1+1
@@ -128,14 +142,29 @@ def main(args):
             X_test[i,0,position_4,0] = 0
             X_test[i,0,position_4,1] = 0
 
+            X_test[i,0,position_5,0] = 0
+            X_test[i,0,position_5,1] = 0
+        elif Y_test[i] == 2:
+            X_test[i,0,position_1,0] = 1
+            X_test[i,0,position_1,1] = 0
+
+            X_test[i,0,position_2,0] = 1
+            X_test[i,0,position_2,1] = 0
+
+            X_test[i,0,position_3,0] = 0
+            X_test[i,0,position_3,1] = 0
+
+            X_test[i,0,position_4,0] = 0
+            X_test[i,0,position_4,1] = 0
+
             X_test[i,0,position_5,0] = 0
             X_test[i,0,position_5,1] = 0
         else:
             X_test[i,0,position_1,0] = 1
             X_test[i,0,position_1,1] = 0
 
             X_test[i,0,position_2,0] = 0
-            X_test[i,0,position_2,1] = 1
+            X_test[i,0,position_2,1] = 0
 
             X_test[i,0,position_3,0] = 0
             X_test[i,0,position_3,1] = 0