X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=expr.py;h=d07d09fde28d3aca60a43443da73844cf9172917;hb=1406521aecaab783d2ea267b0f973ac17e091bf7;hp=baee50275248c8c4a4f234ba8dc1f8f96eabe48b;hpb=495c959114942d07808788e27d9fcaa951a7d21e;p=picoclvr.git

diff --git a/expr.py b/expr.py
index baee502..d07d09f 100755
--- a/expr.py
+++ b/expr.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python
 
-import math
+import math, re
 
 import torch, torchvision
 
@@ -24,7 +24,7 @@ def random_expr(variables, budget):
         else:
             return str(torch.randint(10, (1,)).item())
     else:
-        op = torch.randint(4, (1,)).item()
+        op = torch.randint(3, (1,)).item()
         if op == 0:
             e = random_expr(variables, budget - 2)
             if ("+" in e or "-" in e or "*" in e) and (e[0] != "(" or e[-1] != ")"):
@@ -38,29 +38,42 @@ def random_expr(variables, budget):
             if op == 1:
                 return e1 + "+" + e2
             elif op == 2:
-                return e1 + "+" + e2
-            elif op == 3:
                 return e1 + "*" + e2
 
 
 def generate_program(nb_variables, length):
     s = ""
     variables = set()
+
+    # We take length itself half of the time, and uniform between 1
+    # and length otherwise. The actual length can be slightly greater
+
+    length = min(length, 1+torch.randint(length*2, (1,)).item())
     while len(s) < length:
         v = random_var(nb_variables=nb_variables)
         s += v + "=" + random_expr(variables, budget=20) + ";"
         variables.add(v)
+
     return s, variables
 
 
-def generate_sequences(nb, nb_variables=5, length=20, randomize_length=False):
+def extract_results(seq):
+    f = lambda a: (a[0], -1 if a[1] == "" else int(a[1]))
+    results = [
+        dict([f(tuple(x.split(":"))) for x in re.findall("[A-Z]:[0-9]*", s)])
+        for s in seq
+    ]
+    return results
+
+
+def generate_sequences(nb, nb_variables=5, length=20):
+    assert nb_variables <= 26
     sequences = []
+    result_max=99
     for n in range(nb):
         result = None
-        while result == None or max(result.values()) > 100:
+        while result == None or max(result.values()) > result_max:
             l = length
-            if l > 5 and randomize_length:
-                l = 5 + torch.randint(l-5, (1,)).item()
             p, v = generate_program(nb_variables, l)
             v = ", ".join(['"' + v + '": ' + v for v in v])
             ldict = {}
@@ -69,7 +82,7 @@ def generate_sequences(nb, nb_variables=5, length=20, randomize_length=False):
 
         k = list(result.keys())
         k.sort()
-        sequences.append(p + " " + ";".join([v + ":" + str(result[v]) for v in k]))
+        sequences.append(p + " " + "".join([v + ":" + str(result[v]) + ";" for v in k]))
 
     return sequences
 
@@ -78,8 +91,10 @@ if __name__ == "__main__":
     import time
 
     start_time = time.perf_counter()
-    sequences = generate_sequences(1000, randomize_length=True)
+    sequences = generate_sequences(1000, length=40)
     end_time = time.perf_counter()
     for s in sequences[:10]:
         print(s)
     print(f"{len(sequences) / (end_time - start_time):.02f} samples per second")
+
+    print(extract_results(sequences[:10]))