#!/usr/bin/env python
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
import math
import torch, torchvision
######################################################################
-def generate(nb_values=3, max_input=9, prog_len=6, nb_runs=5):
- prog_len = 1 + torch.randint(prog_len - 1, (1,)).item()
- prog = [rpl_ops[k] for k in torch.randint(len(rpl_ops), (prog_len,))]
+def generate(
+ nb_starting_values=3, nb_result_values_max=None, max_input=9, prog_len=6, nb_runs=5
+):
+ prog_len = (1 + torch.randint(2 * prog_len, (1,))).clamp(max=prog_len).item()
+
+ while True:
+ no_empty_stack = True
+ prog = [rpl_ops[k] for k in torch.randint(len(rpl_ops), (prog_len,))]
+
+ result = []
+ for _ in range(nb_runs):
+ stack = [
+ x.item() for x in torch.randint(max_input + 1, (nb_starting_values,))
+ ]
+ result_stack = rpl_exec(prog, stack)
+ if len(result_stack) == 0:
+ no_empty_stack = False
+ result = result + ["<input>"] + stack + ["<output>"] + result_stack
- result = []
- for _ in range(nb_runs):
- stack = [x.item() for x in torch.randint(max_input + 1, (nb_values,))]
- result_stack = rpl_exec(prog, stack)
- result = result + ["<input>"] + stack + ["<output>"] + result_stack
+ result = result + ["<prog>"] + prog
+ result = result + ["<end>"]
+
+ if no_empty_stack and (
+ nb_result_values_max is None or len(result_stack) <= nb_result_values_max
+ ):
+ break
- result = result + ["<prog>"] + prog
- result = result + ["<end>"]
return result
k = 0
while seq[k] == "<input>":
o = next_marker(seq, ["<output>"], start=k + 1)
+ if o is None:
+ raise ValueError("Missing output markers (should be correct in the prompt)")
e = next_marker(seq, ["<input>", "<prog>"], start=o)
- if o is None or e is None:
- raise ValueError("Invalid input/output")
+ if e is None:
+ raise ValueError(
+ "Missing input/output markers (should be correct in the prompt)"
+ )
try:
io.append(
([int(x) for x in seq[k + 1 : o]], [int(x) for x in seq[o + 1 : e]])
)
except ValueError:
- raise ValueError("Invalid input/output")
+ raise ValueError(
+ "Invalid input/output value (should be correct in the prompt)"
+ )
k = e
prog = []
else:
prog = seq[k + 1 : e]
+ else:
+ raise ValueError("Missing <prog> (it should be in the prompt)")
+
return prog, io
+def stack_distance(target_stack, result_stack):
+ return abs(len(result_stack) - len(target_stack)) + sum(
+ [0 if x == y else 1 for x, y in zip(result_stack, target_stack)]
+ )
+
+
def compute_nb_errors(seq):
prog, io = decompose(seq)
if len(set(prog) - set(rpl_ops)) > 0:
# Program is not valid, we count 100% error
for start_stack, target_stack in io:
- stacks.append((start_stack, target_stack, "N/A", False))
+ stacks.append((start_stack, target_stack, ["N/A"], False))
nb_total += len(target_stack)
nb_errors += len(target_stack)
for start_stack, target_stack in io:
result_stack = rpl_exec(prog, start_stack)
nb_total += len(target_stack)
- e = abs(len(result_stack) - len(target_stack)) + sum(
- [0 if x == y else 1 for x, y in zip(result_stack, target_stack)]
- )
+ e = stack_distance(target_stack, result_stack)
nb_errors += e
stacks.append((start_stack, target_stack, result_stack, e == 0))