5 import torch, torchvision
8 from torch.nn import functional as F
10 ######################################################################
13 def rpl_exec(program, stack):
18 a, b = stack.pop(), stack.pop()
22 a, b = stack.pop(), stack.pop()
23 stack.append(min(a, b))
26 a, b = stack.pop(), stack.pop()
27 stack.append(max(a, b))
30 a, b = stack.pop(), stack.pop()
35 a, b = stack.pop(), stack.pop()
46 raise ValueError(f"Unknown instruction {op}")
51 rpl_ops = ["add", "min", "max", "swp", "rep", "dup", "del"]
53 ######################################################################
56 def generate(nb_starting_values=3, max_input=9, prog_len=6, nb_runs=5):
57 prog_len = (1 + torch.randint(2 * prog_len, (1,))).clamp(max=prog_len).item()
61 prog = [rpl_ops[k] for k in torch.randint(len(rpl_ops), (prog_len,))]
64 for _ in range(nb_runs):
66 x.item() for x in torch.randint(max_input + 1, (nb_starting_values,))
68 result_stack = rpl_exec(prog, stack)
69 if len(result_stack) == 0:
70 no_empty_stack = False
71 result = result + ["<input>"] + stack + ["<output>"] + result_stack
73 result = result + ["<prog>"] + prog
74 result = result + ["<end>"]
81 def next_marker(seq, tokens, start=0):
85 i = seq.index(t, start)
86 if pos is None or i < pos:
96 while seq[k] == "<input>":
97 o = next_marker(seq, ["<output>"], start=k + 1)
98 e = next_marker(seq, ["<input>", "<prog>"], start=o)
99 if o is None or e is None:
100 raise ValueError("Invalid input/output")
103 ([int(x) for x in seq[k + 1 : o]], [int(x) for x in seq[o + 1 : e]])
106 raise ValueError("Invalid input/output")
110 if seq[k] == "<prog>":
111 e = next_marker(seq, ["<end>"], start=k)
115 prog = seq[k + 1 : e]
119 def compute_nb_errors(seq):
120 prog, io = decompose(seq)
122 nb_total, nb_errors = 0, 0
126 if len(set(prog) - set(rpl_ops)) > 0:
127 # Program is not valid, we count 100% error
128 for start_stack, target_stack in io:
129 stacks.append((start_stack, target_stack, ["N/A"], False))
130 nb_total += len(target_stack)
131 nb_errors += len(target_stack)
135 for start_stack, target_stack in io:
136 result_stack = rpl_exec(prog, start_stack)
137 nb_total += len(target_stack)
138 e = abs(len(result_stack) - len(target_stack)) + sum(
139 [0 if x == y else 1 for x, y in zip(result_stack, target_stack)]
142 stacks.append((start_stack, target_stack, result_stack, e == 0))
144 return nb_total, nb_errors, prog, stacks
147 ######################################################################
149 if __name__ == "__main__":
154 print(compute_nb_errors(seq))