def count_lengths(args: argparse.Namespace, filename: str):
print(f"Counting {filename}")
full_filename = args.prelude + "/" + filename
scraped_commands = list(
read_all_text_data(Path2(full_filename + ".scrape")))
scraped_iter = iter(scraped_commands)
if args.post_linear:
original_commands = serapi_instance.load_commands_preserve(
args, 0, full_filename + ".lin")
else:
original_commands = serapi_instance.load_commands_preserve(
args, 0, full_filename)
with open(full_filename + ".csv", 'w') as fout:
rowwriter = csv.writer(fout)
lemma_statement = ""
in_proof = False
cur_len = 0
for cmd in original_commands:
if not serapi_instance.possibly_starting_proof(
cmd) and not in_proof:
continue
if serapi_instance.possibly_starting_proof(cmd) and not in_proof:
normalized_command = norm(cmd)
cur_scraped = norm(next(scraped_iter))
while cur_scraped != normalized_command:
cur_scraped = norm(next(scraped_iter))
try:
next_after_start = next(scraped_iter)
except StopIteration:
next_after_start = ""
if isinstance(next_after_start, ScrapedTactic):
lemma_statement = norm(cmd)
in_proof = True
cur_len = 0
else:
scraped_iter = itertools.chain([next_after_start],
scraped_iter)
elif serapi_instance.ending_proof(cmd):
assert in_proof
rowwriter.writerow([lemma_statement.strip(), cur_len])
cur_len = -1
in_proof = False
elif in_proof:
assert cur_len >= 0
if re.match("[{}]|[*-+]*$", norm(cmd)):
continue
if re.match("Proof\.", norm(cmd)):
continue
cur_len += 1
if args.add_semis or args.post_linear:
cur_len += count_outside_matching("\{\|", "\|\}", ";",
norm(cmd))
return full_filename + ".csv"
def main():
parser = argparse.ArgumentParser(description="linearize a set of files")
parser.add_argument('--prelude', default=".")
parser.add_argument('--hardfail',
default=False,
const=True,
action='store_const')
parser.add_argument('-v', '--verbose', action='count', default=0)
parser.add_argument('--skip-nochange-tac',
default=False,
const=True,
action='store_const',
dest='skip_nochange_tac')
parser.add_argument("--progress",
action='store_const',
const=True,
default=False)
parser.add_argument('filenames', nargs="+", help="proof file name (*.v)")
arg_values = parser.parse_args()
base = os.path.dirname(os.path.abspath(__file__)) + "/.."
coqargs = ["sertop", "--implicit"]
for filename in arg_values.filenames:
if arg_values.verbose:
eprint("Linearizing {}".format(filename))
local_filename = arg_values.prelude + "/" + filename
original_commands = serapi_instance.load_commands_preserve(
arg_values, 0, arg_values.prelude + "/" + filename)
fresh_commands = preprocess_file_commands(arg_values, 0,
original_commands, coqargs,
arg_values.prelude,
local_filename, filename,
False)
serapi_instance.save_lin(fresh_commands, local_filename)
def get_file_commands(args : argparse.Namespace, file_idx : int,
filename : str) -> List[str]:
local_filename = args.prelude + "/" + filename
loaded_commands = try_load_lin(args, file_idx, local_filename)
if loaded_commands is None:
print("Warning: this version of the reports can't linearize files! "
"Using original commands.")
return load_commands_preserve(args, file_idx, local_filename)
else:
return loaded_commands
def scrape_file(coqargs: List[str], args: argparse.Namespace, includes: str,
file_tuple: Tuple[int, str]) -> Optional[str]:
sys.setrecursionlimit(4500)
file_idx, filename = file_tuple
full_filename = args.prelude + "/" + filename
result_file = full_filename + ".scrape"
temp_file = full_filename + ".scrape.partial"
if args.cont:
with contextlib.suppress(FileNotFoundError):
with open(result_file, 'r') as f:
if args.verbose:
eprint(f"Found existing scrape at {result_file}! Using it")
return result_file
try:
if args.linearize:
commands = serapi_instance.try_load_lin(args, file_idx,
full_filename)
if not commands:
commands = linearize_semicolons.preprocess_file_commands(
args, file_idx,
serapi_instance.load_commands_preserve(
args, 0, full_filename), coqargs, args.prelude,
full_filename, filename, args.skip_nochange_tac)
serapi_instance.save_lin(commands, full_filename)
else:
with Path2(full_filename).open(mode='r') as vf:
commands = serapi_instance.read_commands_preserve(
args, file_idx, vf.read())
with serapi_instance.SerapiContext(
coqargs, serapi_instance.get_module_from_filename(filename),
args.prelude, args.relevant_lemmas == "hammer") as coq:
coq.verbose = args.verbose
try:
with open(temp_file, 'w') as f:
for command in tqdm(commands,
file=sys.stdout,
disable=(not args.progress),
position=file_idx * 2,
desc="Scraping file",
leave=False,
dynamic_ncols=True,
bar_format=mybarfmt):
process_statement(args, coq, command, f)
shutil.move(temp_file, result_file)
return result_file
except serapi_instance.TimeoutError:
eprint("Command in {} timed out.".format(filename))
return temp_file
except Exception as e:
eprint("FAILED: In file {}:".format(filename))
eprint(e)
if args.hardfail or len(args.inputs) == 1 or args.hardfail_scrape:
raise e
return None
def get_commands(self, args: argparse.Namespace, file_idx: int,
filename: str) -> List[str]:
local_filename = self.prelude + "/" + filename
loaded_commands = serapi_instance.try_load_lin(args, file_idx,
local_filename)
if loaded_commands is None:
fresh_commands = linearize_semicolons.preprocess_file_commands(
args, file_idx,
serapi_instance.load_commands_preserve(
args, file_idx, self.prelude + "/" + filename),
self.coqargs, self.includes, filename, local_filename,
self.skip_nochange_tac)
serapi_instance.save_lin(fresh_commands, local_filename)
return fresh_commands
else:
return loaded_commands
def get_linearized(args: argparse.Namespace, coqargs: List[str], bar_idx: int,
filename: str) -> List[str]:
local_filename = args.prelude + "/" + filename
loaded_commands = serapi_instance.try_load_lin(args, bar_idx,
local_filename)
if loaded_commands is None:
original_commands = \
serapi_instance.load_commands_preserve(args, bar_idx,
args.prelude + "/" + filename)
fresh_commands = preprocess_file_commands(args, bar_idx,
original_commands, coqargs,
args.prelude, local_filename,
filename, False)
serapi_instance.save_lin(fresh_commands, local_filename)
return fresh_commands
else:
return loaded_commands
def main():
parser = argparse.ArgumentParser(description="linearize a set of files")
parser.add_argument('--prelude', default=".")
parser.add_argument('--debug',
default=False,
const=True,
action='store_const')
parser.add_argument('--hardfail',
default=False,
const=True,
action='store_const')
parser.add_argument('--verbose',
default=False,
const=True,
action='store_const')
parser.add_argument('--skip-nochange-tac',
default=False,
const=True,
action='store_const',
dest='skip_nochange_tac')
parser.add_argument("--progress",
action='store_const',
const=True,
default=False)
parser.add_argument('filenames', nargs="+", help="proof file name (*.v)")
arg_values = parser.parse_args()
base = os.path.dirname(os.path.abspath(__file__)) + "/.."
includes = subprocess.Popen(
['make', '-C', arg_values.prelude, 'print-includes'],
stdout=subprocess.PIPE).communicate()[0].decode('utf-8')
coqargs = ["sertop"]
for filename in arg_values.filenames:
if arg_values.verbose:
eprint("Linearizing {}".format(filename))
local_filename = arg_values.prelude + "/" + filename
original_commands = serapi_instance.load_commands_preserve(
arg_values, 0, arg_values.prelude + "/" + filename)
fresh_commands = preprocess_file_commands(arg_values, 0,
original_commands, coqargs,
includes, local_filename,
filename, False)
serapi_instance.save_lin(fresh_commands, local_filename)
def reinforce(args: argparse.Namespace) -> None:
# Load the scraped (demonstrated) samples, the proof environment
# commands, and the predictor
replay_memory = assign_rewards(
dataloader.tactic_transitions_from_file(args.scrape_file,
args.buffer_size))
predictor = predict_tactic.loadPredictorByFile(args.predictor_weights)
q_estimator = FeaturesQEstimator(args.learning_rate,
args.batch_step,
args.gamma)
signal.signal(
signal.SIGINT,
lambda signal, frame:
progn(q_estimator.save_weights(
args.out_weights, args), # type: ignore
exit()))
if args.start_from:
q_estimator_name, *saved = \
torch.load(args.start_from)
q_estimator.load_saved_state(*saved)
elif args.pretrain:
pre_train(args, q_estimator,
dataloader.tactic_transitions_from_file(
args.scrape_file, args.buffer_size * 10))
epsilon = 0.3
gamma = 0.9
if args.proof is not None:
assert len(args.environment_files) == 1, \
"Can't use multiple env files with --proof!"
env_commands = serapi_instance.load_commands_preserve(
args, 0, args.prelude / args.environment_files[0])
num_proofs = len([cmd for cmd in env_commands
if cmd.strip() == "Qed."
or cmd.strip() == "Defined."])
with serapi_instance.SerapiContext(
["sertop", "--implicit"],
serapi_instance.get_module_from_filename(
args.environment_files[0]),
str(args.prelude)) as coq:
coq.quiet = True
coq.verbose = args.verbose
rest_commands, run_commands = coq.run_into_next_proof(env_commands)
lemma_statement = run_commands[-1]
while coq.cur_lemma_name != args.proof:
if not rest_commands:
eprint("Couldn't find lemma {args.proof}! Exiting...")
return
rest_commands, _ = coq.finish_proof(rest_commands)
rest_commands, run_commands = coq.run_into_next_proof(
rest_commands)
lemma_statement = run_commands[-1]
reinforce_lemma(args, predictor, q_estimator, coq,
lemma_statement,
epsilon, gamma, replay_memory)
q_estimator.save_weights(args.out_weights, args)
else:
for env_file in args.environment_files:
env_commands = serapi_instance.load_commands_preserve(
args, 0, args.prelude / env_file)
num_proofs = len([cmd for cmd in env_commands
if cmd.strip() == "Qed."
or cmd.strip() == "Defined."])
rest_commands = env_commands
all_run_commands: List[str] = []
with tqdm(total=num_proofs, disable=(not args.progress),
leave=True, desc=env_file.stem) as pbar:
while rest_commands:
with serapi_instance.SerapiContext(
["sertop", "--implicit"],
serapi_instance.get_module_from_filename(
env_file),
str(args.prelude),
log_outgoing_messages=args.log_outgoing_messages) \
as coq:
coq.quiet = True
coq.verbose = args.verbose
for command in all_run_commands:
coq.run_stmt(command)
while rest_commands:
rest_commands, run_commands = \
coq.run_into_next_proof(rest_commands)
if not rest_commands:
break
all_run_commands += run_commands[:-1]
lemma_statement = run_commands[-1]
# Check if the definition is
# proof-relevant. If it is, then finishing
# subgoals doesn't necessarily mean you've
# solved the problem, so don't try to
# train on it.
proof_relevant = False
for cmd in rest_commands:
if serapi_instance.ending_proof(cmd):
if cmd.strip() == "Defined.":
proof_relevant = True
break
proof_relevant = proof_relevant or \
bool(re.match(r"\s*Derive", lemma_statement))
for sample in replay_memory:
sample.graph_node = None
if not proof_relevant:
try:
reinforce_lemma(args, predictor,
q_estimator,
coq,
lemma_statement,
epsilon, gamma,
replay_memory)
except serapi_instance.CoqAnomaly:
if args.log_anomalies:
with args.log_anomalies.open('a') as f:
traceback.print_exc(file=f)
if args.hardfail:
eprint(
"Hit an anomaly!"
"Quitting due to --hardfail")
raise
eprint(
"Hit an anomaly! "
"Restarting coq instance")
rest_commands.insert(0, lemma_statement)
break
pbar.update(1)
rest_commands, run_commands = \
coq.finish_proof(rest_commands)
all_run_commands.append(lemma_statement)
all_run_commands += run_commands
q_estimator.save_weights(args.out_weights, args)
