def distribute(node):
"""
Wrap a copy of the given element around the contents of each of its
children, removing the node in the process.
"""
children = list(c for c in node.childNodes if is_element(c))
unwrap(node)
tag_name = node.tagName
for c in children:
wrap_inner(c, tag_name)
def distribute(node):
"""
Wrap a copy of the given element around the contents of each of its
children, removing the node in the process.
"""
children = list(c for c in node.childNodes if is_element(c))
unwrap(node)
tag_name = node.tagName
for c in children:
wrap_inner(c, tag_name)
def remove_nesting(dom, tag_name):
"""
Unwrap items in the node list that have ancestors with the same tag.
"""
for node in dom.getElementsByTagName(tag_name):
for ancestor in ancestors(node):
if ancestor is node:
continue
if ancestor is dom.documentElement:
break
if ancestor.tagName == tag_name:
unwrap(node)
break
def remove_nesting(dom, tag_name):
"""
Unwrap items in the node list that have ancestors with the same tag.
"""
for node in dom.getElementsByTagName(tag_name):
for ancestor in ancestors(node):
if ancestor is node:
continue
if ancestor is dom.documentElement:
break
if ancestor.tagName == tag_name:
unwrap(node)
break
def read(self, read_size, decode_hdlc=False):
buf = b''
buf = self.port.read(read_size)
buf = bytes(buf)
if decode_hdlc:
buf = util.unwrap(write_buf)
return buf
def dubins_trajectory(q0,
qf,
radius,
velocity_real,
step_size,
contact_angle=0):
velocity = velocity_real * 10
#1. compute 2d trajectory
configurations, N_star = compute_dubins_base(q0,
qf,
radius,
velocity_real,
step_size,
contact_angle=contact_angle)
#2. convert trajectory back to original coordinate system (account for different object sides)
configurations_transformed = []
for counter, configuration in enumerate(configurations):
configurations_transformed.append(
rotate_2d_pose_origin(configuration, contact_angle, q0[0:2]))
#3. reformat object trajectory
t_star = np.array(N_star) / velocity
x_star = np.array(configurations_transformed)
if len(configurations) > 0:
x_star[:, 2] = util.unwrap(x_star[:, 2])
return configurations_transformed, N_star, x_star, t_star
def read(self, read_size, decode_hdlc = False):
buf = b''
try:
buf = self.r_handle.read(read_size)
buf = bytes(buf)
except usb.core.USBError:
return b''
if decode_hdlc:
buf = util.unwrap(write_buf)
return buf
def read(self, read_size, decode_hdlc=False):
buf = b''
try:
buf = self.f.read(read_size)
buf = bytes(buf)
except:
return b''
if decode_hdlc:
buf = util.unwrap(write_buf)
return buf
def lazy_multiprocessing_imap(worker: Callable[[T], O], in_data : Iterable[T],
num_threads : Optional[int]=os.cpu_count(),
chunk_size : Optional[int]=None) -> Iterable[O]:
if chunk_size == None:
if num_threads:
chunk_size = num_threads * 10
else:
chunk_size = 100
with multiprocessing.Pool(num_threads) as pool:
for chunk in chunks(in_data, unwrap(chunk_size)):
yield from list(pool.imap(worker, chunk))
def fix_lists(dom):
# <ins> and <del> tags are not allowed within <ul> or <ol> tags.
# Move them to the nearest li, so that the numbering isn't interrupted.
# Find all del > li and ins > li sets.
del_tags = set()
ins_tags = set()
for node in list(dom.getElementsByTagName('li')):
parent = node.parentNode
if parent.tagName == 'del':
del_tags.add(parent)
elif parent.tagName == 'ins':
ins_tags.add(parent)
# Change ins > li into li > ins.
for ins_tag in ins_tags:
distribute(ins_tag)
# Change del > li into li.del-li > del.
for del_tag in del_tags:
children = list(del_tag.childNodes)
unwrap(del_tag)
for c in children:
if c.nodeName == 'li':
c.setAttribute('class', 'del-li')
wrap_inner(c, 'del')
def fix_lists(dom):
# <ins> and <del> tags are not allowed within <ul> or <ol> tags.
# Move them to the nearest li, so that the numbering isn't interrupted.
# Find all del > li and ins > li sets.
del_tags = set()
ins_tags = set()
for node in list(dom.getElementsByTagName('li')):
parent = node.parentNode
if parent.tagName == 'del':
del_tags.add(parent)
elif parent.tagName == 'ins':
ins_tags.add(parent)
# Change ins > li into li > ins.
for ins_tag in ins_tags:
distribute(ins_tag)
# Change del > li into li.del-li > del.
for del_tag in del_tags:
children = list(del_tag.childNodes)
unwrap(del_tag)
for c in children:
if c.nodeName == 'li':
c.setAttribute('class', 'del-li')
wrap_inner(c, 'del')
def parse_diag(self, pkt, hdlc_encoded=True, check_crc=True, radio_id=0):
# Should contain DIAG command and CRC16
# pkt should not contain trailing 0x7E, and either HDLC encoded or not
# When the pkt is not HDLC encoded, hdlc_encoded should be set to True
# radio_id = 0 for default, larger than 1 for SIM 1 and such
if len(pkt) < 3:
return
if hdlc_encoded:
pkt = util.unwrap(pkt)
# Check and strip CRC if existing
if check_crc:
crc = util.dm_crc16(pkt[:-2])
crc_pkt = (pkt[-1] << 8) | pkt[-2]
if crc != crc_pkt:
self.logger.log(
logging.WARNING,
"CRC mismatch: expected 0x{:04x}, got 0x{:04x}".format(
crc, crc_pkt))
self.logger.log(logging.DEBUG, util.xxd(pkt))
pkt = pkt[:-2]
if pkt[0] == diagcmd.DIAG_LOG_F:
self.parse_diag_log(pkt, radio_id)
elif pkt[0] == diagcmd.DIAG_EVENT_REPORT_F and self.parse_events:
self.parse_diag_event(pkt, radio_id)
elif pkt[0] == diagcmd.DIAG_EXT_MSG_F and self.parse_msgs:
self.parse_diag_ext_msg(pkt, radio_id)
elif pkt[0] == diagcmd.DIAG_QSR_EXT_MSG_TERSE_F and self.parse_msgs:
#self.parse_diag_qsr_ext_msg(pkt, radio_id)
pass
elif pkt[0] == diagcmd.DIAG_QSR4_EXT_MSG_TERSE_F and self.parse_msgs:
#self.parse_diag_qsr4_ext_msg(pkt, radio_id)
pass
elif pkt[0] == diagcmd.DIAG_MULTI_RADIO_CMD_F:
self.parse_diag_multisim(pkt)
else:
#print("Not parsing non-Log packet %02x" % pkt[0])
#util.xxd(pkt)
return
def action(self) -> str:
return unwrap(self.transition).action
def reinforce_lemma_multithreaded(
args: argparse.Namespace,
coq: serapi_instance.SerapiInstance,
lock: Lock,
namespace: multiprocessing.managers.Namespace,
worker_idx: int,
samples: Queue[LabeledTransition],
lemma_statement: str,
_module_prefix: str,
demonstration: Optional[Demonstration]) -> Tuple[str, ReinforceGraph]:
lemma_name = serapi_instance.lemma_name_from_statement(lemma_statement)
graph = ReinforceGraph(lemma_name)
lemma_memory = []
for i in trange(args.num_episodes, disable=(not args.progress),
leave=False, position=worker_idx + 1):
cur_node = graph.start_node
proof_contexts_seen = [unwrap(coq.proof_context)]
episode_memory: List[LabeledTransition] = []
reached_qed = False
for t in range(args.episode_length):
context_before = coq.tactic_context(coq.local_lemmas[:-1])
proof_context_before = unwrap(coq.proof_context)
context_trunced = truncate_tactic_context(
context_before, args.max_term_length)
if (demonstration and
t < len(demonstration) -
((i//args.demonstration_steps)+1)):
eprint("Getting demonstration", guard=args.verbose >= 2)
ordered_actions = [(demonstration[t],
certainty_of(namespace.predictor,
args.num_predictions * 2,
context_trunced,
demonstration[t]))]
else:
with print_time("Getting predictions", guard=args.verbose >= 2):
with lock:
eprint(f"Locked in thread {worker_idx}",
guard=args.verbose >= 2)
predictor = namespace.predictor
estimator = namespace.estimator
with print_time("Making predictions",
guard=args.verbose >= 3):
predictions = predictor.predictKTactics(
context_trunced, args.num_predictions)
if random.random() < args.exploration_factor:
eprint("Picking random action",
guard=args.verbose >= 2)
ordered_actions = order_by_score(predictions)
else:
with print_time("Picking actions with q_estimator",
guard=args.verbose >= 2):
q_choices = zip(estimator(
[(context_trunced,
p.prediction, p.certainty)
for p in predictions]),
predictions)
ordered_actions = [p[1] for p in
sorted(q_choices,
key=lambda q: q[0],
reverse=True)]
eprint(f"Unlocked in thread {worker_idx}",
guard=args.verbose >= 2)
with print_time("Running actions", guard=args.verbose >= 2):
action = None
original_certainty = None
for try_action, try_original_certainty in ordered_actions:
try:
coq.run_stmt(try_action)
proof_context_after = unwrap(coq.proof_context)
if any([serapi_instance.contextSurjective(
proof_context_after, path_context)
for path_context in proof_contexts_seen]):
coq.cancel_last()
if args.ghosts:
transition = assign_failed_reward(
context_trunced.relevant_lemmas,
context_trunced.prev_tactics,
proof_context_before,
proof_context_after,
try_action,
try_original_certainty,
0)
ghost_node = graph.addGhostTransition(
cur_node, transition)
transition.graph_node = ghost_node
continue
action = try_action
original_certainty = try_original_certainty
break
except (serapi_instance.ParseError,
serapi_instance.CoqExn,
serapi_instance.TimeoutError):
if args.ghosts:
transition = assign_failed_reward(
context_trunced.relevant_lemmas,
context_trunced.prev_tactics,
proof_context_before,
proof_context_before,
try_action,
try_original_certainty,
0)
ghost_node = graph.addGhostTransition(cur_node,
transition)
transition.graph_node = ghost_node
if action is None:
# We'll hit this case of we tried all of the
# predictions, and none worked
graph.setNodeColor(cur_node, "red")
transition = assign_failed_reward(
context_trunced.relevant_lemmas,
context_trunced.prev_tactics,
proof_context_before,
proof_context_before,
"Abort.",
try_original_certainty,
-25)
samples.put(transition)
episode_memory.append(transition)
break # Break from episode
transition = assign_reward(args,
context_trunced.relevant_lemmas,
context_trunced.prev_tactics,
proof_context_before,
proof_context_after,
action,
unwrap(original_certainty))
cur_node = graph.addTransition(cur_node, transition)
transition.graph_node = cur_node
assert transition.reward < 2000
samples.put(transition)
episode_memory.append(transition)
proof_contexts_seen.append(proof_context_after)
lemma_memory += episode_memory
if coq.goals == "":
eprint("QED!", guard=args.verbose >= 2)
graph.mkQED(cur_node)
for sample in (episode_memory *
(args.success_repetitions - 1)):
samples.put(sample)
reached_qed = True
break
if not reached_qed:
# We'll hit this case of we tried all of the
# predictions, and none worked
graph.setNodeColor(cur_node, "red")
transition = assign_failed_reward(
context_trunced.relevant_lemmas,
context_trunced.prev_tactics,
proof_context_before,
proof_context_before,
"Abort.",
try_original_certainty,
-25)
samples.put(transition)
episode_memory.append(transition)
# Clean up episode
if lemma_name:
coq.run_stmt("Admitted.")
coq.run_stmt(f"Reset {lemma_name}.")
else:
coq.cancel_last()
while coq.goals:
coq.cancel_last()
coq.run_stmt(lemma_statement)
# Write out lemma memory to progress file for resuming
for sample in lemma_memory:
with args.out_weights.with_suffix('.tmp').open('a') as f:
f.write(json.dumps(sample.to_dict()) + "\n")
graphpath = (args.graphs_dir / lemma_name).with_suffix(".png")
return str(graphpath), graph
def _strip_changes_old(node):
for ins_node in node.getElementsByTagName('ins'):
remove_node(ins_node)
for del_node in node.getElementsByTagName('del'):
unwrap(del_node)
def dataloader_args(self) -> DataloaderArgs:
return extract_dataloader_args(unwrap(self.training_args))
def hyp_encoder(self) -> HypArgEncoder:
return unwrap(self._model).hyp_model.arg_encoder
def entire_goal_encoder(self) -> EncoderRNN:
return unwrap(self._model).goal_encoder
def goal_token_encoder(self) -> GoalTokenEncoderModel:
return unwrap(self._model).goal_args_model.encoder_model
def reward(self) -> float:
return unwrap(self.transition).reward
def _strip_changes_old(node):
for ins_node in node.getElementsByTagName('ins'):
remove_node(ins_node)
for del_node in node.getElementsByTagName('del'):
unwrap(del_node)
def reinforce_lemma(args: argparse.Namespace,
predictor: tactic_predictor.TacticPredictor,
estimator: q_estimator.QEstimator,
coq: serapi_instance.SerapiInstance,
lemma_statement: str,
epsilon: float,
gamma: float,
memory: List[LabeledTransition]) -> None:
lemma_name = coq.cur_lemma_name
graph = ReinforceGraph(lemma_name)
for episode in trange(args.num_episodes, disable=(not args.progress),
leave=False):
cur_node = graph.start_node
proof_contexts_seen = [unwrap(coq.proof_context)]
episode_memory = []
for t in range(args.episode_length):
with print_time("Getting predictions", guard=args.verbose):
context_before = coq.tactic_context(coq.local_lemmas[:-1])
proof_context_before = unwrap(coq.proof_context)
predictions = predictor.predictKTactics(
context_before, args.num_predictions)
if random.random() < epsilon:
ordered_actions = [p.prediction for p in
random.sample(predictions,
len(predictions))]
else:
with print_time("Picking actions using q_estimator",
guard=args.verbose):
q_choices = zip(estimator(
[(context_before, prediction.prediction)
for prediction in predictions]),
[p.prediction for p in predictions])
ordered_actions = [p[1] for p in
sorted(q_choices,
key=lambda q: q[0],
reverse=True)]
with print_time("Running actions", guard=args.verbose):
action = None
for try_action in ordered_actions:
try:
coq.run_stmt(try_action)
proof_context_after = unwrap(coq.proof_context)
if any([serapi_instance.contextSurjective(
proof_context_after, path_context)
for path_context in proof_contexts_seen]):
coq.cancel_last()
transition = assign_failed_reward(
context_before.relevant_lemmas,
context_before.prev_tactics,
proof_context_before,
proof_context_after,
try_action,
-50)
assert transition.reward < 2000
memory.append(transition)
if args.ghosts:
ghost_node = graph.addGhostTransition(
cur_node, try_action)
transition.graph_node = ghost_node
continue
action = try_action
break
except (serapi_instance.ParseError,
serapi_instance.CoqExn,
serapi_instance.TimeoutError):
transition = assign_failed_reward(
context_before.relevant_lemmas,
context_before.prev_tactics,
proof_context_before,
proof_context_before,
try_action,
-500)
assert transition.reward < 2000
memory.append(transition)
if args.ghosts:
ghost_node = graph.addGhostTransition(cur_node,
try_action)
transition.graph_node = ghost_node
pass
if action is None:
# We'll hit this case of we tried all of the
# predictions, and none worked
graph.setNodeColor(cur_node, "red")
break # Break from episode
transition = assign_reward(context_before.relevant_lemmas,
context_before.prev_tactics,
proof_context_before,
proof_context_after,
action)
cur_node = graph.addTransition(cur_node, action,
transition.reward)
transition.graph_node = cur_node
assert transition.reward < 2000
episode_memory.append(transition)
memory.append(transition)
proof_contexts_seen.append(proof_context_after)
if coq.goals == "":
graph.mkQED(cur_node)
memory += (episode_memory * (args.success_repetitions - 1))
break
with print_time("Assigning scores", guard=args.verbose):
transition_samples = sample_batch(memory,
args.batch_size)
training_samples = assign_scores(transition_samples,
estimator, predictor,
args.num_predictions,
gamma,
# Passing this graph
# in so we can
# maintain a record
# of the most recent
# q score estimates
# in the graph
graph)
with print_time("Training", guard=args.verbose):
estimator.train(training_samples)
# Clean up episode
coq.run_stmt("Admitted.")
coq.run_stmt(f"Reset {lemma_name}.")
coq.run_stmt(lemma_statement)
graphpath = (args.graphs_dir / lemma_name).with_suffix(".png")
graph.draw(str(graphpath))
pass