def __init__(self, model_file=None, model_type=config.SPARSE_PERCEPTRON, beam=1):
self.state = None # State object created at each parse
self.oracle = None # Oracle object created at each parse
self.scores = None # NumPy array of action scores at each action
self.action_count = 0
self.correct_count = 0
self.total_actions = 0
self.total_correct = 0
self.model = Model(model_type, model_file, Actions().all)
self.beam = beam
self.learning_rate = Config().args.learningrate
self.decay_factor = Config().args.decayfactor
self.state_hash_history = None # For loop checking
# Used in verify_passage to optionally ignore a mismatch in linkage nodes:
self.ignore_node = lambda n: n.tag == layer1.NodeTags.Linkage if Config().args.nolinkage else None
self.best_score = self.best_model = self.dev = self.iteration = self.batch = None
class Parser(object):
"""
Main class to implement transition-based UCCA parser
"""
def __init__(self, model_file=None, model_type=config.SPARSE_PERCEPTRON, beam=1):
self.state = None # State object created at each parse
self.oracle = None # Oracle object created at each parse
self.scores = None # NumPy array of action scores at each action
self.action_count = 0
self.correct_count = 0
self.total_actions = 0
self.total_correct = 0
self.model = Model(model_type, model_file, Actions().all)
self.beam = beam
self.learning_rate = Config().args.learningrate
self.decay_factor = Config().args.decayfactor
self.state_hash_history = None # For loop checking
# Used in verify_passage to optionally ignore a mismatch in linkage nodes:
self.ignore_node = lambda n: n.tag == layer1.NodeTags.Linkage if Config().args.nolinkage else None
self.best_score = self.best_model = self.dev = self.iteration = self.batch = None
def train(self, passages, dev=None, iterations=1, folds=None):
"""
Train parser on given passages
:param passages: iterable of passages to train on
:param dev: iterable of passages to tune on
:param iterations: number of iterations to perform
:param folds: whether we are inside cross-validation with this many folds
:return: trained model
"""
if not passages:
self.model.load() # Nothing to train on; pre-trained model given
return
self.best_score = 0
self.best_model = None
self.dev = dev
last = False
if Config().args.devscores:
with open(Config().args.devscores, "w") as f:
print(",".join(["iteration"] + evaluation.Scores.field_titles()), file=f)
for self.iteration in range(iterations):
if last:
break
last = self.iteration == iterations - 1
self.batch = 0
print("Training iteration %d of %d: " % (self.iteration + 1, iterations))
passages = [passage for _, passage in self.parse(passages, mode="train")]
if last:
if folds is None: # Free some memory, as these are not needed any more
del passages[:]
else:
self.learning_rate *= self.decay_factor
Config().random.shuffle(passages)
last = self.eval_dev_and_save_model(last)
if self.dev and last and folds is None: # Free more memory
del self.dev[:]
print("Trained %d iterations" % iterations)
self.model = self.best_model
def eval_dev_and_save_model(self, last=False):
model = self.model # Save non-finalize model
self.model = self.model.finalize() # To evaluate finalized model on dev
save_model = True
if self.dev:
print("Evaluating on dev passages")
self.dev, scores = zip(*[(passage, evaluate_passage(predicted_passage, passage))
for predicted_passage, passage in
self.parse(self.dev, mode="dev")])
self.dev = list(self.dev)
scores = evaluation.Scores.aggregate(scores)
score = scores.average_f1()
print("Average labeled F1 score on dev: %.3f" % score)
if Config().args.devscores:
prefix = [self.iteration]
if Config().args.saveeverybatch:
prefix.append(self.batch)
with open(Config().args.devscores, "a") as f:
print(",".join([".".join(map(str, prefix))] + scores.fields()), file=f)
if score >= self.best_score:
print("Better than previous best score (%.3f)" % self.best_score)
self.best_score = score
save_model = True
else:
print("Not better than previous best score (%.3f)" % self.best_score)
save_model = False
if score >= 1: # Score cannot go any better, so no point in more training
last = True
if save_model or self.best_model is None:
self.best_model = self.model # This is the finalized model
self.best_model.save()
if not last:
self.model = model # Restore non-finalized model
return last
def parse(self, passages, mode="test"):
"""
Parse given passages
:param passages: iterable of passages to parse
:param mode: "train", "test" or "dev".
If "train", use oracle to train on given passages.
Otherwise, just parse with classifier.
:return: generator of pairs of (parsed passage, original passage)
"""
train = mode == "train"
dev = mode == "dev"
test = mode == "test"
assert train or dev or test, "Invalid parse mode: %s" % mode
passage_word = "sentence" if Config().args.sentences else \
"paragraph" if Config().args.paragraphs else \
"passage"
self.total_actions = 0
self.total_correct = 0
total_duration = 0
total_tokens = 0
num_passages = 0
for passage in passages:
l0 = passage.layer(layer0.LAYER_ID)
num_tokens = len(l0.all)
l1 = passage.layer(layer1.LAYER_ID)
labeled = len(l1.all) > 1
assert not train or labeled, "Cannot train on unannotated passage"
print("%s %-7s" % (passage_word, passage.ID), end=Config().line_end, flush=True)
started = time.time()
self.action_count = 0
self.correct_count = 0
self.state = State(passage, callback=self.pos_tag)
self.state_hash_history = set()
self.oracle = Oracle(passage) if train else None
failed = False
try:
self.parse_passage(train) # This is where the actual parsing takes place
except ParserException as e:
if train:
raise
Config().log("%s %s: %s" % (passage_word, passage.ID, e))
if not test:
print("failed")
failed = True
predicted_passage = passage
if not train or Config().args.verify:
predicted_passage = self.state.create_passage(assert_proper=Config().args.verify)
duration = time.time() - started
total_duration += duration
num_tokens -= len(self.state.buffer)
total_tokens += num_tokens
if train: # We have an oracle to verify by
if not failed and Config().args.verify:
self.verify_passage(passage, predicted_passage, train)
if self.action_count:
print("%-16s" % ("%d%% (%d/%d)" %
(100 * self.correct_count / self.action_count,
self.correct_count, self.action_count)), end=Config().line_end)
print("%0.3fs" % duration, end="")
print("%-15s" % ("" if failed else " (%d tokens/s)" % (num_tokens / duration)), end="")
print(Config().line_end, end="")
if train:
print(Config().line_end, flush=True)
self.model.finish(train=train)
self.total_correct += self.correct_count
self.total_actions += self.action_count
num_passages += 1
if train and Config().args.saveeverybatch and num_passages % Config().args.batchsize == 0:
self.eval_dev_and_save_model()
self.batch += 1
yield predicted_passage, passage
if num_passages > 1:
print("Parsed %d %ss" % (num_passages, passage_word))
if self.oracle and self.total_actions:
print("Overall %d%% correct transitions (%d/%d) on %s" %
(100 * self.total_correct / self.total_actions,
self.total_correct, self.total_actions,
mode))
print("Total time: %.3fs (average time/%s: %.3fs, average tokens/s: %d)" % (
total_duration, passage_word, total_duration / num_passages,
total_tokens / total_duration), flush=True)
def parse_passage(self, train=False):
"""
Internal method to parse a single passage
:param train: use oracle to train on given passages, or just parse with classifier?
"""
if Config().args.verbose:
print(" initial state: %s" % self.state)
while True:
if Config().args.checkloops:
self.check_loop(print_oracle=train)
true_actions = []
if self.oracle is not None:
try:
true_actions = self.oracle.get_actions(self.state)
except (AttributeError, AssertionError) as e:
if train:
raise ParserException("Error in oracle during training") from e
features = self.model.extract_features(self.state)
predicted_action = self.predict_action(features, true_actions) # sets self.scores
action = predicted_action
correct_action = False
if not true_actions:
true_actions = "?"
elif predicted_action in true_actions:
self.correct_count += 1
correct_action = True
elif train:
action = Config().random.choice(true_actions)
if train and not (correct_action and self.model.update_only_on_error):
best_true_action = true_actions[0] if len(true_actions) == 1 else \
true_actions[self.scores[[a.id for a in true_actions]].argmax()]
rate = self.learning_rate
if best_true_action.is_swap:
rate *= Config().args.importance
self.model.update(features, predicted_action.id, best_true_action.id, rate)
self.model.advance()
self.action_count += 1
try:
self.state.transition(action)
except AssertionError as e:
raise ParserException("Invalid transition (%s): %s" % (action, e)) from e
if Config().args.verbose:
if self.oracle is None:
print(" action: %-15s %s" % (action, self.state))
else:
print(" predicted: %-15s true: %-15s taken: %-15s %s" % (
predicted_action, "|".join(map(str, true_actions)), action, self.state))
for line in self.state.log:
print(" " + line)
if self.state.finished or train and not correct_action and Config().args.earlyupdate:
return # action is FINISH
def check_loop(self, print_oracle):
"""
Check if the current state has already occurred, indicating a loop
:param print_oracle: whether to print the oracle in case of an assertion error
"""
h = hash(self.state)
assert h not in self.state_hash_history,\
"\n".join(["Transition loop", self.state.str("\n")] +
[self.oracle.str("\n")] if print_oracle else ())
self.state_hash_history.add(h)
def predict_action(self, features, true_actions):
"""
Choose action based on classifier
:param features: extracted feature values
:param true_actions: from the oracle, to copy orig_node if the same action is selected
:return: valid action with maximum probability according to classifier
"""
self.scores = self.model.score(features) # Returns a NumPy array
best_action = self.select_action(self.scores.argmax(), true_actions)
if self.state.is_valid(best_action):
return best_action
# Usually the best action is valid, so max is enough to choose it in O(n) time
# Otherwise, sort all the other scores to choose the best valid one in O(n lg n)
sorted_ids = self.scores.argsort()[::-1]
actions = (self.select_action(i, true_actions) for i in sorted_ids)
try:
return next(a for a in actions if self.state.is_valid(a))
except StopIteration as e:
raise ParserException("No valid actions available\n" +
("True actions: %s" % true_actions if true_actions
else self.oracle.log if self.oracle is not None
else "") +
"\nReturned actions: %s" %
[self.select_action(i) for i in sorted_ids] +
"\nScores: %s" % self.scores
) from e
@staticmethod
def select_action(i, true_actions=()):
"""
Find action with the given ID in true actions (if exists) or in all actions
:param i: ID to lookup
:param true_actions: preferred set of actions to look in first
:return: Action with id=i
"""
try:
return next(a for a in true_actions if a.id == i)
except StopIteration:
return Actions().all[i]
def verify_passage(self, passage, predicted_passage, show_diff):
"""
Compare predicted passage to true passage and die if they differ
:param passage: true passage
:param predicted_passage: predicted passage to compare
:param show_diff: if passages differ, show the difference between them?
Depends on predicted_passage having the original node IDs annotated
in the "remarks" field for each node.
"""
assert passage.equals(predicted_passage, ignore_node=self.ignore_node),\
"Failed to produce true passage" + \
(diffutil.diff_passages(
passage, predicted_passage) if show_diff else "")
@staticmethod
def pos_tag(state):
"""
Function to pass to State to POS tag the tokens when created
:param state: State object to modify
"""
tokens = [token for tokens in state.tokens for token in tokens]
tokens, tags = zip(*pos_tag(tokens))
if Config().args.verbose:
print(" ".join("%s/%s" % (token, tag) for (token, tag) in zip(tokens, tags)))
for node, tag in zip(state.nodes, tags):
node.pos_tag = tag
