def check_eval(eval_treebank, ep, flag='dev'):
# nonlocal best_eval_fscore
# nonlocal best_eval_model_path
# nonlocal best_eval_processed
dev_start_time = time.time()
eval_predicted = []
for dev_start_index in range(0, len(eval_treebank),
args.eval_batch_size):
subbatch_trees = eval_treebank[dev_start_index:dev_start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word)
for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(subbatch_sentences)
del _
eval_predicted.extend([p.convert() for p in predicted])
eval_fscore = evaluate.evalb(args.evalb_dir, eval_treebank,
eval_predicted)
logger.info(flag + ' eval '
'epoch {} '
"fscore {} "
"elapsed {} "
"total-elapsed {}".format(
ep,
eval_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
))
return eval_fscore
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = treebanks.load_trees(args.test_path, args.test_path_text,
args.text_processing)
print("Loaded {:,} test examples.".format(len(test_treebank)))
if len(args.model_path) != 1:
raise NotImplementedError("Ensembling multiple parsers is not "
"implemented in this version of the code.")
model_path = args.model_path[0]
print("Loading model from {}...".format(model_path))
parser = parse_chart.ChartParser.from_trained(model_path)
if args.no_predict_tags and parser.f_tag is not None:
print("Removing part-of-speech tagging head...")
parser.f_tag = None
if args.parallelize:
parser.parallelize()
elif torch.cuda.is_available():
parser.cuda()
print("Parsing test sentences...")
start_time = time.time()
test_predicted = parser.parse(
test_treebank.without_gold_annotations(),
subbatch_max_tokens=args.subbatch_max_tokens,
)
if args.output_path == "-":
for tree in test_predicted:
print(tree.pformat(margin=1e100))
elif args.output_path:
with open(args.output_path, "w") as outfile:
for tree in test_predicted:
outfile.write("{}\n".format(tree.pformat(margin=1e100)))
# The tree loader does some preprocessing to the trees (e.g. stripping TOP
# symbols or SPMRL morphological features). We compare with the input file
# directly to be extra careful about not corrupting the evaluation. We also
# allow specifying a separate "raw" file for the gold trees: the inputs to
# our parser have traces removed and may have predicted tags substituted,
# and we may wish to compare against the raw gold trees to make sure we
# haven't made a mistake. As far as we can tell all of these variations give
# equivalent results.
ref_gold_path = args.test_path
if args.test_path_raw is not None:
print("Comparing with raw trees from", args.test_path_raw)
ref_gold_path = args.test_path_raw
test_fscore = evaluate.evalb(args.evalb_dir,
test_treebank.trees,
test_predicted,
ref_gold_path=ref_gold_path)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
# model = dy.ParameterCollection()
# [parser] = dy.load(args.model_path_base, model)
parser = torch.load(args.model_path_base)
print("Parsing test sentences...")
start_time = time.time()
test_predicted = []
for tree in test_treebank:
# dy.renew_cg()
parser.eval()
sentence = [(leaf.tag, leaf.word) for leaf in tree.leaves()]
predicted, _ = parser.parse(sentence)
test_predicted.append(predicted.convert())
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, test_predicted)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def check_dev(epoch_num):
nonlocal best_dev_fscore
nonlocal best_model_path
nonlocal best_dev_processed
dev_start_time = time.time()
parser.eval()
dev_predicted = []
for dev_start_index in range(0, len(dev_treebank),
args.eval_batch_size):
subbatch_trees = dev_treebank[dev_start_index:dev_start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word)
for leaf in tree.leaves()]
for tree in subbatch_trees]
(
predicted,
_,
) = parser.parse_batch(subbatch_sentences)
del _
dev_predicted.extend([p.convert() for p in predicted])
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
print("\n"
"dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time)))
if dev_fscore.fscore > best_dev_fscore:
if best_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_model_path + ext
if os.path.exists(path):
print(
"Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
best_model_path = "{}_best_dev={:.2f}".format(
args.model_path_base, dev_fscore.fscore)
best_dev_processed = total_processed
print("Saving new best model to {}...".format(best_model_path))
torch.save(
{
"spec": parser.spec,
"state_dict": parser.state_dict(),
"trainer": trainer.state_dict(),
},
best_model_path + ".pt",
)
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
nonlocal dev_efscore
dev_start_time = time.time()
dev_predicted = []
for dev_start_index in range(0, len(dev_treebank),
args.eval_batch_size):
subbatch_trees = dev_treebank[dev_start_index:dev_start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word)
for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(subbatch_sentences)
del _
dev_predicted.extend([p.convert() for p in predicted])
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
print(" dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
),
flush=True)
dev_efscore = evaluate_EDITED.Evaluate(dev_treebank, dev_predicted)
print(" dev-Efscore: {}".format(dev_efscore), flush=True)
# MJ - keep model with best efscore
if dev_efscore.efscore > best_dev_fscore:
best_dev_fscore = dev_efscore.efscore
if best_dev_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_dev_model_path + ext
if os.path.exists(path):
print(
" Removing previous model file {}...".format(path),
flush=True)
os.remove(path)
best_dev_model_path = "{}_Edev={:.4}".format(
args.model_path_base, best_dev_fscore)
print(" Saving new best model to {}...".format(
best_dev_model_path),
flush=True)
torch.save(
{
'spec': parser.spec,
'state_dict': parser.state_dict(),
'trainer': trainer.state_dict(),
}, best_dev_model_path + ".pt")
def run_parse_extra(args):
if args.output_path != '-' and os.path.exists(args.output_path):
print("Error: output file already exists:", args.output_path)
return
print("Loading parse trees from {}...".format(args.input_path))
treebank = trees.load_trees(args.input_path)
if args.max_len_eval > 0:
treebank = [
tree for tree in treebank
if len(list(tree.leaves())) <= args.max_len_eval
]
print("Loaded {:,} parse tree examples.".format(len(treebank)))
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(
".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = parse_nk.NKChartParser.from_spec(info['spec'], info['state_dict'])
print("Parsing test sentences...")
start_time = time.time()
new_treebank = []
for start_index in range(0, len(treebank), args.eval_batch_size):
subbatch_trees = treebank[start_index:start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(subbatch_sentences)
del _
new_treebank.extend([p.convert() for p in predicted])
assert len(treebank) == len(new_treebank), (len(treebank),
len(new_treebank))
if args.write_parse is not None:
print('writing to {}'.format(args.write_parse))
f = open(args.write_parse, 'w')
for x, y in zip(new_treebank, treebank):
gold = '(ROOT {})'.format(y.linearize())
pred = '(ROOT {})'.format(x.linearize())
ex = dict(gold=gold, pred=pred)
f.write(json.dumps(ex) + '\n')
f.close()
test_fscore = evaluate.evalb(args.evalb_dir,
treebank,
new_treebank,
ref_gold_path=None)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(
".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = SAPar_model.SAChartParser.from_spec(info['spec'],
info['state_dict'])
print("Parsing test sentences...")
start_time = time.time()
test_predicted = []
for start_index in tqdm(range(0, len(test_treebank),
args.eval_batch_size)):
subbatch_trees = test_treebank[start_index:start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(subbatch_sentences)
del _
test_predicted.extend([p.convert() for p in predicted])
# The tree loader does some preprocessing to the trees (e.g. stripping TOP
# symbols or SPMRL morphological features). We compare with the input file
# directly to be extra careful about not corrupting the evaluation. We also
# allow specifying a separate "raw" file for the gold trees: the inputs to
# our parser have traces removed and may have predicted tags substituted,
# and we may wish to compare against the raw gold trees to make sure we
# haven't made a mistake. As far as we can tell all of these variations give
# equivalent results.
ref_gold_path = args.test_path
if args.test_path_raw is not None:
print("Comparing with raw trees from", args.test_path_raw)
ref_gold_path = args.test_path_raw
test_fscore = evaluate.evalb(args.evalb_dir,
test_treebank,
test_predicted,
ref_gold_path=ref_gold_path)
model_name = args.model_path_base[args.model_path_base.rfind('/') +
1:args.model_path_base.rfind('.')]
output_file = './results/' + model_name + '.txt'
with open(output_file, "w") as outfile:
for tree in test_predicted:
outfile.write("{}\n".format(tree.linearize()))
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
nonlocal best_dev_processed
dev_start_time = time.time()
dev_predicted = []
for dev_start_index in range(0, len(dev_treebank),
args.eval_batch_size):
subbatch_trees = dev_treebank[dev_start_index:dev_start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word)
for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(
subbatch_sentences,
span_index=span_index,
k=K,
zero_empty=parser.zero_empty,
train_nn=args.train_through_nn,
)
del _
dev_predicted.extend([p.convert() for p in predicted])
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
print("dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
))
if dev_fscore.fscore > best_dev_fscore:
if best_dev_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_dev_model_path + ext
if os.path.exists(path):
print(
"Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
best_dev_model_path = "{}_dev={:.2f}".format(
args.model_path_base, dev_fscore.fscore)
best_dev_processed = total_processed
print("Saving new best model to {}...".format(best_dev_model_path))
torch.save(
{
'spec': parser.spec,
'state_dict': parser.state_dict(),
'trainer': trainer.state_dict(),
}, best_dev_model_path + ".pt")
def run_ensemble(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
parsers = []
for model_path_base in args.model_path_base:
print("Loading model from {}...".format(model_path_base))
assert model_path_base.endswith(".pt"), "Only pytorch savefiles supported"
info = torch_load(model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = parse_nk.NKChartParser.from_spec(info['spec'], info['state_dict'])
parsers.append(parser)
# Ensure that label scores charts produced by the models can be combined
# using simple averaging
ref_label_vocab = parsers[0].label_vocab
for parser in parsers:
assert parser.label_vocab.indices == ref_label_vocab.indices
print("Parsing test sentences...")
start_time = time.time()
test_predicted = []
# Ensemble by averaging label score charts from different models
# We did not observe any benefits to doing weighted averaging, probably
# because all our parsers output label scores of around the same magnitude
for start_index in range(0, len(test_treebank), args.eval_batch_size):
subbatch_trees = test_treebank[start_index:start_index+args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()] for tree in subbatch_trees]
chart_lists = []
for parser in parsers:
charts = parser.parse_batch(subbatch_sentences, return_label_scores_charts=True)
chart_lists.append(charts)
subbatch_charts = [np.mean(list(sentence_charts), 0) for sentence_charts in zip(*chart_lists)]
predicted, _ = parsers[0].decode_from_chart_batch(subbatch_sentences, subbatch_charts)
del _
test_predicted.extend([p.convert() for p in predicted])
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, test_predicted, ref_gold_path=args.test_path)
print(
"test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
)
)
def run_test(args):
test_path = args.test_ptb_path
if args.dataset == "ctb":
test_path = args.test_ctb_path
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(
".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert "hparams" in info["spec"], "Older savefiles not supported"
parser = Lparser.ChartParser.from_spec(info["spec"], info["state_dict"])
parser.eval()
print("Loading test trees from {}...".format(test_path))
test_treebank = trees.load_trees(test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Parsing test sentences...")
start_time = time.time()
punct_set = "." "``" "''" ":" ","
parser.eval()
test_predicted = []
for start_index in range(0, len(test_treebank), args.eval_batch_size):
subbatch_trees = test_treebank[start_index:start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in subbatch_trees]
(
predicted,
_,
) = parser.parse_batch(subbatch_sentences)
del _
test_predicted.extend([p.convert() for p in predicted])
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, test_predicted)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
nonlocal best_dev_processed
dev_start_time = time.time()
dev_predicted = parser.parse(
dev_treebank.without_gold_annotations(),
subbatch_max_tokens=args.subbatch_max_tokens,
)
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank.trees, dev_predicted)
print(
"dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
)
)
if dev_fscore.fscore > best_dev_fscore:
if best_dev_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_dev_model_path + ext
if os.path.exists(path):
print("Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
best_dev_model_path = "{}_dev={:.2f}".format(
args.model_path_base, dev_fscore.fscore
)
best_dev_processed = total_processed
print("Saving new best model to {}...".format(best_dev_model_path))
torch.save(
{
"config": parser.config,
"state_dict": parser.state_dict(),
"optimizer": optimizer.state_dict(),
},
best_dev_model_path + ".pt",
)
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
dev_start_time = time.time()
dev_predicted = []
for tree in dev_treebank:
# dy.renew_cg()
parser.eval()
sentence = [(leaf.tag, leaf.word) for leaf in tree.leaves()]
predicted, _ = parser.parse(sentence)
dev_predicted.append(predicted.convert())
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
print("dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
))
if dev_fscore.fscore > best_dev_fscore:
if best_dev_model_path is not None:
# for ext in [".data", ".meta"]:
# path = best_dev_model_path + ext
# if os.path.exists(path):
# print("Removing previous model file {}...".format(path))
# os.remove(path)
path = best_dev_model_path
if os.path.exists(path):
print("Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
# best_dev_model_path = "{}_dev={:.2f}".format(
best_dev_model_path = "{}_dev={:.2f}.pth".format(
args.model_path_base, dev_fscore.fscore)
print("Saving new best model to {}...".format(best_dev_model_path))
# dy.save(best_dev_model_path, [parser])
torch.save(parser, best_dev_model_path)
def check_performance(parser, treebank, sentence_embeddings, args):
dev_start_time = time.time()
dev_predicted = []
for tree_index, tree in enumerate(treebank):
if tree_index % 100 == 0:
dy.renew_cg()
if sentence_embeddings is not None:
elmo_embeddings = dy.inputTensor(sentence_embeddings[str(tree_index)][:, :, :])
else:
elmo_embeddings = None
sentence = [(leaf.tag, leaf.word) for leaf in tree.leaves]
predicted = parser.span_parser(sentence,
is_train=False,
elmo_embeddings=elmo_embeddings)
dev_predicted.append(predicted.convert())
dev_fscore = evaluate.evalb('EVALB', treebank, dev_predicted, args=args, name="dev")
return dev_fscore, dev_start_time
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = treebanks.load_trees(
args.test_path, args.test_path_text, args.text_processing
)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path))
parser = Parser(args.model_path, batch_size=args.batch_size)
print("Parsing test sentences...")
start_time = time.time()
if args.output_path == "-":
output_file = sys.stdout
elif args.output_path:
output_file = open(args.output_path, "w")
else:
output_file = None
test_predicted = []
for predicted_tree in parser.parse_sents(
inputs_from_treebank(test_treebank, predict_tags=args.predict_tags)
):
test_predicted.append(predicted_tree)
if output_file is not None:
print(tree.pformat(margin=1e100), file=output_file)
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank.trees, test_predicted)
print(
"test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
)
)
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
dev_start_time = time.time()
dev_predicted = run_eval(parser, dev_treebank)
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
print("dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
))
if dev_fscore.fscore > best_dev_fscore:
if best_dev_model_path is not None:
# for ext in [".data", ".meta"]:
# path = best_dev_model_path + ext
# if os.path.exists(path):
# print("Removing previous model file {}...".format(path))
# os.remove(path)
path = best_dev_model_path
if os.path.exists(path):
print("Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
# best_dev_model_path = "{}_dev={:.2f}".format(
best_dev_model_path = "{}_dev={:.2f}.pth".format(
args.model_path_base, dev_fscore.fscore)
print("Saving new best model to {}...".format(best_dev_model_path))
# dy.save(best_dev_model_path, [parser])
torch.save(parser, best_dev_model_path)
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
# model = dy.ParameterCollection()
# [parser] = dy.load(args.model_path_base, model)
parser = torch.load(args.model_path_base)
if torch.cuda.is_available():
parser = parser.cuda()
print("Parsing test sentences...")
start_time = time.time()
test_predicted = run_eval(parser, test_treebank)
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, test_predicted)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def main(args):
parser = Parser(args.grammar, args.expand_binaries)
print(
'Grammar rules:',
f'{parser.grammar.num_lexical_rules:,} lexical,',
f'{parser.grammar.num_unary_rules:,} unary,',
f'{parser.grammar.num_binary_rules:,} binary.'
)
if args.infile:
print(f'Predicting trees for tokens in `{args.infile}`.')
print(f'Writing trees to file `{args.outfile}`...')
if args.parallel:
trees = predict_from_file_parallel(
parser, args.infile, args.num_lines, args.tokenize)
else:
trees = predict_from_file(
parser, args.infile, args.num_lines, args.tokenize)
with open(args.outfile, 'w') as fout:
print('\n'.join(trees), file=fout)
if args.show:
show(args.outfile)
print('Evaluating bracket score...')
if args.goldfile:
try:
evalb(args.evalb_dir, args.outfile, args.goldfile, args.result, args.ignore_empty)
if args.show:
show(args.result)
except:
exit('Could not evaluate trees. Maybe you did not parse the entire file?')
print(f'Finished. Results saved to `{args.result}`.')
elif args.treefile:
num_trees = 10 if args.num_lines == None else args.num_lines
parses = predict_from_trees(parser, args.treefile)
fscores = []
for i in range(num_trees):
gold, pred, prec, rec, fscore = next(parses)
fscores.append(fscore)
print(f'Tree {i}, f1={fscore:.3f}.')
print()
print('Gold:')
gold.pretty_print()
print()
print('Pred:')
pred.pretty_print()
print()
print()
print('All F1 =', ' '.join([f'{fscore:.3f}' for fscore in fscores]))
print('Avg F1 = ', sum(fscores) / len(fscores))
elif args.syneval:
syneval(parser, args.syneval, args.outfile, parallel=args.parallel, short=args.short)
else:
if args.sent:
sentence = tokenize.word_tokenize(args.sent)
else:
# Demo: use a default test-sentence with gold tree.
sentence, gold = SENT.split(), GOLD
print('Parsing sentence...')
start = time.time()
tree, score = parser.parse(sentence, use_numpy=args.use_numpy)
elapsed = time.time() - start
tree.un_chomsky_normal_form()
print('Predicted.')
print()
tree.pretty_print()
print('Logprob:', score)
print()
if not args.sent:
gold = Tree.fromstring(gold)
prec, recall, fscore = parser.evalb(
gold.pformat(margin=np.inf), tree.pformat(margin=np.inf))
print('Gold.')
gold.pretty_print()
print(f'Precision = {prec:.3f}')
print(f'Recall = {recall:.3f}')
print(f'F1 = {fscore:.3f}')
print()
print(f'Parse-time: {elapsed:.3f}s.')
if args.perplexity:
perplexity = parser.perplexity(sentence)
print('Perplexity:', round(perplexity, 2))
def check_dev():
nonlocal best_dev_fscore
nonlocal best_dev_model_path
nonlocal best_dev_processed
dev_start_time = time.time()
dev_predicted = []
eval_batch_size = args.eval_batch_size
for dev_start_index in range(0, len(dev_treebank), eval_batch_size):
subbatch_trees = dev_treebank[dev_start_index:dev_start_index \
+ eval_batch_size]
subbatch_sent_ids = dev_sent_ids[dev_start_index:dev_start_index \
+ eval_batch_size]
if hparams.seg:
subbatch_txt = [tree[0] for tree in subbatch_trees]
subbatch_lbl = [tree[1] for tree in subbatch_trees]
subbatch_sentences = [[(lbl,txt) for lbl,txt in zip(sent_lbl,sent_txt)] \
for sent_lbl,sent_txt in zip(subbatch_lbl,subbatch_txt)]
else:
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in \
tree.leaves()] for tree in subbatch_trees]
subbatch_features = load_features(subbatch_sent_ids, dev_feat_dict)
predicted, _ = parser.parse_batch(subbatch_sentences, \
subbatch_sent_ids, subbatch_features)
del _
if hparams.seg:
dev_predicted.extend(predicted)
else:
dev_predicted.extend([p.convert() for p in predicted])
if hparams.seg:
dev_fscore = evaluate.seg_fscore(dev_treebank, dev_predicted)
else:
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank,
dev_predicted)
"""
with open('tmp_preds.txt','w') as f:
for pred in dev_predicted:
f.write(pred.linearize())
f.write('\n')
with open('tmp_gold.txt','w') as f:
for gold in dev_treebank:
f.write(gold.linearize())
f.write('\n')
"""
print("dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
))
sys.stdout.flush()
if dev_fscore.fscore > best_dev_fscore:
if best_dev_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_dev_model_path + ext
if os.path.exists(path):
print(
"Removing previous model file {}...".format(path))
os.remove(path)
best_dev_fscore = dev_fscore.fscore
best_dev_model_path = "{}_dev={:.2f}".format(
args.model_path_base, dev_fscore.fscore)
best_dev_processed = total_processed
print("Saving new best model to {}...".format(best_dev_model_path))
torch.save(
{
'spec': parser.spec,
'state_dict': parser.state_dict(),
'trainer': trainer.state_dict(),
}, best_dev_model_path + ".pt")
sys.stdout.flush()
print("Parsing test sentences using tensorflow...")
start_time = time.time()
test_predicted = []
for start_index in range(0, len(test_treebank), args.eval_batch_size):
# for start_index in range(0, 2, 2):
print(start_index, format_elapsed(start_time))
subbatch_trees = test_treebank[start_index:start_index+args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()] for tree in subbatch_trees]
predicted, _ = tf_parse_batch(subbatch_sentences)
del _
test_predicted.extend([p.convert() for p in predicted])
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank[:len(test_predicted)], test_predicted)
print('Done', format_elapsed(start_time))
str(test_fscore)
#%%
input_node_names = [the_inp_tokens.name.split(':')[0], the_inp_mask.name.split(':')[0]]
output_node_names = [the_out_chart.name.split(':')[0], the_out_tags.name.split(':')[0]]
print("Input node names:", input_node_names)
print("Output node names:", output_node_names)
graph_def = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, output_node_names)
#%%
def test_on_parses(args):
if not os.path.exists(args.experiment_directory):
os.mkdir(args.experiment_directory)
model = dy.ParameterCollection()
[parser] = dy.load(args.model_path_base, model)
treebank = trees.load_trees(args.input_file, strip_top=True, filter_none=True)
output = [tree.linearize() for tree in treebank]
with open(os.path.join(args.experiment_directory, 'parses.txt'), 'w') as f:
f.write('\n'.join(output))
sentence_embeddings = h5py.File(args.elmo_embeddings_file_path, 'r')
test_predicted = []
start_time = time.time()
total_log_likelihood = 0
total_confusion_matrix = {}
total_turned_off = 0
ranks = []
num_correct = 0
total = 0
for tree_index, tree in enumerate(treebank):
if tree_index % 100 == 0:
print(tree_index)
dy.renew_cg()
sentence = [(leaf.tag, leaf.word) for leaf in tree.leaves]
elmo_embeddings_np = sentence_embeddings[str(tree_index)][:, :, :]
assert elmo_embeddings_np.shape[1] == len(sentence), (
elmo_embeddings_np.shape[1], len(sentence), [word for pos, word in sentence])
elmo_embeddings = dy.inputTensor(elmo_embeddings_np)
predicted, (additional_info, c, t) = parser.span_parser(sentence, is_train=False,
elmo_embeddings=elmo_embeddings)
num_correct += c
total += t
rank = additional_info[3]
ranks.append(rank)
total_log_likelihood += additional_info[-1]
test_predicted.append(predicted.convert())
print('pos accuracy', num_correct / total)
print("total time", time.time() - start_time)
print("total loglikelihood", total_log_likelihood)
print("total turned off", total_turned_off)
print(total_confusion_matrix)
print(ranks)
print("avg", np.mean(ranks), "median", np.median(ranks))
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, test_predicted,
args=args,
erase_labels=True,
name="without-labels")
print("dev-fscore without labels", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, test_predicted,
args=args,
erase_labels=True,
flatten=True,
name="without-label-flattened")
print("dev-fscore without labels and flattened", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, test_predicted,
args=args,
erase_labels=False,
flatten=True,
name="flattened")
print("dev-fscore with labels and flattened", dev_fscore_without_labels)
test_fscore = evaluate.evalb('EVALB/', treebank, test_predicted, args=args,
name="regular")
print(
"test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
)
)
with open(os.path.join(args.experiment_directory, "confusion_matrix.pickle"), "wb") as f:
pickle.dump(total_confusion_matrix, f)
def run_test_qbank(args):
if not os.path.exists(args.experiment_directory):
os.mkdir(args.experiment_directory)
print("Loading model from {}...".format(args.model_path_base))
model = dy.ParameterCollection()
[parser] = dy.load(args.model_path_base, model)
all_trees = trees.load_trees(args.question_bank_trees_path)
if args.stanford_split == 'true':
print('using stanford split')
split_to_indices = {
'train': list(range(0, 1000)) + list(range(2000, 3000)),
'dev': list(range(1000, 1500)) + list(range(3000, 3500)),
'test': list(range(1500, 2000)) + list(range(3500, 4000))
}
else:
print('not using stanford split')
split_to_indices = {
'train': range(0, 2000),
'dev': range(2000, 3000),
'test': range(3000, 4000)
}
test_indices = split_to_indices[args.split]
qb_embeddings_file = h5py.File('../question-bank.hdf5', 'r')
dev_predicted = []
for test_index in test_indices:
if len(dev_predicted) % 100 == 0:
dy.renew_cg()
tree = all_trees[test_index]
sentence = [(leaf.tag, leaf.word) for leaf in tree.leaves]
test_embeddings_np = qb_embeddings_file[str(test_index)][:, :, :]
assert test_embeddings_np.shape[1] == len(sentence)
test_embeddings = dy.inputTensor(test_embeddings_np)
predicted, _ = parser.span_parser(sentence, is_train=False,
elmo_embeddings=test_embeddings)
dev_predicted.append(predicted.convert())
test_treebank = [all_trees[index] for index in test_indices]
dev_fscore_without_labels = evaluate.evalb(args.evalb_dir, test_treebank, dev_predicted,
args=args,
erase_labels=True,
name="without-labels")
print("dev-fscore without labels", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb(args.evalb_dir, test_treebank, dev_predicted,
args=args,
erase_labels=True,
flatten=True,
name="without-label-flattened")
print("dev-fscore without labels and flattened", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb(args.evalb_dir, test_treebank, dev_predicted,
args=args,
erase_labels=False,
flatten=True,
name="flattened")
print("dev-fscore with labels and flattened", dev_fscore_without_labels)
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, dev_predicted, args=args,
name="regular")
print("regular", test_fscore)
def run_export(args):
if args.test_path is not None:
print("Loading test trees from {}...".format(args.test_path))
test_treebank = treebanks.load_trees(
args.test_path, args.test_path_text, args.text_processing
)
print("Loaded {:,} test examples.".format(len(test_treebank)))
else:
test_treebank = None
print("Loading model from {}...".format(args.model_path))
parser = Parser(args.model_path, batch_size=args.batch_size)
model = parser._parser
if model.pretrained_model is None:
raise ValueError(
"Exporting is only defined when using a pre-trained transformer "
"encoder. For CharLSTM-based model, just distribute the pytorch "
"checkpoint directly. You may manually delete the 'optimizer' "
"field to reduce file size by discarding the optimizer state."
)
if test_treebank is not None:
print("Parsing test sentences (predicting tags)...")
start_time = time.time()
test_inputs = inputs_from_treebank(test_treebank, predict_tags=True)
test_predicted = list(parser.parse_sents(test_inputs))
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank.trees, test_predicted)
test_elapsed = format_elapsed(start_time)
print("test-fscore {} test-elapsed {}".format(test_fscore, test_elapsed))
print("Parsing test sentences (not predicting tags)...")
start_time = time.time()
test_inputs = inputs_from_treebank(test_treebank, predict_tags=False)
notags_test_predicted = list(parser.parse_sents(test_inputs))
notags_test_fscore = evaluate.evalb(
args.evalb_dir, test_treebank.trees, notags_test_predicted
)
notags_test_elapsed = format_elapsed(start_time)
print(
"test-fscore {} test-elapsed {}".format(notags_test_fscore, notags_test_elapsed)
)
print("Exporting tokenizer...")
model.retokenizer.tokenizer.save_pretrained(args.output_dir)
print("Exporting config...")
config = model.pretrained_model.config
config.benepar = model.config
config.save_pretrained(args.output_dir)
if args.compress:
print("Compressing weights...")
state_dict = get_compressed_state_dict(model.cpu())
print("Saving weights...")
else:
print("Exporting weights...")
state_dict = model.cpu().state_dict()
torch.save(state_dict, os.path.join(args.output_dir, "benepar_model.bin"))
del model, parser, state_dict
print("Loading exported model from {}...".format(args.output_dir))
exported_parser = Parser(args.output_dir, batch_size=args.batch_size)
if test_treebank is None:
print()
print("Export complete.")
print("Did not verify model accuracy because no treebank was provided.")
return
print("Parsing test sentences (predicting tags)...")
start_time = time.time()
test_inputs = inputs_from_treebank(test_treebank, predict_tags=True)
exported_predicted = list(exported_parser.parse_sents(test_inputs))
exported_fscore = evaluate.evalb(
args.evalb_dir, test_treebank.trees, exported_predicted
)
exported_elapsed = format_elapsed(start_time)
print(
"exported-fscore {} exported-elapsed {}".format(
exported_fscore, exported_elapsed
)
)
print("Parsing test sentences (not predicting tags)...")
start_time = time.time()
test_inputs = inputs_from_treebank(test_treebank, predict_tags=False)
notags_exported_predicted = list(exported_parser.parse_sents(test_inputs))
notags_exported_fscore = evaluate.evalb(
args.evalb_dir, test_treebank.trees, notags_exported_predicted
)
notags_exported_elapsed = format_elapsed(start_time)
print(
"exported-fscore {} exported-elapsed {}".format(
notags_exported_fscore, notags_exported_elapsed
)
)
print()
print("Export and verification complete.")
fscore_delta = evaluate.FScore(
recall=notags_exported_fscore.recall - notags_test_fscore.recall,
precision=notags_exported_fscore.precision - notags_test_fscore.precision,
fscore=notags_exported_fscore.fscore - notags_test_fscore.fscore,
complete_match=(
notags_exported_fscore.complete_match - notags_test_fscore.complete_match
),
tagging_accuracy=(
exported_fscore.tagging_accuracy - test_fscore.tagging_accuracy
),
)
print("delta-fscore {}".format(fscore_delta))
def check_dev(epoch_num):
nonlocal best_dev_score
nonlocal best_model_path
dev_start_time = time.time()
parser.eval()
dev_predicted = []
for dev_start_index in range(0, len(dev_treebank), args.eval_batch_size):
subbatch_trees = dev_treebank[dev_start_index:dev_start_index+args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()] for tree in subbatch_trees]
predicted, _,= parser.parse_batch(subbatch_sentences)
del _
dev_predicted.extend([p.convert() for p in predicted])
dev_fscore = evaluate.evalb(args.evalb_dir, dev_treebank, dev_predicted)
print(
"dev-fscore {} "
"dev-elapsed {} "
"total-elapsed {}".format(
dev_fscore,
format_elapsed(dev_start_time),
format_elapsed(start_time),
)
)
dev_pred_head = [[leaf.father for leaf in tree.leaves()] for tree in dev_predicted]
dev_pred_type = [[leaf.type for leaf in tree.leaves()] for tree in dev_predicted]
assert len(dev_pred_head) == len(dev_pred_type)
assert len(dev_pred_type) == len(dep_dev_type)
stats, stats_nopunc, stats_root, num_inst = dep_eval.eval(len(dev_pred_head), dep_dev_word, dep_dev_pos,
dev_pred_head, dev_pred_type,
dep_dev_headid, dep_dev_type,
dep_dev_lengs, punct_set=punct_set,
symbolic_root=False)
dev_ucorr, dev_lcorr, dev_total, dev_ucomlpete, dev_lcomplete = stats
dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc, dev_ucomlpete_nopunc, dev_lcomplete_nopunc = stats_nopunc
dev_root_corr, dev_total_root = stats_root
dev_total_inst = num_inst
print(
'W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
dev_ucorr, dev_lcorr, dev_total, dev_ucorr * 100 / dev_total, dev_lcorr * 100 / dev_total,
dev_ucomlpete * 100 / dev_total_inst, dev_lcomplete * 100 / dev_total_inst))
print(
'Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
dev_ucorr_nopunc, dev_lcorr_nopunc, dev_total_nopunc,
dev_ucorr_nopunc * 100 / dev_total_nopunc,
dev_lcorr_nopunc * 100 / dev_total_nopunc,
dev_ucomlpete_nopunc * 100 / dev_total_inst, dev_lcomplete_nopunc * 100 / dev_total_inst))
print('Root: corr: %d, total: %d, acc: %.2f%%' % (
dev_root_corr, dev_total_root, dev_root_corr * 100 / dev_total_root))
dev_uas = dev_ucorr_nopunc * 100 / dev_total_nopunc
dev_las = dev_lcorr_nopunc * 100 / dev_total_nopunc
if dev_fscore.fscore + dev_las > best_dev_score :
if best_model_path is not None:
extensions = [".pt"]
for ext in extensions:
path = best_model_path + ext
if os.path.exists(path):
print("Removing previous model file {}...".format(path))
os.remove(path)
best_dev_score = dev_fscore.fscore + dev_las
best_model_path = "{}_best_dev={:.2f}_devuas={:.2f}_devlas={:.2f}".format(
args.model_path_base, dev_fscore.fscore, dev_uas,dev_las)
print("Saving new best model to {}...".format(best_model_path))
torch.save({
'spec': parser.spec,
'state_dict': parser.state_dict(),
'trainer' : trainer.state_dict(),
}, besthh_model_path + ".pt")
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
if args.test_lbls:
test_txt = [
l.strip().split() for l in open(args.test_path, 'r').readlines()
]
test_lbls = [
l.strip().split() for l in open(args.test_lbls, 'r').readlines()
]
test_sent_ids = [
l.strip() for l in open(args.test_sent_id_path, 'r').readlines()
]
test_treebank = [(txt, lbl) for txt, lbl in zip(test_txt, test_lbls)]
else:
test_treebank, test_sent_ids = trees.load_trees_with_idx(args.test_path, \
args.test_sent_id_path, strip_top=False)
if not args.new_set:
test_pause_path = os.path.join(args.feature_path, args.test_prefix + \
'_pause.pickle')
with open(test_pause_path, 'rb') as f:
test_pause_data = pickle.load(f, encoding='latin1')
# to_remove = set(test_sent_ids).difference(set(test_pause_data.keys()))
# to_remove = sorted([test_sent_ids.index(i) for i in to_remove])
# for x in to_remove[::-1]:
# test_treebank.pop(x)
# test_sent_ids.pop(x)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(".pt"), "Only pytorch files supported"
info = torch_load(args.model_path_base)
print(info.keys())
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = parse_model.SpeechParser.from_spec(info['spec'], \
info['state_dict'])
from prettytable import PrettyTable
total_params = 0
table = PrettyTable(["Modules", "Parameters"])
for name, parameter in parser.named_parameters():
if not parameter.requires_grad: continue
param = parameter.numel()
table.add_row([name, param])
total_params += param
parser.eval() # turn off dropout at evaluation time
label_vocab = parser.label_vocab
#print("{} ({:,}): {}".format("label", label_vocab.size, \
# sorted(value for value in label_vocab.values)))
test_feat_dict = {}
if info['spec']['speech_features'] is not None:
speech_features = info['spec']['speech_features']
print("Loading speech features for test set...")
for feat_type in speech_features:
print("\t", feat_type)
feat_path = os.path.join(args.feature_path, \
args.test_prefix + '_' + feat_type + '.pickle')
with open(feat_path, 'rb') as f:
feat_data = pickle.load(f, encoding='latin1')
test_feat_dict[feat_type] = feat_data
print("Parsing test sentences...")
start_time = time.time()
test_predicted = []
test_scores = []
pscores = []
gscores = []
with torch.no_grad():
for start_index in range(0, len(test_treebank), args.eval_batch_size):
subbatch_treebank = test_treebank[start_index:start_index \
+ args.eval_batch_size]
subbatch_sent_ids = test_sent_ids[start_index:start_index \
+ args.eval_batch_size]
if args.test_lbls: # EKN using this instead of the seg flag bc it's an hparam
subbatch_txt = [turn[0] for turn in subbatch_treebank]
subbatch_lbl = [turn[1] for turn in subbatch_treebank]
subbatch_sentences = [[(lbl,txt) for lbl,txt in zip(sent_lbl,sent_txt)] for \
sent_lbl,sent_txt in zip(subbatch_lbl,subbatch_txt)]
else:
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in \
tree.leaves()] for tree in subbatch_treebank]
subbatch_trees = [t.convert() for t in subbatch_treebank]
subbatch_features = load_features(subbatch_sent_ids, test_feat_dict\
, args.sp_off)
predicted, scores = parser.parse_batch(subbatch_sentences, \
subbatch_sent_ids, subbatch_features)
if not args.get_scores:
del scores
else:
charts = parser.parse_batch(subbatch_sentences, \
subbatch_sent_ids, subbatch_features, subbatch_trees, True)
for i in range(len(charts)):
decoder_args = dict(sentence_len=len(subbatch_sentences[i]),\
label_scores_chart=charts[i],\
gold=subbatch_trees[i],\
label_vocab=parser.label_vocab, \
is_train=False, \
backoff=True)
p_score, _, _, _, _ = chart_helper.decode(
False, **decoder_args)
g_score, _, _, _, _ = chart_helper.decode(
True, **decoder_args)
pscores.append(p_score)
gscores.append(g_score)
test_scores += scores
if args.test_lbls:
test_predicted.extend(predicted)
else:
test_predicted.extend([p.convert() for p in predicted])
# DEBUG
# print(test_scores)
#print(test_score_offsets)
with open(args.output_path, 'w') as output_file:
for tree in test_predicted:
if args.test_lbls:
#import pdb;pdb.set_trace()
lbls = ' '.join(tree)
output_file.write("{}\n".format(lbls))
else:
output_file.write("{}\n".format(tree.linearize()))
print("Output written to:", args.output_path)
if args.get_scores:
with open(args.output_path + '.scores', 'w') as output_file:
for score1, score2, score3 in zip(test_scores, pscores, gscores):
output_file.write("{}\t{}\t{}\n".format(
score1, score2, score3))
print("Output scores written to:", args.output_path + '.scores')
if args.write_gold:
with open(args.test_prefix + '_sent_ids.txt', 'w') as sid_file:
for sent_id in test_sent_ids:
sid_file.write("{}\n".format(sent_id))
print("Sent ids written to:", args.test_prefix + '_sent_ids.txt')
with open(args.test_prefix + '_gold.txt', 'w') as gold_file:
for tree in test_treebank:
gold_file.write("{}\n".format(tree.linearize()))
print("Gold trees written to:", args.test_prefix + '_gold.txt')
# The tree loader does some preprocessing to the trees (e.g. stripping TOP
# symbols or SPMRL morphological features). We compare with the input file
# directly to be extra careful about not corrupting the evaluation. We also
# allow specifying a separate "raw" file for the gold trees: the inputs to
# our parser have traces removed and may have predicted tags substituted,
# and we may wish to compare against the raw gold trees to make sure we
# haven't made a mistake. As far as we can tell all of these variations give
# equivalent results.
ref_gold_path = args.test_path
if args.test_path_raw is not None:
print("Comparing with raw trees from", args.test_path_raw)
ref_gold_path = args.test_path_raw
else:
# Need this since I'm evaluating on subset
ref_gold_path = None
if args.test_lbls:
test_fscore = evaluate.seg_fscore(test_treebank,
test_predicted,
is_train=False)
else:
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, \
test_predicted, ref_gold_path=ref_gold_path, is_train=False)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def run_test(args):
const_test_path = args.consttest_ptb_path
dep_test_path = args.deptest_ptb_path
if args.dataset == 'ctb':
const_test_path = args.consttest_ctb_path
dep_test_path = args.deptest_ctb_path
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = Zparser.ChartParser.from_spec(info['spec'], info['state_dict'])
parser.eval()
dep_test_reader = CoNLLXReader(dep_test_path, parser.type_vocab)
print('Reading dependency parsing data from %s' % dep_test_path)
dep_test_data = []
test_inst = dep_test_reader.getNext()
dep_test_headid = np.zeros([40000, 300], dtype=int)
dep_test_type = []
dep_test_word = []
dep_test_pos = []
dep_test_lengs = np.zeros(40000, dtype=int)
cun = 0
while test_inst is not None:
inst_size = test_inst.length()
dep_test_lengs[cun] = inst_size
sent = test_inst.sentence
dep_test_data.append((sent.words, test_inst.postags, test_inst.heads, test_inst.types))
for i in range(inst_size):
dep_test_headid[cun][i] = test_inst.heads[i]
dep_test_type.append(test_inst.types)
dep_test_word.append(sent.words)
dep_test_pos.append(sent.postags)
# dep_sentences.append([(tag, word) for i, (word, tag) in enumerate(zip(sent.words, sent.postags))])
test_inst = dep_test_reader.getNext()
cun = cun + 1
dep_test_reader.close()
print("Loading test trees from {}...".format(const_test_path))
test_treebank = trees.load_trees(const_test_path, dep_test_headid, dep_test_type, dep_test_word)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Parsing test sentences...")
start_time = time.time()
punct_set = '.' '``' "''" ':' ','
parser.eval()
test_predicted = []
for start_index in range(0, len(test_treebank), args.eval_batch_size):
subbatch_trees = test_treebank[start_index:start_index + args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()] for tree in subbatch_trees]
predicted, _, = parser.parse_batch(subbatch_sentences)
del _
test_predicted.extend([p.convert() for p in predicted])
test_fscore = evaluate.evalb(args.evalb_dir, test_treebank, test_predicted)
print(
"test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
)
)
test_pred_head = [[leaf.father for leaf in tree.leaves()] for tree in test_predicted]
test_pred_type = [[leaf.type for leaf in tree.leaves()] for tree in test_predicted]
assert len(test_pred_head) == len(test_pred_type)
assert len(test_pred_type) == len(dep_test_type)
stats, stats_nopunc, stats_root, test_total_inst = dep_eval.eval(len(test_pred_head), dep_test_word, dep_test_pos,
test_pred_head,
test_pred_type, dep_test_headid, dep_test_type,
dep_test_lengs, punct_set=punct_set,
symbolic_root=False)
test_ucorrect, test_lcorrect, test_total, test_ucomlpete_match, test_lcomplete_match = stats
test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc, test_ucomlpete_match_nopunc, test_lcomplete_match_nopunc = stats_nopunc
test_root_correct, test_total_root = stats_root
print(
'best test W. Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%%' % (
test_ucorrect, test_lcorrect, test_total, test_ucorrect * 100 / test_total,
test_lcorrect * 100 / test_total,
test_ucomlpete_match * 100 / test_total_inst, test_lcomplete_match * 100 / test_total_inst
))
print(
'best test Wo Punct: ucorr: %d, lcorr: %d, total: %d, uas: %.2f%%, las: %.2f%%, ucm: %.2f%%, lcm: %.2f%% ' % (
test_ucorrect_nopunc, test_lcorrect_nopunc, test_total_nopunc,
test_ucorrect_nopunc * 100 / test_total_nopunc,
test_lcorrect_nopunc * 100 / test_total_nopunc,
test_ucomlpete_match_nopunc * 100 / test_total_inst,
test_lcomplete_match_nopunc * 100 / test_total_inst))
print('best test Root: corr: %d, total: %d, acc: %.2f%%' % (
test_root_correct, test_total_root, test_root_correct * 100 / test_total_root))
print(
'============================================================================================================================')
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(
".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = parse_jc.NKChartParser.from_spec(info['spec'], info['state_dict'])
if args.redo_vocab:
print(
"Loading memory bank trees from {} for generating label vocab...".
format(args.train_path))
train_treebank = trees.load_trees(args.train_path)
parser.label_vocab = gen_label_vocab(
[tree.convert() for tree in train_treebank])
print("Parsing test sentences...")
start_time = time.time()
if args.use_neighbours:
index_const = index.FaissIndex if args.library == "faiss" else index.AnnoyIndex
span_index = index_const(
num_labels=len(parser.label_vocab.values),
metric=parser.metric,
)
prefix = index.get_index_prefix(
index_base_path=args.index_path,
full_model_path=args.model_path_base,
nn_prefix=args.nn_prefix,
)
span_index.load(prefix)
# also remove relu
parser.no_relu = args.no_relu
if args.no_relu:
parser.remove_relu()
test_predicted = []
for start_index in range(0, len(test_treebank), args.eval_batch_size):
subbatch_trees = test_treebank[start_index:start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves()]
for tree in subbatch_trees]
predicted, _ = parser.parse_batch(
subbatch_sentences,
span_index=span_index if args.use_neighbours else None,
k=args.k,
zero_empty=args.zero_empty,
)
del _
test_predicted.extend([p.convert() for p in predicted])
# The tree loader does some preprocessing to the trees (e.g. stripping TOP
# symbols or SPMRL morphological features). We compare with the input file
# directly to be extra careful about not corrupting the evaluation. We also
# allow specifying a separate "raw" file for the gold trees: the inputs to
# our parser have traces removed and may have predicted tags substituted,
# and we may wish to compare against the raw gold trees to make sure we
# haven't made a mistake. As far as we can tell all of these variations give
# equivalent results.
ref_gold_path = args.test_path
if args.test_path_raw is not None:
print("Comparing with raw trees from", args.test_path_raw)
ref_gold_path = args.test_path_raw
test_fscore = evaluate.evalb(
args.evalb_dir,
test_treebank,
test_predicted,
ref_gold_path=ref_gold_path,
)
print("test-fscore {} "
"test-elapsed {}".format(
test_fscore,
format_elapsed(start_time),
))
def evaluate_on_brown_corpus(args):
if not os.path.exists(args.experiment_directory):
os.mkdir(args.experiment_directory)
model = dy.ParameterCollection()
[parser] = dy.load(args.model_path_base, model)
assert parser.use_elmo == args.use_elmo, (parser.use_elmo, args.use_elmo)
directories = ['cf', 'cg', 'ck', 'cl', 'cm', 'cn', 'cp', 'cr']
for directory in directories:
print('-' * 100)
print(directory)
input_file = '../brown/' + directory + '/' + directory + '.all.mrg'
expt_name = args.experiment_directory + '/' + directory
if not os.path.exists(expt_name):
os.mkdir(expt_name)
cleaned_corpus_path = trees.cleanup_text(input_file)
treebank = trees.load_trees(cleaned_corpus_path, strip_top=True, filter_none=True)
sentences = [[(leaf.tag, leaf.word) for leaf in tree.leaves] for tree in treebank]
tokenized_lines = [' '.join([word for pos, word in sentence]) for sentence in sentences]
if args.use_elmo:
embedding_file = compute_elmo_embeddings(tokenized_lines, expt_name)
else:
embedding_file = None
dev_predicted = []
num_correct = 0
total = 0
for tree_index, tree in enumerate(treebank):
if tree_index % 100 == 0:
print(tree_index)
dy.renew_cg()
sentence = sentences[tree_index]
if args.use_elmo:
embeddings_np = embedding_file[str(tree_index)][:, :, :]
assert embeddings_np.shape[1] == len(sentence), (
embeddings_np.shape[1], len(sentence))
embeddings = dy.inputTensor(embeddings_np)
else:
embeddings = None
predicted, (additional_info, c, t) = parser.span_parser(sentence, is_train=False,
elmo_embeddings=embeddings)
num_correct += c
total += t
dev_predicted.append(predicted.convert())
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, dev_predicted,
args=args,
erase_labels=True,
name="without-labels",
expt_name=expt_name)
print("dev-fscore without labels", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, dev_predicted,
args=args,
erase_labels=True,
flatten=True,
name="without-label-flattened",
expt_name=expt_name)
print("dev-fscore without labels and flattened", dev_fscore_without_labels)
dev_fscore_without_labels = evaluate.evalb('EVALB/', treebank, dev_predicted,
args=args,
erase_labels=False,
flatten=True,
name="flattened",
expt_name=expt_name)
print("dev-fscore with labels and flattened", dev_fscore_without_labels)
test_fscore = evaluate.evalb('EVALB/', treebank, dev_predicted, args=args,
name="regular",
expt_name=expt_name)
print("regular", test_fscore)
pos_fraction = num_correct / total
print('pos fraction', pos_fraction)
with open(expt_name + '/pos_accuracy.txt', 'w') as f:
f.write(str(pos_fraction))
def run_test(args):
print("Loading test trees from {}...".format(args.test_path))
test_treebank = trees.load_trees(args.test_path)
print("Loaded {:,} test examples.".format(len(test_treebank)))
print("Loading model from {}...".format(args.model_path_base))
assert args.model_path_base.endswith(
".pt"), "Only pytorch savefiles supported"
info = torch_load(args.model_path_base)
assert 'hparams' in info['spec'], "Older savefiles not supported"
parser = parse_nk.NKChartParser.from_spec(info['spec'], info['state_dict'])
hparams = info['spec']['hparams']
if ('use_extra_info' in hparams) and hparams['use_extra_info']:
loaded_test_info = h5py.File(args.test_path + '.hdf5', 'r')
test_info = list()
for i in range(len(test_treebank)):
item_info = list()
for key in sorted(loaded_test_info.keys()):
item_info.append(loaded_test_info[key + '/' + str(i)])
item_info = np.array(item_info)
item_info = np.concatenate([
-1e8 * np.ones(
(item_info.shape[0], item_info.shape[1], 1)), item_info
],
axis=2)
item_info = np.concatenate([
item_info, -1e8 * np.ones(
(item_info.shape[0], 1, item_info.shape[2]))
],
axis=1)
test_info.append(item_info)
print("Loaded and processed extra info from constituency test.")
else:
test_info = None
print("Parsing test sentences...")
start_time = time.time()
test_predicted = []
for start_index in range(0, len(test_treebank), args.eval_batch_size):
try:
subbatch_trees = test_treebank[start_index:start_index +
args.eval_batch_size]
subbatch_sentences = [[(leaf.tag, leaf.word)
for leaf in tree.leaves()]
for tree in subbatch_trees]
subbatch_info = test_info[
start_index:start_index +
args.eval_batch_size] if test_info is not None else None
predicted, _ = parser.parse_batch(subbatch_sentences,
extra_info=subbatch_info)
except:
from IPython import embed
embed(using=False)
del _
test_predicted.extend([p.convert() for p in predicted])
# The tree loader does some preprocessing to the trees (e.g. stripping TOP
# symbols or SPMRL morphological features). We compare with the input file
# directly to be extra careful about not corrupting the evaluation. We also
# allow specifying a separate "raw" file for the gold trees: the inputs to
# our parser have traces removed and may have predicted tags substituted,
# and we may wish to compare against the raw gold trees to make sure we
# haven't made a mistake. As far as we can tell all of these variations give
# equivalent results.
ref_gold_path = args.test_path
if args.test_path_raw is not None:
print("Comparing with raw trees from", args.test_path_raw)
ref_gold_path = args.test_path_raw
try:
test_fscore = evaluate.evalb(args.evalb_dir,
test_treebank,
test_predicted,
ref_gold_path=ref_gold_path)
print("labeled-fscore {} ".format(test_fscore))
except:
print('Failed to predict labeled score.')
for rm_punct in [True]:
for compute_level in ["corpus"]:
clean_test_treebank = [
tokens2list(
tree.linearize().replace('(',
' ( ').replace(')',
' ) ').split(),
rm_punct) for tree in test_treebank
]
clean_test_predicted = [
tokens2list(
tree.linearize().replace('(',
' ( ').replace(')',
' ) ').split(),
rm_punct) for tree in test_predicted
]
unlabeled_fscore = evaluate.evaluate_unlabeled(
clean_test_treebank, clean_test_predicted, compute_level)
print("unlabeled-fscore ({} {}) {}".format(rm_punct, compute_level,
unlabeled_fscore))
gold_bracket_num = sum(
[len(get_brackets(x)[0]) for x in clean_test_treebank])
predicted_bracket_num = sum(
[len(get_brackets(x)[0]) for x in clean_test_predicted])
print("# gold brackets: {}".format(gold_bracket_num))
print("# predicted brackets: {}".format(predicted_bracket_num))
for i, tree in enumerate(clean_test_treebank):
print(tree2str(tree))
print(tree2str(clean_test_predicted[i]))