def build_cartesian_domain_sent_corpora(self):
"""
aggregate descriptive sentences by their labels.
"""
des_pattern_stripped = re.compile(self.DES_PATTERN_STRIPPED)
sent_pattern = re.compile(self.SENT_PATTERN_TAGGED) # reserve start and end
dom_sents_dict = dict()
src_fns = [fn for fn in listdir(self.dp_dataset_test_filtered)
if isfile(join(self.dp_dataset_test_filtered, fn))]
for src_fn in tqdm(src_fns):
label = '_'.join(src_fn.split('_')[1:])
src_fp = join(self.dp_dataset_test_filtered, src_fn)
with io.open(src_fp, encoding='utf-8') as src_f:
text_stripped = re.findall(des_pattern_stripped, src_f.read())[0]
sents_stripped = re.findall(sent_pattern, text_stripped)
if label in dom_sents_dict:
dom_sents_dict[label].extend(sents_stripped)
else:
dom_sents_dict[label] = sents_stripped
for dom, sents in dom_sents_dict.items():
logger.info('{0} sents: {1}'.format(dom, str(len(sents))))
out_fp = join(self.dp_dataset_dom_sent_corpora, dom)
with io.open(out_fp, mode='a', encoding='utf-8') as out_f:
out_f.write('\n'.join(sents))
def _build_components(cid, query):
sim_items = tfidf_tools.build_sim_items_e2e(cid,
query,
mask_intra=mask_intra)
rel_scores = sim_items['rel_scores']
sim_mat = sim_items['doc_sim_mat']
processed_sents = sim_items['processed_sents']
rel_vec = rel_scores / np.sum(rel_scores) # l1 norm to make a distribution
np.fill_diagonal(sim_mat, 0.0) # avoid self-transition
logger.info('sim_mat: {}'.format(sim_mat))
sid2abs = {}
sid_abs = 0
for doc_idx, doc in enumerate(processed_sents):
for sent_idx, sent in enumerate(doc):
sid = config.SEP.join((str(doc_idx), str(sent_idx)))
sid2abs[sid] = sid_abs
sid_abs += 1
components = {
'sim_mat': sim_mat,
'rel_vec': rel_vec,
'sid2abs': sid2abs,
}
return components
def compute_rouge_for_oracle():
"""
The rec dp for oracle saves text for comparing against refecence.
:return:
"""
ir_rec_dp = join(path_parser.summary_rank,
ir_config.IR_RECORDS_DIR_NAME_TF)
if exists(ir_rec_dp):
raise ValueError('ir_rec_dp exists: {}'.format(ir_rec_dp))
os.mkdir(ir_rec_dp)
cids = tools.get_test_cc_ids()
for cid in tqdm(cids):
retrieval_params = {
'model_name': ir_config.IR_MODEL_NAME_TF,
'cid': cid,
'filter_var': ir_config.FILTER_VAR,
'filter': ir_config.FILTER,
'deduplicate': ir_config.DEDUPLICATE,
'prune': True,
}
retrieved_items = ir_tools.retrieve(**retrieval_params)
summary = '\n'.join([item[-1] for item in retrieved_items])
with open(join(ir_rec_dp, cid), mode='a', encoding='utf-8') as out_f:
out_f.write(summary)
performance = rouge.compute_rouge_for_ablation_study(ir_rec_dp)
logger.info(performance)
def init():
# parse args
parser = ArgumentParser()
parser.add_argument(
'n_devices',
nargs='?',
default=4,
help='num of devices on which model will be running on')
args = parser.parse_args()
all_device_ids = [0, 1, 2, 3]
device = all_device_ids[:int(args.n_devices)]
config_meta['device'] = device
if not torch.cuda.is_available():
placement = 'cpu'
logger.info('[MAIN INIT] path mode: {0}, placement: {1}'.format(
config.path_type, placement))
else:
if len(device) == 1:
placement = 'single'
torch.cuda.set_device(device[0])
elif config_meta['auto_parallel']:
placement = 'auto'
else:
placement = 'manual'
logger.info(
'[MAIN INIT] path mode: {0}, placement: {1}, n_devices: {2}'.
format(config.path_type, placement, args.n_devices))
config_meta['placement'] = placement
def _compute_rel_scores_tf_dot(processed_sents, query):
"""
:param processed_sents:
:param query_sents:
:param mask_intra:
:return:
"""
doc_sents = list(itertools.chain(*processed_sents)) # 1d sent list
doc_sents = copy.deepcopy(
doc_sents) # avoid affecting the original doc_sents list
# logger.info('doc_sents: {}'.format(len(doc_sents)))
doc_sents.append(query)
tf_mat = get_tf_mat(sents=doc_sents).toarray()
# logger.info('tf_idf_mat: {}'.format(tf_idf_mat))
# doc_sent_mat = tf_mat[:-1].A
# query_mat = tf_mat[-1].A.reshape(-1, 1)
# logger.info('doc_sent_mat: {}, query_mat: {}'.format(doc_sent_mat.shape, query_mat.shape))
# logger.info('doc_sent_mat: {}'.format(type(doc_sent_mat)))
rel_scores = np.matmul(tf_mat[:-1], tf_mat[-1])
# rel_scores = np.squeeze(doc_query_sim_mat, axis=-1)
logger.info('rel_scores: {}'.format(rel_scores.shape))
return rel_scores
def approx_rand(self):
obs_diff = self._obs_diff()
def _iter(iter_idx):
logger.info('iter_idx: {0}'.format(iter_idx))
reserve = np.random.randint(2, size=[self.n_obs, 1])
reserve_mat = np.tile(reserve, [1, n_ways])
flip_mat = 1 - reserve_mat
x_stat = np.multiply(self.obs_1, reserve_mat) + np.multiply(
self.obs_2, flip_mat)
y_stat = np.multiply(self.obs_2, reserve_mat) + np.multiply(
self.obs_1, flip_mat)
diff = self._diff(x_stat, y_stat)
# logger.info('diff: {0}, x_stat: {1}, y_stat: {2}'.format(diff, x_stat, y_stat))
if diff >= obs_diff:
return 1
else:
return 0
pool = ThreadPool(6)
better = float(sum(pool.map(_iter, range(self.iter))))
self.p = (better + 1) / (self.iter + 1)
logger.info('p-value: {0}'.format(self.p))
def get_doc_by_head(self, head, use_head_id=False):
"""
get a doc as a string by its head (default) or head id (set use_head_id = True).
"""
if lang != 'en' or use_head_id:
head_id = head
else:
head_id = self.get_head_id(head)
if not head_id:
return None
if lang == 'en':
file_id = self.get_file_id(head_id)
fp_wiki_file = self.fp_wiki_files[file_id]
else:
fp_wiki_file = self.fp_wiki_file
with io.open(fp_wiki_file, encoding='utf-8') as f:
data = mmap.mmap(f.fileno(), 0, access=mmap.PROT_READ)
pattern = re.compile(
self.DOC_PATTERN.format(head_id).encode('utf-8'))
match = re.search(pattern, data)
if match:
logger.info('succeeded to extract doc: {0}'.format(head_id))
return match.group().decode('utf-8')
else:
logger.info('failed to extract doc: {0}'.format(head_id))
return None
def _proc_sent(self,
sent,
rm_dialog,
rm_stop,
stem,
rm_short=None,
min_nw_sent=3):
sent = sent.lower()
sent = re.sub(r'\s+', ' ', sent).strip() # remove extra spaces
if not sent:
return None
if rm_short and len(nltk.tokenize.word_tokenize(sent)) < min_nw_sent:
return None
if rm_dialog:
dialog_tokens = ["''", "``"]
for tk in dialog_tokens:
if tk in sent:
logger.info('Remove dialog')
return None
if config.test_year == '2005' and sent[0] == "'" and (
'says' in sent or 'said' in sent):
logger.info('Remove dialog')
return None
if rm_stop:
sent = remove_stopwords(sent)
if stem:
sent = self.porter_stemmer.stem_sentence(sent)
return sent
def compute_rouge(model_name,
n_iter=None,
diversity_param_tuple=None,
cos_threshold=None,
extra=None):
rouge_args = '-a -l 250 -n 2 -m -2 4 -u -c 95 -r 1000 -f A -p 0.5 -t 0 -d -e {} -x'.format(
path_parser.rouge_dir)
r = Rouge155(rouge_args=rouge_args)
baselines_wo_config = ['lead', 'lead-2006', 'lead-2007', 'lead_2007']
if model_name in baselines_wo_config or model_name.startswith('duc'):
text_dp = join(path_parser.summary_text, model_name)
else:
text_dp = tools.get_text_dp(
model_name,
cos_threshold=cos_threshold,
n_iter=n_iter,
diversity_param_tuple=diversity_param_tuple,
extra=extra)
r.system_dir = text_dp
r.model_dir = join(path_parser.data_summary_targets, config.test_year)
gen_sys_file_pat = '(\w*)'
gen_model_file_pat = '#ID#_[\d]'
r.system_filename_pattern = gen_sys_file_pat
r.model_filename_pattern = gen_model_file_pat
output = r.convert_and_evaluate()
output = proc_output(output)
logger.info(output)
return output
def rank_e2e():
"""
:param pool_func: avg, max, or None (for integrated query).
:return:
"""
rank_dp = join(path_parser.summary_rank, ir_config.IR_MODEL_NAME_TF)
test_cid_query_dicts = general_tools.build_test_cid_query_dicts(
tokenize_narr=False,
concat_title_narr=ir_config.CONCAT_TITLE_NARR,
query_type=ir_config.QUERY_TYPE)
if exists(rank_dp):
raise ValueError('rank_dp exists: {}'.format(rank_dp))
os.mkdir(rank_dp)
for cid_query_dict in tqdm(test_cid_query_dicts):
params = {
**cid_query_dict,
}
rank_records = _rank(**params)
rank_sent.dump_rank_records(rank_records,
out_fp=join(rank_dp, params['cid']),
with_rank_idx=False)
logger.info('Successfully dumped rankings to: {}'.format(rank_dp))
def check_cross_du_2(self):
dom_ids_list = list()
unique_ids = list()
for dom in doms_final:
id_fp = os.path.join(self.dp_doc_dedu, dom + '_ids.txt')
with io.open(id_fp, encoding='utf-8') as id_f:
dom_ids = [l.rstrip('\n') for l in id_f.readlines() if l]
dom_ids_list.append(dom_ids)
unique_ids.extend(
[dom_id for dom_id in dom_ids if dom_id not in unique_ids])
n_label_list = [0] * len(doms_final)
for unique_id in unique_ids:
n_labels = 0
for dom_ids in dom_ids_list:
if unique_id in dom_ids:
n_labels += 1
n_label_list[n_labels] += 1
n_label_ratio = [
'{0:.2f}'.format(float(n_label) / sum(n_label_list))
for n_label in n_label_list
]
logger.info('#labels: {}'.format(n_label_list))
logger.info('ratio of #labels: {}'.format(n_label_ratio))
def get_head_id(self, head):
if head in self.head_to_id:
return self.head_to_id[head]
# return self.heads.index(head) + 1 # index from 1
else:
logger.info('doc out of list: {0}'.format(head))
return None
def __init__(self):
self.fp_non_ge_ids = path_parser.data_non_ge_ids
self.fp_clean_stats = path_parser.data_clean_stats
# join: doc & doms (8 classes)
self.dp_doc = path_parser.data_doc
self.dom_dp = dict()
for dom in doms:
self.dom_dp[dom] = os.path.join(self.dp_doc, dom)
# join: doc_dedu & doms_final (7 classes)
self.dp_doc_dedu = path_parser.data_doc_dedu
self.dom_dedu_dp = dict()
for dom in doms_final:
self.dom_dedu_dp[dom] = os.path.join(self.dp_doc_dedu, dom)
self.dp_proj_root = path_parser.proj_root
self.SUBCATEGORIES = 'subcategories'
self.PAGES = 'pages'
if lang == 'en':
self.START_DOC, self.END_DOC = '#s-doc\t{0}', '#e-doc\t{0}'
self.START_SEC, self.END_SEC = 's-headline', '#e-headline'
self.DOC_PATTERN = '#s-doc\t{0}[\s\S]*#e-doc\t{0}'
self.SEC_PATTERN = '#s-headline[\s\S]*?(?=#s-headline)'
self.SEC_PATTERN_LAST = '#s-headline\t{0}[\s\S]*?(?=#e-doc)'
self.fp_wiki_corpus_head = path_parser.wiki_corpus_head
self.fp_wiki_files = list()
for i in range(24):
if i < 10:
i = '0{0}'.format(i)
self.fp_wiki_files.append(
os.path.join(path_parser.wiki_corpus,
'wikipedia-tagged_{}.txt'.format(i)))
# self.heads = list()
self.id_to_head, self.head_to_id = dict(), dict()
with io.open(self.fp_wiki_corpus_head, encoding='utf-8') as f:
for line in f:
items = line.split('\t')
if len(items) >= 2:
head_id, head = int(items[0]), items[1]
self.id_to_head[head_id] = head
self.head_to_id[head] = head_id
# self.heads.append(head)
logger.info('#articles: {0}'.format(len(self.id_to_head)))
else:
self.START_DOC, self.END_DOC = '<article name="{0}">', '</article>'
self.START_PARAGRAPH, self.END_PARAGRAPH = '<p>', '</p>'
self.START_HEADLINE, self.END_HEADLINE = '<h>', '</h>'
self.DOC_PATTERN = self.START_DOC + '[\s\S]*?' + self.END_DOC
self.fp_wiki_file = path_parser.wiki_corpus
# multi-threads
self.n_extraction_threads = 5 # 10
self.n_file_threads = 5 # 6
def check_cross_du(self):
for target_dom in doms_final:
doms_final_copy = list(doms_final)
doms_final_copy.remove(target_dom)
other_dom_ids = list()
for dom in doms_final_copy:
id_fp = os.path.join(self.dp_doc_dedu, dom + '_ids.txt')
with io.open(id_fp, encoding='utf-8') as id_f:
ids = [l.rstrip('\n') for l in id_f.readlines() if l]
other_dom_ids.extend(ids)
target_id_fp = os.path.join(self.dp_doc_dedu,
target_dom + '_ids.txt')
with io.open(target_id_fp, encoding='utf-8') as target_id_f:
target_ids = [
l.rstrip('\n') for l in target_id_f.readlines() if l
]
du_ids = [
target_id for target_id in target_ids
if target_id in other_dom_ids
]
ratio = float(len(du_ids)) / len(target_ids)
logger.info('{0}: {1}/{2}, {3:.2f}'.format(
target_dom, len(du_ids), len(target_ids), ratio))
def __init__(self,
dataset_type,
transform=None,
collect_doc_ids=False,
doc_ids=None,
in_para_only=False):
super(DomDetDataset, self).__init__()
self.doms = doms_final
self.score_func = config_model['score_func']
self.dataset_parser = DatasetParser()
self.dataset_type = dataset_type
assert dataset_type in ('train', 'dev', 'test', 'test_filtered')
if mode == 'debug':
dataset_type = 'debug'
self.root = path_parser.dataset_type_dp[dataset_type]
self.fns = [
fn for fn in listdir(self.root) if isfile(join(self.root, fn))
]
if doc_ids: # only one sample
logger.info('Load only {} target docs'.format(len(doc_ids)))
self.fns = [
fn for fn in self.fns if tools.get_doc_id(fn) in doc_ids
]
self.transform = transform
self.collect_doc_ids = collect_doc_ids
self.in_para_only = in_para_only
def compute_exam_p_and_r(y_true_3d, y_pred_3d, n_sents, n_words, silent=False):
"""
return a list of precision values.
:param y_true_4d: d_batch * max_n_sents * max_n_words * 1
:param y_pred_4d: d_batch * max_n_sents * max_n_words * 1
:param n_sents: d_batch * 1
:param n_words: d_batch * max_n_sents * 1
"""
p_list = list()
r_list = list()
d_batch = y_true_3d.shape[0]
for sample_idx in range(d_batch):
n_sent = n_sents[sample_idx, 0]
for sent_idx in range(n_sent):
n_word = n_words[sample_idx, sent_idx]
y_true = y_true_3d[sample_idx, sent_idx, :n_word]
y_pred = y_pred_3d[sample_idx, sent_idx, :n_word]
if not silent and not y_pred.any() == 1:
logger.info('No pred is made for: {0}.{1}. y_pred: {2}'.format(
sample_idx, sent_idx, y_pred))
p_list.append(precision_score(y_true, y_pred))
r_list.append(recall_score(y_true, y_pred))
return p_list, r_list
def get_wiki_summ_text_dp(order_subdir,
accurate_only,
multi_label_only,
variation=None):
if not variation:
variation = config_model['variation']
text_root = join(path_parser.summary_text_dps['wiki'], variation)
if not exists(text_root):
logger.info("Saving root does not exist. Making directory {}".format(
text_root))
mkdir(text_root)
if accurate_only:
subdir = '{}_accurate'.format(order_subdir)
else:
subdir = '{}_all'.format(order_subdir)
if multi_label_only:
subdir = '{}_multilabel'.format(subdir)
summ_text_dp = join(text_root, subdir)
if not exists(summ_text_dp):
logger.info("Saving root does not exist. Making directory {}".format(
summ_text_dp))
mkdir(summ_text_dp)
return summ_text_dp
def build_rel_scores_tf_passage(cid,
query,
retrieved_dp=None,
tdqfs_data=False):
_, proc_passages, passage_ids = load_retrieved_passages(
cid,
get_sents=False,
retrieved_dp=retrieved_dp,
passage_ids=None,
tdqfs_data=tdqfs_data)
# passage_ids, passage_fps = get_passage_fps(cid, retrieved_dp=retrieved_dp)
# #
# # proc_passages = []
# # passage_ids = []
# for fp in passage_fps:
# # po = load_obj(fp)
# with open(fp, 'rb') as f:
# po = dill.load(f)
# # print('po: {}, type(po): {}'.format(po, type(po)))
# # passage_ids.append(po.pid)
#
# passage = po.get_proc_passage()
# proc_passages.append(passage)
# # logger.info('{}: {}'.format(po.pid, passage))
rel_scores = _compute_rel_scores_tf(
[proc_passages], query
) # nest proc_passages again for compatibility; todo: double check the nest level
logger.info('rel_scores: {}'.format(rel_scores))
pid2score = {}
for pid, score in zip(passage_ids, rel_scores):
pid2score[pid] = score
return pid2score
def _build_components(cid, query):
sim_items = tfidf_tools.build_sim_items_e2e_tfidf_with_lexrank(cid, query, rm_dialog=RM_DIALOG)
sim_mat = vec_tools.norm_sim_mat(sim_mat=sim_items['doc_sim_mat'], max_min_scale=False)
rel_vec = vec_tools.norm_rel_scores(rel_scores=sim_items['rel_scores'], max_min_scale=False)
logger.info('rel_vec: {}'.format(rel_vec))
if len(rel_vec) != len(sim_mat):
raise ValueError('Incompatible sim_mat size: {} and rel_vec size: {} for cid: {}'.format(
sim_mat.shape, rel_vec.shape, cid))
processed_sents = sim_items['processed_sents']
sid2abs = {}
sid_abs = 0
for doc_idx, doc in enumerate(processed_sents):
for sent_idx, sent in enumerate(doc):
sid = config.SEP.join((str(doc_idx), str(sent_idx)))
sid2abs[sid] = sid_abs
sid_abs += 1
components = {
'sim_mat': sim_mat,
'rel_vec': rel_vec,
'sid2abs': sid2abs,
}
return components
def _passage_core(cid, query, narr, passage_size, stride):
original_sents, processed_sents = dataset_parser.cid2sents(cid, max_ns_doc=None) # 2d lists, docs => sents
logger.info('#doc: {}'.format(len(original_sents)))
# build sent_objs
sent_objs = [] # organized by doc
sent_idx = 0
for doc_idx in range(len(original_sents)):
sent_objs_doc = []
for original_s, proc_s in zip(original_sents[doc_idx], processed_sents[doc_idx]):
sid = config.SEP.join([cid, str(sent_idx)])
so = SentObj(sid=sid, original_sent=original_s, proc_sent=proc_s)
sent_objs_doc.append(so)
sent_idx += 1
sent_objs.append(sent_objs_doc)
# build passage objs
passage_objs = []
for sent_objs_doc in sent_objs:
start = 0
# make sure the last sentence whose length < stride will be discarded
while start + stride < len(sent_objs_doc):
pid = config.SEP.join([cid, str(len(passage_objs))])
target_sent_objs = sent_objs_doc[start:start+passage_size]
po = PassageObj(pid=pid, query=query, narr=narr, sent_objs=target_sent_objs)
passage_objs.append(po)
start += stride
return passage_objs
def compute_rouge():
tg2recall_all = {}
tg2f1_all = {}
for idx in range(N_REFS):
build_eval_dirs(summary_index=idx + 1)
tg2recall, tg2f1 = compute_rouge_for_human(system_dp=SYSTEM_DP_TEMP,
model_dp=MODEL_DP_TEMP)
logger.info('tg2f1: {}'.format(tg2f1))
for metric in ROUGE_METRICS:
if metric in tg2recall_all:
tg2recall_all[metric] += tg2recall[metric]
else:
tg2recall_all[metric] = tg2recall[metric]
if metric in tg2f1_all:
tg2f1_all[metric] += tg2f1[metric]
else:
tg2f1_all[metric] = tg2f1[metric]
for metric in ROUGE_METRICS:
tg2recall_all[metric] /= N_REFS
tg2f1_all[metric] /= N_REFS
recall_str = 'Recall:\t{}'.format('\t'.join(
['{0:.2f}'.format(tg2recall_all[metric]) for metric in ROUGE_METRICS]))
f1_str = 'F1:\t{}'.format('\t'.join(
['{0:.2f}'.format(tg2f1_all[metric]) for metric in ROUGE_METRICS]))
output = '\n' + '\n'.join((f1_str, recall_str))
logger.info(output)
def build_tdqfs_oracle_test_cid_query_dicts(query_fp):
def _get_ref(cid):
REF_DP = path_parser.data_tdqfs_summary_targets
fp = join(REF_DP, '{}_{}'.format(cid, 0))
ref = ''
# for fn in fns:
lines = io.open(fp, encoding='utf-8').readlines()
for line in lines:
ref += line.rstrip('\n')
return ref
assert config_meta['test_year'] == 'tdqfs'
lines = io.open(query_fp).readlines()
cids = [line.rstrip('\n').split('\t')[0] for line in lines]
test_cid_query_dicts = []
for cid in cids:
ref = _get_ref(cid)
logger.info('cid {}: {}'.format(cid, ref))
test_cid_query_dicts.append({
'cid': cid,
'query': ref,
})
return test_cid_query_dicts
def test_major_sent(synthese):
# logger.info('START: model testing...')
dataset_type = 'synthese' if synthese else 'lead'
data_loader = pipe.DomDetDataLoader(dataset_type=dataset_type)
n_iter, total_loss = 0, 0.0
n_samples, total_hamming = 0, 0.0
cf_mats, precision_list, recall_list = list(), list(), list()
for batch_idx, batch in enumerate(data_loader):
n_iter += 1
y_true = batch['labels'].cpu().numpy()
d_batch = len(y_true)
des_sent_info = batch['des_sent_info'].cpu().numpy()
n_samples += np.sum(des_sent_info[:, -1])
# logger.info('batch_size: {0}'.format(y_true.shape[0]))
if synthese:
hyp_scores = np.tile(y_pred_vec, (d_batch, 1))
fids = batch['fids'].cpu().numpy()
eval_args = {
'hyp_scores': hyp_scores,
'fids': fids,
'is_hiernet': True
}
eval_res = metrics.metric_eval_for_syn_doc(**eval_args)
else:
hyp_scores = np.tile(y_pred_vec, (d_batch, max_n_sents, 1))
eval_args = {
'y_true': y_true,
'hyp_scores': hyp_scores,
'des_sent_info': des_sent_info,
}
eval_res = metrics.metric_eval(**eval_args)
cf_mats.append(eval_res['cf_mat_list'])
precision_list.extend(eval_res['precision_list'])
recall_list.extend(eval_res['recall_list'])
total_hamming += eval_res['hamming']
cls_f1, avg_f1 = metrics.compute_f1_with_confusion_mats(cf_mats)
example_based_f1 = metrics.compute_example_based_f1(
precision_list=precision_list, recall_list=recall_list)
hamming = total_hamming / n_samples
eval_log_info = {
'example_based_f1': example_based_f1,
'avg_f1': avg_f1,
'cls_f1': cls_f1,
'hamming': hamming,
}
res_str = 'example_based_f1: {example_based_f1:.6f},' \
'avg_f1: {avg_f1:.6f}, cls_f1: {cls_f1}, hamming: {hamming:.6f}'
logger.info(res_str.format(**eval_log_info))
def dump_retrieval(fp, retrieved_items):
retrieve_records = ['\t'.join(items) for items in retrieved_items]
n_sents = rank_sent.dump_rank_records(rank_records=retrieve_records,
out_fp=fp,
with_rank_idx=False)
logger.info('successfully dumped {0} retrieved items to {1}'.format(
n_sents, fp))
def validate(data_type, window):
dump_fp = join(path_parser.squad_proc, 'window_{}'.format(window),
'{}.tsv'.format(data_type))
with io.open(dump_fp, encoding='utf-8') as dump_f:
line = dump_f.readline()
if len(line.split('\t')) != 5:
raise ValueError('Invalid line: {}'.format(line))
logger.info('Validated!')
def sample_human_eval_fns(self, max_dom_doc=2):
selected_fns = []
dict_dom_n_files = dict()
for dom in self.doms_final:
dict_dom_n_files[dom] = 0
root_path = path_parser.dataset_test
target_path = path_parser.dataset_test_human_eval
fns = [fn for fn in listdir(root_path) if
not fn.startswith('.') and len(fn.split('_')) >= 2 and isfile(join(root_path, fn))]
logger.info('#files: {}'.format(len(fns)))
existing_fns = [fn for fn in listdir(target_path) if
not fn.startswith('.') and len(fn.split('_')) >= 2 and isfile(join(target_path, fn))]
for fn in existing_fns:
labels = fn.split('_')[1:]
# logger.info('labels: {}'.format(labels))
for label in labels:
dict_dom_n_files[label] += 1
for k, v in dict_dom_n_files.items():
logger.info('Existing: {0} - {1}'.format(k, v))
if min(dict_dom_n_files.values()) == max_dom_doc:
logger.error('Already full!')
raise FileExistsError
random.shuffle(fns)
full_status = False
for fn in fns:
labels = fn.split('_')[1:]
# logger.info('labels: {0}'.format(labels))
ahead_overflow = False
for label in labels:
if dict_dom_n_files[label] == max_dom_doc:
ahead_overflow = True
break
if not ahead_overflow and self.check_existence_of_paras(fp=join(root_path, fn)):
selected_fns.append(fn)
for label in labels:
dict_dom_n_files[label] += 1
if min(dict_dom_n_files.values()) == max_dom_doc:
full_status = True
if full_status:
break
for fn in selected_fns:
old_path = join(root_path, fn)
new_path = join(target_path, fn)
sp.call(['cp', old_path, new_path])
def __init__(self, model_out_fp_1, model_out_fp_2, gold_fp, iter=10000):
self.obs_1 = get_obs(fp=model_out_fp_1)
self.obs_2 = get_obs(fp=model_out_fp_2)
self.gold = get_obs(fp=gold_fp)
assert self.obs_1.shape == self.obs_2.shape == self.gold.shape
self.n_obs = self.obs_1.shape[0]
logger.info('n_obs: {0}'.format(self.n_obs))
self.iter = iter
def _prune_rank_items(rank_items, threshold=1e-10):
if float(rank_items[-1][1]) > threshold:
logger.info('Prune ratio: 0.00')
return rank_items
for i in range(len(rank_items)):
if float(rank_items[i][1]) <= threshold:
logger.info('Prune ratio: {0:.2f}'.format(
float(i) / len(rank_items)))
return rank_items[:i]
def get_bert_in_func():
if config.meta_model_name == 'bert_qa':
from frame.bert_qa import bert_input
bert_in_func = bert_input.build_bert_x
else:
from data import bert_input_sep
bert_in_func = bert_input_sep.build_bert_x_sep
logger.info('Using bert_in_func: {}'.format(config.meta_model_name))
return bert_in_func
def _dump_passages(year, cid, passage_objs):
cc_dp = join(path_parser.data_passages, year, cid)
if not exists(cc_dp): # remove previous output
os.mkdir(cc_dp)
for po in passage_objs:
with open(join(cc_dp, po.pid), 'wb') as f:
dill.dump(po, f)
logger.info('[_dump_passages] Dump {} passage objects to {}'.format(len(passage_objs), cc_dp))
