def generate(self, qids) -> Iterator[SRPerQuery]:
for qid in qids:
query_tokens = self.resource.get_q_tokens(qid)
content_len = self.max_seq_length - 3 - len(query_tokens)
try:
docs: List[MSMarcoDoc] = load_per_query_docs(qid, None)
docs_d = {d.doc_id: d for d in docs}
sr_per_query_doc_list = []
for doc_id in self.resource.get_doc_for_query_d()[qid]:
label = self.resource.get_label(qid, doc_id)
try:
doc = docs_d[doc_id]
segs: List[SegmentRepresentation] = self.get_segs(
query_tokens, doc, content_len)
sr_per_query_doc = SRPerQueryDoc(doc_id, segs, label)
sr_per_query_doc_list.append(sr_per_query_doc)
except KeyError:
pass
sr_per_query = SRPerQuery(qid, sr_per_query_doc_list)
yield sr_per_query
except KeyError as e:
print(e)
print(doc_id)
except FileNotFoundError as e:
print(e)
print(qid)
def work(self, job_id):
qid_list = self.query_group[job_id]
missing_rel_cnt = 0
missing_nrel_cnt = 0
def empty_doc_fn(query_id, doc_id):
rel_docs = self.ms_reader.qrel[query_id]
nonlocal missing_rel_cnt
nonlocal missing_nrel_cnt
if doc_id in rel_docs:
missing_rel_cnt += 1
else:
missing_nrel_cnt += 1
for qid in qid_list:
docs: List[MSMarcoDoc] = load_per_query_docs(qid, empty_doc_fn)
if qid not in self.candidate_docs_d:
continue
target_docs = self.candidate_docs_d[qid]
tokens_d = {}
for d in docs:
if d.doc_id in target_docs:
tokens_d[d.doc_id] = []
if len(tokens_d) < len(target_docs):
log_variables(job_id, qid, tokens_d, target_docs)
not_found_docs = list([
doc_id for doc_id in target_docs if doc_id not in tokens_d
])
print("{} of {} not found: {}".format(len(not_found_docs),
len(target_docs),
not_found_docs))
def work(self, job_id):
qid_list = self.query_group[job_id]
ticker = TimeEstimator(len(qid_list))
missing_rel_cnt = 0
missing_nrel_cnt = 0
def empty_doc_fn(query_id, doc_id):
rel_docs = self.ms_reader.qrel[query_id]
nonlocal missing_rel_cnt
nonlocal missing_nrel_cnt
if doc_id in rel_docs:
missing_rel_cnt += 1
else:
missing_nrel_cnt += 1
for qid in qid_list:
if qid not in self.candidate_docs_d:
continue
docs: List[MSMarcoDoc] = load_per_query_docs(qid, empty_doc_fn)
ticker.tick()
target_docs = self.candidate_docs_d[qid]
text_d = {}
bert_tokens_d = {}
stemmed_tokens_d = {}
for d in docs:
if d.doc_id in target_docs:
title = d.title
title = crop_to_space(title, self.max_title_length)
body_sents = sent_tokenize(d.body)
new_body_sents = self.resplit_body_sents(body_sents)
text_d[d.doc_id] = title, new_body_sents
for tokenize_fn, save_dict in [
(self.bert_tokenizer.tokenize, bert_tokens_d),
(self.stem_tokenizer.tokenize_stem, stemmed_tokens_d)
]:
title_tokens = tokenize_fn(title)
body_tokens_list = lmap(tokenize_fn, new_body_sents)
save_dict[d.doc_id] = (title_tokens, body_tokens_list)
todo = [
(text_d, self.text_dir_name),
(bert_tokens_d, self.bert_tokens_dir_name),
(stemmed_tokens_d, self.stemmed_tokens_dir_name),
]
for tokens_d, dir_name in todo:
save_path = os.path.join(self.out_dir, dir_name, str(qid))
pickle.dump(tokens_d, open(save_path, "wb"))
def get_todo() -> List[Tuple[QueryID, MSMarcoDoc]]:
print("get_todo()")
doc_queries = load_train_queries()
doc_qrels: Dict[QueryID, List[str]] = load_msmarco_raw_qrels("train")
todo: List[Tuple[QueryID, MSMarcoDoc]] = []
doc_id_to_find = []
n_item = 1000
for qid, q_text in doc_queries[:n_item]:
docs = load_per_query_docs(qid, None)
for doc in docs:
if doc.doc_id in doc_qrels[qid]:
todo.append((qid, doc))
doc_id_to_find.append(doc.doc_id)
return todo
def generate(self, data_id_manager, qids):
missing_cnt = 0
success_docs = 0
missing_doc_qid = []
ticker = TimeEstimator(len(qids))
for qid in qids:
if qid not in self.resource.get_doc_for_query_d():
continue
ticker.tick()
docs: List[MSMarcoDoc] = load_per_query_docs(qid, None)
docs_d = {d.doc_id: d for d in docs}
q_tokens = self.resource.get_q_tokens(qid)
pos_doc_id_list, neg_doc_id_list \
= get_pos_neg_doc_ids_for_qid(self.resource, qid)
def iter_passages(doc_id):
doc = docs_d[doc_id]
insts: List[Tuple[List, List]] = self.encoder.encode(q_tokens, doc.title, doc.body)
for passage_idx, passage in enumerate(insts):
yield passage
for pos_doc_id in pos_doc_id_list:
sampled_neg_doc_id = pick1(neg_doc_id_list)
try:
for passage_idx1, passage1 in enumerate(iter_passages(pos_doc_id)):
for passage_idx2, passage2 in enumerate(iter_passages(sampled_neg_doc_id)):
tokens_seg1, seg_ids1 = passage1
tokens_seg2, seg_ids2 = passage2
data_id = data_id_manager.assign({
'doc_id1': pos_doc_id,
'passage_idx1': passage_idx1,
'doc_id2': sampled_neg_doc_id,
'passage_idx2': passage_idx2,
})
inst = PairedInstance(tokens_seg1, seg_ids1, tokens_seg2, seg_ids2, data_id)
yield inst
success_docs += 1
except KeyError:
missing_cnt += 1
missing_doc_qid.append(qid)
if missing_cnt > 10:
print(missing_doc_qid)
print("success: ", success_docs)
raise KeyError
def main():
split = "dev"
query_d = dict(load_queries(split))
bm25_module = get_bm25_module()
ranked_list_path = at_working_dir("msmarco-doc{}-top100".format(split))
run_name = "BM25_df100"
rlg = load_ranked_list_grouped(ranked_list_path)
save_path = at_output_dir("ranked_list", "mmd_dev_{}.txt".format(run_name))
te = TimeEstimator(100)
out_entries = []
for query_id, entries in rlg.items():
doc_ids = list([e.doc_id for e in entries])
docs = load_per_query_docs(query_id, None)
found_doc_ids = list([d.doc_id for d in docs])
not_found_doc_ids = list(
[doc_id for doc_id in doc_ids if doc_id not in found_doc_ids])
doc_id_len = len(not_found_doc_ids)
if doc_id_len:
print("{} docs not found".format(doc_id_len))
query_text = query_d[QueryID(query_id)]
def score(doc: MSMarcoDoc):
content = doc.title + " " + doc.body
return bm25_module.score(query_text, content)
scored_docs = list([(d, score(d)) for d in docs])
scored_docs.sort(key=get_second, reverse=True)
reranked_entries = []
for rank, (doc, score) in enumerate(scored_docs):
e = TrecRankedListEntry(query_id, doc.doc_id, rank, score,
run_name)
reranked_entries.append(e)
out_entries.extend(reranked_entries)
te.tick()
if len(out_entries) > 100 * 100:
break
write_trec_ranked_list_entry(out_entries, save_path)
def work(self, job_id):
qid_list = self.query_group[job_id]
ticker = TimeEstimator(len(qid_list))
missing_rel_cnt = 0
missing_nrel_cnt = 0
def empty_doc_fn(query_id, doc_id):
rel_docs = self.ms_reader.qrel[query_id]
nonlocal missing_rel_cnt
nonlocal missing_nrel_cnt
if doc_id in rel_docs:
missing_rel_cnt += 1
else:
missing_nrel_cnt += 1
def get_tf(text):
tokens = self.tokenizer.tokenize_stem(text)
return Counter(tokens)
for qid in qid_list:
if qid not in self.candidate_docs_d:
continue
docs: List[MSMarcoDoc] = load_per_query_docs(qid, empty_doc_fn)
ticker.tick()
target_docs = self.candidate_docs_d[qid]
tokens_d = {}
for d in docs:
if d.doc_id in target_docs:
title_tokens = self.tokenizer.tokenize_stem(d.title)
body_sents = sent_tokenize(d.body)
body_tf_list = lmap(get_tf, body_sents)
tokens_d[d.doc_id] = (title_tokens, body_tf_list)
if len(tokens_d) < len(target_docs):
log_variables(job_id, qid)
print("{} of {} not found".format(len(tokens_d),
len(target_docs)))
save_path = os.path.join(self.out_dir, str(qid))
pickle.dump(tokens_d, open(save_path, "wb"))
def generate(self, data_id_manager, qids):
missing_cnt = 0
success_docs = 0
missing_doc_qid = []
ticker = TimeEstimator(len(qids))
for qid in qids:
if qid not in self.resource.get_doc_for_query_d():
continue
ticker.tick()
docs: List[MSMarcoDoc] = load_per_query_docs(qid, None)
docs_d = {d.doc_id: d for d in docs}
q_tokens = self.resource.get_q_tokens(qid)
for doc_id in self.resource.get_doc_for_query_d()[qid]:
label = self.resource.get_label(qid, doc_id)
try:
doc = docs_d[doc_id]
insts: Iterable[Tuple[List, List]] = self.encoder.encode(q_tokens, doc.title, doc.body)
for passage_idx, passage in enumerate(insts):
tokens_seg, seg_ids = passage
assert type(tokens_seg[0]) == str
assert type(seg_ids[0]) == int
data_id = data_id_manager.assign({
'doc_id': doc_id,
'passage_idx': passage_idx,
'label': label,
})
inst = ClassificationInstanceWDataID(tokens_seg, seg_ids, label, data_id)
yield inst
success_docs += 1
except KeyError:
missing_cnt += 1
missing_doc_qid.append(qid)
if missing_cnt > 10:
print(missing_doc_qid)
print("success: ", success_docs)
raise KeyError
