def _fix_word(i, index, annot_id, h5fn_graph, h5fn_word, pad_ind):
"""
Turn a detection list into what we want: some text, as well as some tags.
:param tokenized_sent: Tokenized sentence with detections collapsed to a list.
:param old_det_to_new_ind: Mapping of the old ID -> new ID (which will be used as the tag)
:param obj_to_type: [person, person, pottedplant] indexed by the old labels
:return: tokenized sentence
"""
#_fix_adj(i, index, annot_id, h5fn_graph, pad_ind):
##_fix_word(i, index, annot_id, h5fn_word, pad_ind):
new_tokenization_with_tags = []
with h5py.File(h5fn_graph, 'r') as h5:
grp_items = {k: np.array(v) for k, v in h5[str(index)].items()}
if annot_id != grp_items[f'annot_id']:
raise ValueError("annot_id is different!!")
node_num_list = grp_items[f'top_50_{i}']
bert_embs = np.zeros([len(node_num_list), 768])
for i, tok in enumerate(node_num_list):
new_tokenization_with_tags.append((tok, pad_ind))
with h5py.File(h5fn_word, 'r') as h5:
grp_items = {
k: np.array(v, dtype=np.float16)
for k, v in h5[str(tok)].items()
}
bert_embs[i, :] = grp_items[f'word']
text_field = BertField([Token(x[0]) for x in new_tokenization_with_tags],
bert_embs,
padding_value=0)
tags = SequenceLabelField([x[1] for x in new_tokenization_with_tags],
text_field)
return text_field, tags
def _fix_tokenization(tokenized_sent, bert_embs, old_det_to_new_ind, obj_to_type, token_indexers, pad_ind=-1):
"""
Turn a detection list into what we want: some text, as well as some tags.
:param tokenized_sent: Tokenized sentence with detections collapsed to a list.
:param old_det_to_new_ind: Mapping of the old ID -> new ID (which will be used as the tag)
:param obj_to_type: [person, person, pottedplant] indexed by the old labels
:return: tokenized sentence
"""
new_tokenization_with_tags = []
for tok in tokenized_sent:
if isinstance(tok, list):
for int_name in tok:
obj_type = obj_to_type[int_name]
new_ind = old_det_to_new_ind[int_name]
if new_ind < 0:
raise ValueError("Oh no, the new index is negative! that means it's invalid. {} {}".format(
tokenized_sent, old_det_to_new_ind
))
text_to_use = GENDER_NEUTRAL_NAMES[
new_ind % len(GENDER_NEUTRAL_NAMES)] if obj_type == 'person' else obj_type
new_tokenization_with_tags.append((text_to_use, new_ind))
else:
new_tokenization_with_tags.append((tok, pad_ind))
text_field = BertField([Token(x[0]) for x in new_tokenization_with_tags],
bert_embs,
padding_value=0)
tags = SequenceLabelField([x[1] for x in new_tokenization_with_tags], text_field)
return text_field, tags
def __call__(self, tokens, embeddings):
""" tokens: Tokenized sentence with detections collapsed to a list. """
new_tokens_with_tags = []
for tok in tokens:
if isinstance(tok, list):
for int_name in tok:
obj_type = self.obj2type[int_name]
new_ind = self.obj2ind[int_name]
assert new_ind >= 0
# 与R2C保持一致
if obj_type == 'person':
text2use = GENDER_NEUTRAL_NAMES[
(new_ind - self.obj_start_ind) %
len(GENDER_NEUTRAL_NAMES)]
else:
text2use = obj_type
new_tokens_with_tags.append((text2use, new_ind))
else:
new_tokens_with_tags.append((tok, self.pad_ind))
# 记得回来删
assert len(new_tokens_with_tags) == embeddings.shape[0]
text_field = BertField([Token(x[0]) for x in new_tokens_with_tags],
embeddings,
padding_value=0)
tags = SequenceLabelField([x[1] for x in new_tokens_with_tags],
text_field)
return text_field, tags
def _fix_visual_concept(visual_concept, visual_concept_num, h5fn, pad_ind):
"""
Turn a detection list into what we want: some text, as well as some tags.
:param tokenized_sent: Tokenized sentence with detections collapsed to a list.
:param old_det_to_new_ind: Mapping of the old ID -> new ID (which will be used as the tag)
:param obj_to_type: [person, person, pottedplant] indexed by the old labels
:return: tokenized sentence
"""
bert_embs = np.zeros([len(visual_concept),768])
new_tokenization_with_tags = []
for i,tok in enumerate(visual_concept):
new_tokenization_with_tags.append((tok, pad_ind))
with h5py.File(h5fn, 'r') as h5:
grp_items = {k: np.array(v) for k, v in h5[str(visual_concept_num[i])].items()}
bert_embs[i,:] = grp_items[f'word']
text_field = BertField([Token(x[0]) for x in new_tokenization_with_tags],
bert_embs,
padding_value=0)
tags = SequenceLabelField([x[1] for x in new_tokenization_with_tags], text_field)
return text_field, tags
