File size: 7,798 Bytes
3e1d9f3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 |
from ..root import (
DATASETS,
QUESTION_PLACEHOLDER,
IMAGE_PLACEHOLDER,
BOXES_PLACEHOLDER,
)
from ..utils import MInstrDataset
def prepare_sentence(sent):
ret_str = []
ret_box_seq = []
for word in sent:
if isinstance(word, list):
ret_str.append(BOXES_PLACEHOLDER)
ret_box_seq.append(word)
else:
ret_str.append(word)
return " ".join(ret_str), ret_box_seq
def prepare_choice(pack_choices, label_index, *, options='ABCDEFG'):
ret_str = []
ret_box_seq = []
for pack, op in zip(pack_choices, options):
ret_str.append(f"({op}) {pack[0]}")
ret_box_seq.extend(pack[1])
ret_pack = (" ".join(ret_str), ret_box_seq)
label_choice = f"The answer is ({options[label_index]})."
return ret_pack, (label_choice, [])
def merge(packs, *, prefixs, postfixs=None):
if postfixs is None:
postfixs = ['' for _ in range(len(packs))]
assert len(packs) == len(prefixs) == len(postfixs), f"{len(packs)},{len(prefixs)},{len(postfixs)}"
ret_str = []
ret_box_seq = []
for pack, prefix, postfix in zip(packs, prefixs, postfixs):
if prefix:
ret_str.append(prefix)
ret_str.append(pack[0])
if postfix:
ret_str.append(postfix)
ret_box_seq.extend(pack[1])
return " ".join(ret_str), ret_box_seq
@DATASETS.register_module()
class VCRDataset(MInstrDataset):
def __init__(self, *args, version, **kwargs):
super().__init__(*args, **kwargs, placeholders=(IMAGE_PLACEHOLDER, QUESTION_PLACEHOLDER))
self.version = version
assert version in [
'q-a', 'q-ra',
'qc-a', 'qc-ra', 'qc-rac', # for evaluation: A
'qa-r', 'q-a-q-r',
'qac-r', 'qc-a-qc-r', # for evaluation: R
]
# for evaluation:
# A: 'qc-a' 'qc-ra' 'qc-rac'
# R: 'qac-r' 'qc-a-qc-r'
def __getitem__(self, index, force_answer_label=None, force_rationale_label=None):
item = self.get_raw_item(index)
image = self.get_image(item['img_fn'])
boxes_with_prob = item['boxes']
boxes = [box[:4] for box in boxes_with_prob]
question = item['question']
answer_choices = item['answer_choices']
rationale_choices = item['rationale_choices']
if force_answer_label is not None:
answer_label = force_answer_label
else:
answer_label = item['answer_label']
if force_rationale_label is not None:
rationale_label = force_rationale_label
else:
rationale_label = item['rationale_label']
question_pack = prepare_sentence(question)
answer_pack_choices = [prepare_sentence(_) for _ in answer_choices]
rationale_pack_choices = [prepare_sentence(_) for _ in rationale_choices]
answer_choices_pack, answer_choice = prepare_choice(answer_pack_choices, answer_label)
rationale_choices_pack, rationale_choice = prepare_choice(rationale_pack_choices, rationale_label)
answer_gold_pack = answer_pack_choices[answer_label]
rationale_gold_pack = rationale_pack_choices[rationale_label]
version = self.version
if version == 'q-a':
final_packs = [
merge([question_pack], prefixs=['QUESTION:'], ),
answer_gold_pack,
]
elif version == 'q-ra':
final_packs = [
merge([question_pack], prefixs=['QUESTION:'], ),
merge([rationale_gold_pack, answer_gold_pack], prefixs=['', '']),
]
elif version == 'qc-a':
final_packs = [
merge([question_pack, answer_choices_pack], prefixs=['QUESTION:', '\nOPTIONS:'], postfixs=['', 'You should decide on the best choice and output the corresponding letter.']),
answer_choice,
]
elif version == 'qc-ra':
final_packs = [
merge([question_pack, answer_choices_pack], prefixs=['QUESTION:', '\nOPTIONS:'], postfixs=['', 'You should decide on the best choice and output the corresponding letter.']),
merge([rationale_gold_pack, answer_choice], prefixs=['', '']),
]
elif version == 'qc-rac':
final_packs = [
merge([question_pack, answer_choices_pack], prefixs=['QUESTION:', '\nOPTIONS:'], postfixs=['', 'You should decide on the best choice and output the corresponding letter.']),
merge([rationale_gold_pack, answer_gold_pack, answer_choice], prefixs=['', '', '']),
]
elif version == 'qa-r':
final_packs = [
merge([question_pack, answer_gold_pack], prefixs=['QUESTION:', '\nANSWER:'], postfixs=['', 'You should explain the reason for the above answer.']),
rationale_gold_pack,
]
elif version == 'qac-r':
final_packs = [
merge([question_pack, answer_gold_pack, rationale_choices_pack], prefixs=['QUESTION:', '\nANSWER:', '\nRATIONALE OPTIONS:'], postfixs=['', '', 'You should decide on the best choice that explains the above answer and output the corresponding letter.']),
rationale_choice,
]
elif version == 'q-a-q-r':
final_packs = [
merge([question_pack], prefixs=['QUESTION:'], ),
answer_gold_pack,
('You should explain the reason for the above answer.', ()),
rationale_gold_pack,
]
elif version == 'qc-a-qc-r':
final_packs = [
merge([question_pack, answer_choices_pack], prefixs=['QUESTION:', '\nOPTIONS:'], postfixs=['', 'You should decide on the best choice and output the corresponding letter.']),
answer_choice,
merge([rationale_choices_pack], prefixs=['RATIONALE OPTIONS:'], postfixs=['You should decide on the best choice that explains the above answer and output the corresponding letter.']),
rationale_choice,
]
else:
assert False
conversations = []
roles = ['human', 'gpt']
for idx, pack in enumerate(final_packs):
conversations.append({
'from': roles[idx % 2],
'value': pack[0],
'boxes_seq': pack[1],
})
conversations[0]['value'] = self.get_template().replace(QUESTION_PLACEHOLDER, conversations[0]['value'])
ret = {
'image': image,
'target': {'boxes': boxes},
'conversations': conversations,
}
return ret
@DATASETS.register_module()
class VCRPredDataset(VCRDataset):
def __init__(self, *args, version, **kwargs):
super().__init__(*args, version=version, **kwargs)
assert version in [
'qc-a', 'qc-ra', 'qc-rac', # for evaluation: A
'qac-r', 'qc-a-qc-r', # for evaluation: R
]
self.is_pred_for_r = version in [
'qac-r', 'qc-a-qc-r', # for evaluation: R
]
def __len__(self):
if self.is_pred_for_r:
return super().__len__() * 4
else:
return super().__len__()
# noinspection PyMethodOverriding
def __getitem__(self, index):
if self.is_pred_for_r:
item_index = index // 4
answer_index = index % 4
ret = super().__getitem__(item_index, force_answer_label=answer_index, force_rationale_label=0)
else:
ret = super().__getitem__(index, force_answer_label=0, force_rationale_label=0)
ret['conversations'][-1]['value'] += "WARNING: answer and rationale here are just placeholders. we have no real anno."
return ret
|