def proc_ans_tokens(self, ans, token_to_ix_ans, max_token):
ans_ix = np.zeros(max_token, np.int64)
if (self.__C.AUGMENTED_ANSWER):
question = ans['question']
augmented_ans = re.sub(
r"([.,'!?\"()*#:;])",
'',
question.lower()
).replace('-', ' ').replace('/', ' ').replace(ans['question_type'], ans['multiple_choice_answer'])
words = prep_ans(augmented_ans).split()
else :
words = prep_ans(ans['multiple_choice_answer']).split()
for ix, word in enumerate(words):
if word in token_to_ix_ans:
ans_ix[ix] = token_to_ix_ans[word]
else:
ans_ix[ix] = token_to_ix_ans['UNK']
if ix + 1 == max_token:
break
return ans_ix
def ans_stat(self, stat_ans_list, ans_freq):
ans_to_ix = {}
ix_to_ans = {}
ans_freq_dict = {}
'''
ans = {
"image_id": 393714,
"question_id": 393714000,
"question_type": "what color are the",
"answer_type": "other",
"multiple_choice_answer": "white",
"answers": [
{"answer": "light blue", "answer_confidence": "yes", "answer_id": 1},
{"answer": "white", "answer_confidence": "yes", "answer_id": 2},
{"answer": "white", "answer_confidence": "yes", "answer_id": 3},
{"answer": "white", "answer_confidence": "yes", "answer_id": 4},
{"answer": "white", "answer_confidence": "yes", "answer_id": 5},
{"answer": "white", "answer_confidence": "maybe", "answer_id": 6},
{"answer": "white", "answer_confidence": "yes", "answer_id": 7},
{"answer": "white", "answer_confidence": "yes", "answer_id": 8},
{"answer": "white", "answer_confidence": "yes", "answer_id": 9},
{"answer": "white", "answer_confidence": "yes", "answer_id": 10}
]
}
'''
for ans in stat_ans_list:
'''
To get all unique answer for each question
'''
answers = ans['answers']
unprep_ans = set([i['answer'] for i in answers])
all_answers = [prep_ans(i) for i in unprep_ans]
ans_proc = prep_ans(ans['multiple_choice_answer'])
'''
ans_proc is a string and may have multiple words
'''
if ans_proc not in ans_freq_dict:
ans_freq_dict[ans_proc] = 1
else:
ans_freq_dict[ans_proc] += 1
ans_freq_filter = ans_freq_dict.copy()
count_rejected = 0
for ans in ans_freq_dict:
if ans_freq_dict[ans] <= ans_freq:
count_rejected += ans_freq_dict[ans]
ans_freq_filter.pop(ans)
print(' ========== Unanswerable questions:', count_rejected)
for ans in ans_freq_filter:
ix_to_ans[ans_to_ix.__len__()] = ans
ans_to_ix[ans] = ans_to_ix.__len__()
return ans_to_ix, ix_to_ans
def proc_ans(self, ans, ans_to_ix):
ans_score = np.zeros(ans_to_ix.__len__(), np.float32)
ans_emb = None
flag = False # indicates whether the embeddings of most probable answer is already made
ans_prob_dict = {}
for ans_ in ans['answers']:
ans_proc = prep_ans(ans_['answer'])
if ans_proc not in ans_prob_dict:
ans_prob_dict[ans_proc] = 1
else:
ans_prob_dict[ans_proc] += 1
'''
Here we create the embedding of the most probable answer,
in future we can also take the weighted average of embeddings
of all the possible answers.
'''
if self.__C.LOSS_FUNC in ['kld']:
for ans_ in ans_prob_dict:
if ans_ in ans_to_ix:
ans_score[ans_to_ix[ans_]] = ans_prob_dict[ans_] / 10.
else:
for ans_ in ans_prob_dict:
if ans_ in ans_to_ix:
ans_score[ans_to_ix[ans_]] = self.get_score(ans_prob_dict[ans_])
return ans_score
def tokenize_ans(self, ans_list, use_glove):
token_to_ix_ans = {
'PAD': 0,
'UNK': 1,
'CLS': 2,
}
pretrained_emb_ans = []
if use_glove:
pretrained_emb_ans.append(self.spacy_tool('PAD').vector)
pretrained_emb_ans.append(self.spacy_tool('UNK').vector)
pretrained_emb_ans.append(self.spacy_tool('CLS').vector)
for ans in ans_list:
#Taking out the multiple choice/ most voted ans of the question
multi_choice_ans = ans['multiple_choice_answer']
multi_choice_ans = prep_ans(multi_choice_ans)
words = multi_choice_ans.split()
for word in words:
if word not in token_to_ix_ans:
token_to_ix_ans[word] = len(token_to_ix_ans)
if use_glove:
pretrained_emb_ans.append(self.spacy_tool(word).vector)
pretrained_emb_ans = np.array(pretrained_emb_ans)
return token_to_ix_ans, pretrained_emb_ans
def ans_stat(self, stat_ans_dict):
ans_to_ix = {}
ix_to_ans = {}
for qid in stat_ans_dict:
ans = stat_ans_dict[qid]['answer']
ans = prep_ans(ans)
if ans not in ans_to_ix:
ix_to_ans[ans_to_ix.__len__()] = ans
ans_to_ix[ans] = ans_to_ix.__len__()
return ans_to_ix, ix_to_ans
def proc_ans_tokens(self, ans, token_to_ix_ans, max_token):
ans_ix = np.zeros(max_token, np.int64)
words = prep_ans(ans['multiple_choice_answer']).split()
for ix, word in enumerate(words):
if word in token_to_ix_ans:
ans_ix[ix] = token_to_ix_ans[word]
else:
ans_ix[ix] = token_to_ix_ans['UNK']
if ix + 1 == max_token:
break
return ans_ix
def proc_ans(self, ans, ans_to_ix):
ans_score = np.zeros(ans_to_ix.__len__(), np.float32)
ans_prob_dict = {}
for ans_ in ans['answers']:
ans_proc = prep_ans(ans_['answer'])
if ans_proc not in ans_prob_dict:
ans_prob_dict[ans_proc] = 1
else:
ans_prob_dict[ans_proc] += 1
for ans_ in ans_prob_dict:
if ans_ in ans_to_ix:
ans_score[ans_to_ix[ans_]] = self.get_score(ans_prob_dict[ans_])
return ans_score
def ans_stat(stat_ans_list):
ans_to_ix = {}
ix_to_ans = {}
ans_freq_dict = {}
for ans in stat_ans_list:
ans_proc = prep_ans(ans['multiple_choice_answer'])
if ans_proc not in ans_freq_dict:
ans_freq_dict[ans_proc] = 1
else:
ans_freq_dict[ans_proc] += 1
ans_freq_filter = ans_freq_dict.copy()
for ans in ans_freq_dict:
if ans_freq_dict[ans] <= 8:
ans_freq_filter.pop(ans)
for ans in ans_freq_filter:
ix_to_ans[ans_to_ix.__len__()] = ans
ans_to_ix[ans] = ans_to_ix.__len__()
return ans_to_ix, ix_to_ans
def proc_ans(self, ans, ans_to_ix):
ans_ix = np.zeros(1, np.int64)
ans = prep_ans(ans)
ans_ix[0] = ans_to_ix[ans]
return ans_ix
