def vqa_processing(image_dir, annotation_file, question_file, valid_answer_set,
image_set):
print('building vqa %s dataset' % image_set) #image_set对应具体的数据
if image_set in ['train2014', 'val2014']: #存在答案
load_answer = True
with open(annotation_file % image_set) as f:
annotations = json.load(f)[
'annotations'] #列表，每个元素是字典，对应一条Image_QA实例
#建立问题编号与解释字典间的字典
qid2ann_dict = {ann['question_id']: ann for ann in annotations}
else:
load_answer = False
with open(question_file % image_set) as f:
questions = json.load(f)['questions'] #列表。每个元素为字典，对应每个Image_Q
coco_set_name = image_set.replace('-dev', '')
#绝对路径
abs_image_dir = os.path.abspath(
image_dir % coco_set_name) #test-dev2015的Q仍对应图像文件夹test2015
image_name_template = 'COCO_' + coco_set_name + '_%012d' #图像名字模板
dataset = [None] * len(questions) #全部问题数目
unk_ans_count = 0
for n_q, q in enumerate(questions): #对于每一问题
if (n_q + 1) % 10000 == 0:
print('processing %d / %d' % (n_q + 1, len(questions)))
image_id = q['image_id'] #与问题对应的图像id
question_id = q['question_id'] #问题id
image_name = image_name_template % image_id #图像文件名
image_path = os.path.join(abs_image_dir, image_name + '.jpg') #图像路径
question_str = q['question'] #问题文本
question_tokens = text_processing.tokenize(question_str) #返回句子切词列表
#构建信息字典
iminfo = dict(image_name=image_name,
image_path=image_path,
question_id=question_id,
question_str=question_str,
question_tokens=question_tokens)
if load_answer: #如果有回答
ann = qid2ann_dict[question_id] #通过问题id对应到解释字典
#抽取回答集合
all_answers, valid_answers = extract_answers(
ann['answers'], valid_answer_set)
if len(valid_answers) == 0:
valid_answers = ['<unk>']
unk_ans_count += 1 #没有有效回答的例子数目
iminfo['all_answers'] = all_answers #往信息表中添加答案信息
iminfo['valid_answers'] = valid_answers
dataset[n_q] = iminfo #根据问题文件序列存储信息
print('total %d out of %d answers are <unk>' %
(unk_ans_count, len(questions)))
return dataset #返回整合后的数据集，列表型，元素为字典
def vqa_processing(image_dir, annotation_file, question_file, valid_answer_set,
image_set):
print('building vqa %s dataset' % image_set)
if image_set in ['train2014', 'val2014']:
load_answer = True
with open(annotation_file % image_set) as f:
annotations = json.load(f)['annotations']
qid2ann_dict = {ann['question_id']: ann for ann in annotations}
else:
load_answer = False
with open(question_file % image_set) as f:
questions = json.load(f)['questions']
coco_set_name = image_set.replace('-dev', '')
abs_image_dir = os.path.abspath(image_dir % coco_set_name)
image_name_template = 'COCO_' + coco_set_name + '_%012d'
dataset = [None] * len(questions)
unk_ans_count = 0
for n_q, q in enumerate(questions):
if (n_q + 1) % 10000 == 0:
print('processing %d / %d' % (n_q + 1, len(questions)))
image_id = q['image_id']
question_id = q['question_id']
image_name = image_name_template % image_id
image_path = os.path.join(abs_image_dir, image_name + '.jpg')
question_str = q['question']
question_tokens = text_helper.tokenize(question_str)
iminfo = dict(image_name=image_name,
image_path=image_path,
question_id=question_id,
question_str=question_str,
question_tokens=question_tokens)
if load_answer:
ann = qid2ann_dict[question_id]
all_answers, valid_answers = extract_answers(
ann['answers'], valid_answer_set)
if len(valid_answers) == 0:
valid_answers = ['<unk>']
unk_ans_count += 1
iminfo['all_answers'] = all_answers
iminfo['valid_answers'] = valid_answers
dataset[n_q] = iminfo
print('total %d out of %d answers are <unk>' %
(unk_ans_count, len(questions)))
return dataset
def vqa_processing(image_dir, annotation_file, question_file, valid_answer_set,
image_set):
if image_set in ['train2014', 'val2014']: # 存在答案
load_answer = True
with open(annotation_file % image_set) as f:
annotations = json.load(f)['annotations']
'''
a list
annotations[i] = {
'question_type',
'multiple_choice_answer',
'answers': [
{'answer',
'answer_confidence',
'answer_id'},...,{}
]
'image_id',
'answer_type',
'question_id'
}
'''
# 建立问题编号与解释字典间的字典
qid2ann_dict = {ann['question_id']: ann for ann in annotations}
print(
f'annotations in {annotation_file.split("/")[-1:][0] % image_set} loaded.'
)
else:
load_answer = False
with open(question_file % image_set) as f:
questions = json.load(f)['questions'] #列表。每个元素为字典，对应每个Image_Q
'''
a list
questions[i] = {
'image_id',
'question',
'question_id'
}
'''
print(
f'questions of {question_file.split("/")[-1:][0] % image_set} loaded.'
)
coco_set_name = image_set.replace('-dev', '')
# get all questions in variable `question` & all annotations in `annotations`
# 绝对路径
abs_image_dir = os.path.abspath(
image_dir % coco_set_name) # test-dev2015的Questions仍对应图像文件夹test2015
image_name_template = 'COCO_' + coco_set_name + '_%012d' # 图像名字模板
dataset = [None] * len(questions) # 全部问题数目
unk_ans_count = 0 # 没有有效回答的例子数目
for n_q, q in enumerate(questions): #对于每一问题
if (n_q + 1) % 10000 == 0:
print(f'processing {round(10000*(n_q+1)/len(questions))/100}%')
image_id = q['image_id'] # 与问题对应的图像id
question_id = q['question_id'] # 问题id
image_name = image_name_template % image_id # 图像文件名
image_path = os.path.join(abs_image_dir, image_name +
'.jpg') # 图像路径 # all in .jpg format
question_str = q['question'] # 问题文本
question_tokens = tokenize(question_str) # 返回句子切词列表
#构建信息字典
iminfo = dict(image_name=image_name,
image_path=image_path,
question_id=question_id,
question_str=question_str,
question_tokens=question_tokens)
if load_answer: # 如果有回答
ann = qid2ann_dict[question_id] # 通过问题id对应到解释字典
# 抽取回答集合
all_answers, valid_answers = extract_answers(
ann['answers'], valid_answer_set)
if len(valid_answers) == 0:
valid_answers = ['<unk>']
unk_ans_count += 1 # 没有有效回答的例子数目+1
iminfo['all_answers'] = all_answers # 往信息表中添加答案信息
iminfo['valid_answers'] = valid_answers
dataset[n_q] = iminfo # 根据问题文件序列存储信息
print(
f'in {image_set}: total {unk_ans_count} out of {len(questions)} answers are <unk>\n'
)
return dataset # 返回整合后的数据集，列表型，元素为字典