def store_qas(dataset, qas, vocab, max_length=20):
total = len(qas)
questions = dataset.create_dataset('questions', (total, args.max_length),
dtype='i')
answers = dataset.create_dataset('answers', (total, args.max_length),
dtype='i')
categories = dataset.create_dataset('categories', (total, ), dtype='i')
image_indices = dataset.create_dataset('image_indices', (total, ),
dtype='i')
image_ids = []
bar = progressbar.ProgressBar(maxval=len(qas))
for idx, entry in enumerate(qas):
i_image = len(image_ids)
if entry['image_id'] in image_ids:
i_image = image_ids.index(entry['image_id'])
else:
image_ids.append(entry['image_id'])
image_indices[idx] = i_image
categories[idx] = entry['category']
q, length = process_text(entry['question'].encode('utf-8'),
vocab,
max_length=max_length)
questions[idx, :length] = q
a, length = process_text(entry['answer'].encode('utf-8'),
vocab,
max_length=max_length)
answers[idx, :length] = a
bar.update(idx)
return image_ids
def save_dataset(dataset, vocab, output,
max_length=20):
sentences=[]
labels=[]
with open (dataset) as questions_file:
csv_reader=csv.reader(questions_file, delimiter='\t')
line_count=0
column_names=[]
for row in csv_reader:
if line_count==0:
column_names.append(row[0])
column_names.append(row[1])
line_count+=1
else:
sentences.append(row[1])
labels.append(row[2])
total_sentences=len(sentences)
vocab=load_vocab(vocab)
print('Number of sentences to be written: %d',total_sentences)
h5file = h5py.File(output, "w")
d_questions = h5file.create_dataset(
"sentences", (total_sentences, max_length), dtype='i')
d_labels = h5file.create_dataset(
"labels", (total_sentences,), dtype='i')
bar = progressbar.ProgressBar(maxval=total_sentences)
q_index = 0
for sentence,label in zip(sentences,labels):
q, length = process_text(sentence, vocab,
max_length=20)
d_questions[q_index, :length] = q
if label == 'yesno' or label == 'factoid' or label == 'list' or label == 'summary':
x=label
else:
print('error')
l = process_label(label)
d_labels[q_index] = int(l)
q_index += 1
bar.update(q_index)
h5file.close()
def save_dataset(image_dir, questions, annotations, vocab, ans2cat, output,
im_size=224, max_q_length=20, max_a_length=4,
with_answers=False):
"""Saves the Visual Genome images and the questions in a hdf5 file.
Args:
image_dir: Directory with all the images.
questions: Location of the questions.
annotations: Location of all the annotations.
vocab: Location of the vocab file.
ans2cat: Mapping from answers to category.
output: Location of the hdf5 file to save to.
im_size: Size of image.
max_q_length: Maximum length of the questions.
max_a_length: Maximum length of the answers.
with_answers: Whether to also save the answers.
"""
# Load the data.
vocab = load_vocab(vocab)
with open(annotations) as f:
annos = json.load(f)
with open(questions) as f:
questions = json.load(f)
with open(ans2cat) as f:
ans2cat = json.load(f)
# Get the mappings from qid to answers.
qid2ans, image_ids = create_answer_mapping(annos, ans2cat)
total_questions = len(qid2ans.keys())
total_images = len(image_ids)
print "Number of images to be written: %d" % total_images
print "Number of QAs to be written: %d" % total_questions
h5file = h5py.File(output, "w")
d_questions = h5file.create_dataset(
"questions", (total_questions, max_q_length), dtype='i')
d_indices = h5file.create_dataset(
"image_indices", (total_questions,), dtype='i')
d_images = h5file.create_dataset(
"images", (total_images, im_size, im_size, 3), dtype='f')
d_answers = h5file.create_dataset(
"answers", (total_questions, max_a_length), dtype='i')
d_answer_types = h5file.create_dataset(
"answer_types", (total_questions,), dtype='i')
# Create the transforms we want to apply to every image.
transform = transforms.Compose([
transforms.Resize((im_size, im_size))])
# Iterate and save all the questions and images.
bar = progressbar.ProgressBar(maxval=total_questions)
i_index = 0
q_index = 0
done_img2idx = {}
for entry in questions['questions']:
image_id = entry['image_id']
question_id = entry['question_id']
if image_id not in image_ids:
continue
if question_id not in qid2ans:
continue
if image_id not in done_img2idx:
try:
path = "%d.jpg" % (image_id)
image = Image.open(os.path.join(image_dir, path)).convert('RGB')
except IOError:
path = "%012d.jpg" % (image_id)
image = Image.open(os.path.join(image_dir, path)).convert('RGB')
image = transform(image)
d_images[i_index, :, :, :] = np.array(image)
done_img2idx[image_id] = i_index
i_index += 1
q, length = process_text(entry['question'], vocab,
max_length=max_q_length)
d_questions[q_index, :length] = q
answer = qid2ans[question_id]
a, length = process_text(answer, vocab,
max_length=max_a_length)
d_answers[q_index, :length] = a
d_answer_types[q_index] = ans2cat[answer]
d_indices[q_index] = done_img2idx[image_id]
q_index += 1
bar.update(q_index)
h5file.close()
print "Number of images written: %d" % i_index
print "Number of QAs written: %d" % q_index
def save_dataset(image_dir, questions_path, vocab, output,
im_size=224, max_q_length=20, max_a_length=6, max_c_length=20,
with_answers=False):
"""Saves the Visual Genome images and the questions in a hdf5 file.
Args:
image_dir: Directory with all the images.
questions: Location of the questions.
annotations: Location of all the annotations.
vocab: Location of the vocab file.
output: Location of the hdf5 file to save to.
im_size: Size of image.
max_q_length: Maximum length of the questions.
max_a_length: Maximum length of the answers.
with_answers: Whether to also save the answers.
"""
# Load the data.
vocab = load_vocab(vocab)
with open(questions_path) as f:
questions = json.load(f)
# Get the mappings from qid to answers.
qid2ans, image_ids = create_questions_images_ids(questions)
total_questions = len(qid2ans)
total_images = len(image_ids)
print ("Number of images to be written: %d" % total_images)
print ("Number of QAs to be written: %d" % total_questions)
h5file = h5py.File(output, "w")
d_questions = h5file.create_dataset(
"questions", (total_questions, max_q_length), dtype='i')
d_indices = h5file.create_dataset(
"image_indices", (total_questions,), dtype='i')
d_images = h5file.create_dataset(
"images", (total_images, im_size, im_size, 3), dtype='f')
d_answers = h5file.create_dataset(
"answers", (total_questions, max_a_length), dtype='i')
# Create the transforms we want to apply to every image.
transform = transforms.Compose([
transforms.Resize((im_size, im_size))])
# Iterate and save all the questions and images.
bar = progressbar.ProgressBar(maxval=total_questions)
i_index = 0
q_index = 0
done_img2idx = {}
images_ids=[]
for entry in questions:
image_id = entry['image_id']
images=[]
images.append(image_id)
question_id = entry['question_id']
if image_id not in image_ids:
continue
if question_id not in qid2ans:
continue
if image_id not in done_img2idx:
try:
path = image_id
image = Image.open(os.path.join(image_dir, path+".jpg")).convert('RGB')
except IOError:
path = image_id
image = Image.open(os.path.join(image_dir, path)).convert('RGB')
image = transform(image)
d_images[i_index, :, :, :] = np.array(image)
done_img2idx[image_id] = i_index
i_index += 1
q, length = process_text(entry['question'], vocab,
max_length=max_q_length)
d_questions[q_index, :length] = q
a, length = process_text(entry['answer'], vocab,
max_length=max_a_length)
d_answers[q_index, :length] = a
d_indices[q_index] = done_img2idx[image_id]
images.append(int(d_indices[q_index]))
images_ids.append(images)
#print(str(d_indices[q_index]))
q_index += 1
bar.update(q_index)
h5file.close()
print ("Number of images written: %d" % i_index)
print ("Number of QAs written: %d" % q_index)
with open(os.path.join('data/vqa/', 'results_image_v.json'), 'w') as results_file:
json.dump(images_ids, results_file)