def get_tokens(self, sentences):
if not isinstance(sentences, list):
sentences = [sentences]
final_sentences = []
for _, sentence in enumerate(sentences):
tokens = tokenize(sentence)
final_sentences.append(tokens)
return final_sentences
def get_imdb(ann_path: str, quest_path: str, split: str,
answer_vocab_path: str) -> np.ndarray:
imdb = [{"dataset_name": "okvqa"}]
with PathManager.open(answer_vocab_path, "r") as f:
answer_vocab = set(f.read().splitlines())
with PathManager.open(ann_path, "r") as f:
annotations = json.load(f)["annotations"]
with PathManager.open(quest_path, "r") as f:
questions = json.load(f)["questions"]
gt_answers = {}
for ann in annotations:
gt_answers[ann["question_id"]] = ann["answers"]
count = 0
for quest in tqdm(questions):
image_name = f"COCO_{split}_{quest['image_id']:012d}"
q_id = quest["question_id"]
all_answers = [item['answer'] for item in gt_answers[q_id]]
answers = [ans for ans in all_answers if ans in answer_vocab]
if len(answers) == 0:
answers = ["<unk>"]
count += 1
entry = {
"image_name": image_name,
"image_id": quest["image_id"],
"feature_path": f"{image_name}.npy",
"question_id": q_id,
"question_str": quest["question"],
"question_tokens": tokenize(quest["question"]),
"answers": answers,
"all_answers": all_answers,
}
imdb.append(entry)
print("Unknown questions:", count)
return np.array(imdb)
def extract(self):
os.makedirs(self.out_dir, exist_ok=True)
word_count = Counter()
texts = self.get_text()
text_lengths = [None] * len(texts)
for inx, text in enumerate(texts):
words = tokenize(text)
text_lengths[inx] = len(words)
word_count.update(words)
# UNK token will added on fly if you use Vocab class in core/text
vocabulary = [w[0] for w in word_count.items() if w[1] >= self.min_freq]
vocabulary.sort()
self.save_vocabulary(vocabulary)
print("min text len=", min(text_lengths))
print("max text len=", max(text_lengths))
def __getitem__(self, idx):
data = self.questions[idx]
# Each call to __getitem__ from dataloader returns a Sample class object which
# collated by our special batch collator to a SampleList which is basically
# a attribute based batch in layman terms
current_sample = Sample()
question = data["question"]
tokens = tokenize(question, keep=[";", ","], remove=["?", "."])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
processed = self.answer_processor({"answers": [data["answer"]]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
image_path = os.path.join(self.image_path, data["image_filename"])
image = np.true_divide(Image.open(image_path).convert("RGB"), 255)
image = image.astype(np.float32)
current_sample.image = torch.from_numpy(image.transpose(2, 0, 1))
return current_sample
def test_tokenize(self):
tokens = text_utils.tokenize(self.TOKENIZE_EXAMPLE)
self.assertEqual(list(tokens), self.TOKENS)
