EuroPython2022/latr-vqa

1# import random2import torch3import math4from torch.nn.utils.rnn import pad_sequence5 6 7def find_pad_idx(boxes):8  for i, j in enumerate(boxes):9    if int(boxes[i].sum().item()) == 0:10      return i11  return i12 13 14 15# def apply_mask_on_token_bbox(boxes, tokenized_words, only_actual_words = False, span = 4, proportion_to_mask = 0.15, special_token = 103):16  17  # '''18  # code taken from here: https://www.geeksforgeeks.org/python-non-overlapping-random-ranges/19 20  # Note: A more robust solution is to be coded 21  # '''22  # length_to_be_masked = int(proportion_to_mask*len(boxes))23 24  # if only_actual_words:25  #   tot = find_pad_idx(tokenized_words)26  # else:27  #   tot = len(boxes)28  29  # res = set()30  # for _ in range(length_to_be_masked):31  #   temp = random.randint(0, tot - span) 32  #   while any(((temp >= idx) and (temp <= idx + span)) for idx in res):33  #     temp = random.randint(0, tot - span) 34  #   res.add(temp)35 36  #   ## Applying the mask on token37  #   tokenized_words[temp] = special_token38 39  #   ## Applying the masking on the box40  #   boxes[temp, 0] = torch.min(boxes[temp: temp+span, 0])41  #   boxes[temp, 1] = torch.min(boxes[temp: temp+span, 1])42  #   boxes[temp, 2] = torch.max(boxes[temp: temp+span, 2])43  #   boxes[temp, 3] = torch.max(boxes[temp: temp+span, 3])44  #   boxes[temp, 4] = boxes[temp, 2] - boxes[temp, 0]45  #   boxes[temp, 5] = boxes[temp, 3] - boxes[temp, 1]46 47  # return res,boxes, tokenized_words48 49 50def convert_ans_to_token(answer, label2id, max_seq_length = 512 ):51 52  ## Simple Trick to pad a sequence to deired length53  dummy_array = torch.zeros(max_seq_length)54  actual_ans_array = []55 56  answer = answer.split(" ")57  for token in answer:58    actual_ans_array.append(label2id[token]['id'])59  60  actual_ans_array = torch.tensor(actual_ans_array, dtype = torch.int32)61  actual_ans_array = pad_sequence([actual_ans_array,dummy_array], batch_first  = True)[0]62 63  return actual_ans_array64 65 66def convert_ques_to_token(question, tokenizer, pad_token_id = 0, max_seq_len = 512):67 68  question_array = []69  question = question.split(" ")70  71  for token in question:72    question_array.extend(tokenizer(token, add_special_tokens = False).input_ids)73  74  if len(question_array)< max_seq_len:75        question_array.extend([pad_token_id]* (max_seq_len-len(question_array)))76 77  question_array = torch.tensor(question_array, dtype = torch.int32)78  return question_array[:max_seq_len]79 80 81## To be taken from here82## https://logicatcore.github.io/scratchpad/lidar/sensor-fusion/jupyter/2021/04/20/3D-Oriented-Bounding-Box.html83 84def rotate(origin, point, angle):85    """86    Rotate a point counterclockwise by a given angle around a given origin.87    The angle should be given in radians.88    89    modified from answer here: https://stackoverflow.com/questions/34372480/rotate-point-about-another-point-in-degrees-python90    """91    # angle = np.deg2rad(angle)92    ox, oy = origin93    px, py = point94 95    qx = ox + math.cos(angle) * (px - ox) - math.sin(angle) * (py - oy)96    qy = oy + math.sin(angle) * (px - ox) + math.cos(angle) * (py - oy)97    return int(qx), int(qy)98 99 100def convert_token_to_ques(ques, tokenizer):101  decoded_ques = tokenizer.decode(ques, skip_special_tokens=True)102  return decoded_ques103 104 105def convert_token_to_answer(ans, id2label):106  non_zero_argument = torch.nonzero(ans,as_tuple = False).view(-1)107 108  actual_answer = ans[non_zero_argument].cpu().numpy()109  decoded_answer = []110  111  for token in actual_answer:112    decoded_answer.append(id2label[token])113  114  decoded_answer = " ".join(decoded_answer)115  return decoded_answer116 117