车位角点检测代码
選択できるのは25トピックまでです。 トピックは、先頭が英数字で、英数字とダッシュ('-')を使用した35文字以内のものにしてください。

100 行
4.4KB

  1. """Train directional marking point detector."""
  2. import math
  3. import random
  4. import torch
  5. from torch.utils.data import DataLoader
  6. import config
  7. import data
  8. import util
  9. from model import DirectionalPointDetector
  10. def plot_prediction(logger, image, marking_points, prediction):
  11. """Plot the ground truth and prediction of a random sample in a batch."""
  12. rand_sample = random.randint(0, image.size(0)-1)
  13. sampled_image = util.tensor2im(image[rand_sample])
  14. logger.plot_marking_points(sampled_image, marking_points[rand_sample],
  15. win_name='gt_marking_points')
  16. sampled_image = util.tensor2im(image[rand_sample])
  17. pred_points = data.get_predicted_points(prediction[rand_sample], 0.01)
  18. if pred_points:
  19. logger.plot_marking_points(sampled_image,
  20. list(list(zip(*pred_points))[1]),
  21. win_name='pred_marking_points')
  22. def generate_objective(marking_points_batch, device):
  23. """Get regression objective and gradient for directional point detector."""
  24. batch_size = len(marking_points_batch)
  25. objective = torch.zeros(batch_size, config.NUM_FEATURE_MAP_CHANNEL,
  26. config.FEATURE_MAP_SIZE, config.FEATURE_MAP_SIZE,
  27. device=device)
  28. gradient = torch.zeros_like(objective)
  29. gradient[:, 0].fill_(1.)
  30. for batch_idx, marking_points in enumerate(marking_points_batch):
  31. for marking_point in marking_points:
  32. col = math.floor(marking_point.x * config.FEATURE_MAP_SIZE)
  33. row = math.floor(marking_point.y * config.FEATURE_MAP_SIZE)
  34. # Confidence Regression
  35. objective[batch_idx, 0, row, col] = 1.
  36. # Makring Point Shape Regression
  37. objective[batch_idx, 1, row, col] = marking_point.shape
  38. # Offset Regression
  39. objective[batch_idx, 2, row, col] = marking_point.x*16 - col
  40. objective[batch_idx, 3, row, col] = marking_point.y*16 - row
  41. # Direction Regression
  42. direction = marking_point.direction
  43. objective[batch_idx, 4, row, col] = math.cos(direction)
  44. objective[batch_idx, 5, row, col] = math.sin(direction)
  45. # Assign Gradient
  46. gradient[batch_idx, 1:6, row, col].fill_(1.)
  47. return objective, gradient
  48. def train_detector(args):
  49. """Train directional point detector."""
  50. args.cuda = not args.disable_cuda and torch.cuda.is_available()
  51. device = torch.device('cuda:' + str(args.gpu_id) if args.cuda else 'cpu')
  52. torch.set_grad_enabled(True)
  53. dp_detector = DirectionalPointDetector(
  54. 3, args.depth_factor, config.NUM_FEATURE_MAP_CHANNEL).to(device)
  55. if args.detector_weights:
  56. print("Loading weights: %s" % args.detector_weights)
  57. dp_detector.load_state_dict(torch.load(args.detector_weights))
  58. dp_detector.train()
  59. optimizer = torch.optim.Adam(dp_detector.parameters(), lr=args.lr)
  60. if args.optimizer_weights:
  61. print("Loading weights: %s" % args.optimizer_weights)
  62. optimizer.load_state_dict(torch.load(args.optimizer_weights))
  63. logger = util.Logger(args.enable_visdom, ['train_loss'])
  64. data_loader = DataLoader(data.ParkingSlotDataset(args.dataset_directory),
  65. batch_size=args.batch_size, shuffle=True,
  66. num_workers=args.data_loading_workers,
  67. collate_fn=lambda x: list(zip(*x)))
  68. for epoch_idx in range(args.num_epochs):
  69. for iter_idx, (images, marking_points) in enumerate(data_loader):
  70. images = torch.stack(images).to(device)
  71. optimizer.zero_grad()
  72. prediction = dp_detector(images)
  73. objective, gradient = generate_objective(marking_points, device)
  74. loss = (prediction - objective) ** 2
  75. loss.backward(gradient)
  76. optimizer.step()
  77. train_loss = torch.sum(loss*gradient).item() / loss.size(0)
  78. logger.log(epoch=epoch_idx, iter=iter_idx, train_loss=train_loss)
  79. if args.enable_visdom:
  80. plot_prediction(logger, images, marking_points, prediction)
  81. torch.save(dp_detector.state_dict(),
  82. 'weights/dp_detector_%d.pth' % epoch_idx)
  83. torch.save(optimizer.state_dict(), 'weights/optimizer.pth')
  84. if __name__ == '__main__':
  85. train_detector(config.get_parser_for_training().parse_args())