In Masterarbeit:"Anomalie-Detektion in Zellbildern zur Anwendung der Leukämieerkennung" verwendete CSI Methode.
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

simclr.py 3.2KB

2 years ago
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101
  1. import time
  2. import torch.optim
  3. import models.transform_layers as TL
  4. from training.contrastive_loss import get_similarity_matrix, NT_xent
  5. from utils.utils import AverageMeter, normalize
  6. device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  7. hflip = TL.HorizontalFlipLayer().to(device)
  8. def train(P, epoch, model, criterion, optimizer, scheduler, loader, logger=None,
  9. simclr_aug=None, linear=None, linear_optim=None):
  10. assert simclr_aug is not None
  11. assert P.sim_lambda == 1.0
  12. if logger is None:
  13. log_ = print
  14. else:
  15. log_ = logger.log
  16. batch_time = AverageMeter()
  17. data_time = AverageMeter()
  18. losses = dict()
  19. losses['cls'] = AverageMeter()
  20. losses['sim'] = AverageMeter()
  21. check = time.time()
  22. for n, (images, labels) in enumerate(loader):
  23. model.train()
  24. count = n * P.n_gpus # number of trained samples
  25. data_time.update(time.time() - check)
  26. check = time.time()
  27. ### SimCLR loss ###
  28. if P.dataset != 'imagenet':
  29. batch_size = images.size(0)
  30. images = images.to(device)
  31. images_pair = hflip(images.repeat(2, 1, 1, 1)) # 2B with hflip
  32. else:
  33. batch_size = images[0].size(0)
  34. images1, images2 = images[0].to(device), images[1].to(device)
  35. images_pair = torch.cat([images1, images2], dim=0) # 2B
  36. labels = labels.to(device)
  37. images_pair = simclr_aug(images_pair) # transform
  38. _, outputs_aux = model(images_pair, simclr=True, penultimate=True)
  39. simclr = normalize(outputs_aux['simclr']) # normalize
  40. sim_matrix = get_similarity_matrix(simclr, multi_gpu=P.multi_gpu)
  41. loss_sim = NT_xent(sim_matrix, temperature=0.5) * P.sim_lambda
  42. ### total loss ###
  43. loss = loss_sim
  44. optimizer.zero_grad()
  45. loss.backward()
  46. optimizer.step()
  47. scheduler.step(epoch - 1 + n / len(loader))
  48. lr = optimizer.param_groups[0]['lr']
  49. batch_time.update(time.time() - check)
  50. ### Post-processing stuffs ###
  51. simclr_norm = outputs_aux['simclr'].norm(dim=1).mean()
  52. ### Linear evaluation ###
  53. outputs_linear_eval = linear(outputs_aux['penultimate'].detach())
  54. loss_linear = criterion(outputs_linear_eval, labels.repeat(2))
  55. linear_optim.zero_grad()
  56. loss_linear.backward()
  57. linear_optim.step()
  58. ### Log losses ###
  59. losses['cls'].update(0, batch_size)
  60. losses['sim'].update(loss_sim.item(), batch_size)
  61. if count % 50 == 0:
  62. log_('[Epoch %3d; %3d] [Time %.3f] [Data %.3f] [LR %.5f]\n'
  63. '[LossC %f] [LossSim %f]' %
  64. (epoch, count, batch_time.value, data_time.value, lr,
  65. losses['cls'].value, losses['sim'].value))
  66. check = time.time()
  67. log_('[DONE] [Time %.3f] [Data %.3f] [LossC %f] [LossSim %f]' %
  68. (batch_time.average, data_time.average,
  69. losses['cls'].average, losses['sim'].average))
  70. if logger is not None:
  71. logger.scalar_summary('train/loss_cls', losses['cls'].average, epoch)
  72. logger.scalar_summary('train/loss_sim', losses['sim'].average, epoch)
  73. logger.scalar_summary('train/batch_time', batch_time.average, epoch)