loss.py 9.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216
  1. # Loss functions
  2. import torch
  3. import torch.nn as nn
  4. from utils.general import bbox_iou
  5. from utils.torch_utils import is_parallel
  6. def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
  7. # return positive, negative label smoothing BCE targets
  8. return 1.0 - 0.5 * eps, 0.5 * eps
  9. class BCEBlurWithLogitsLoss(nn.Module):
  10. # BCEwithLogitLoss() with reduced missing label effects.
  11. def __init__(self, alpha=0.05):
  12. super(BCEBlurWithLogitsLoss, self).__init__()
  13. self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss()
  14. self.alpha = alpha
  15. def forward(self, pred, true):
  16. loss = self.loss_fcn(pred, true)
  17. pred = torch.sigmoid(pred) # prob from logits
  18. dx = pred - true # reduce only missing label effects
  19. # dx = (pred - true).abs() # reduce missing label and false label effects
  20. alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
  21. loss *= alpha_factor
  22. return loss.mean()
  23. class FocalLoss(nn.Module):
  24. # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
  25. def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
  26. super(FocalLoss, self).__init__()
  27. self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
  28. self.gamma = gamma
  29. self.alpha = alpha
  30. self.reduction = loss_fcn.reduction
  31. self.loss_fcn.reduction = 'none' # required to apply FL to each element
  32. def forward(self, pred, true):
  33. loss = self.loss_fcn(pred, true)
  34. # p_t = torch.exp(-loss)
  35. # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability
  36. # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
  37. pred_prob = torch.sigmoid(pred) # prob from logits
  38. p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
  39. alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
  40. modulating_factor = (1.0 - p_t) ** self.gamma
  41. loss *= alpha_factor * modulating_factor
  42. if self.reduction == 'mean':
  43. return loss.mean()
  44. elif self.reduction == 'sum':
  45. return loss.sum()
  46. else: # 'none'
  47. return loss
  48. class QFocalLoss(nn.Module):
  49. # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
  50. def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
  51. super(QFocalLoss, self).__init__()
  52. self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
  53. self.gamma = gamma
  54. self.alpha = alpha
  55. self.reduction = loss_fcn.reduction
  56. self.loss_fcn.reduction = 'none' # required to apply FL to each element
  57. def forward(self, pred, true):
  58. loss = self.loss_fcn(pred, true)
  59. pred_prob = torch.sigmoid(pred) # prob from logits
  60. alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
  61. modulating_factor = torch.abs(true - pred_prob) ** self.gamma
  62. loss *= alpha_factor * modulating_factor
  63. if self.reduction == 'mean':
  64. return loss.mean()
  65. elif self.reduction == 'sum':
  66. return loss.sum()
  67. else: # 'none'
  68. return loss
  69. class ComputeLoss:
  70. # Compute losses
  71. def __init__(self, model, autobalance=False):
  72. super(ComputeLoss, self).__init__()
  73. device = next(model.parameters()).device # get model device
  74. h = model.hyp # hyperparameters
  75. # Define criteria
  76. BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
  77. BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
  78. # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
  79. self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
  80. # Focal loss
  81. g = h['fl_gamma'] # focal loss gamma
  82. if g > 0:
  83. BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
  84. det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
  85. self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
  86. self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
  87. self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
  88. for k in 'na', 'nc', 'nl', 'anchors':
  89. setattr(self, k, getattr(det, k))
  90. def __call__(self, p, targets): # predictions, targets, model
  91. device = targets.device
  92. lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
  93. tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
  94. # Losses
  95. for i, pi in enumerate(p): # layer index, layer predictions
  96. b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
  97. tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
  98. n = b.shape[0] # number of targets
  99. if n:
  100. ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
  101. # Regression
  102. pxy = ps[:, :2].sigmoid() * 2. - 0.5
  103. pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
  104. pbox = torch.cat((pxy, pwh), 1) # predicted box
  105. iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target)
  106. lbox += (1.0 - iou).mean() # iou loss
  107. # Objectness
  108. tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
  109. # Classification
  110. if self.nc > 1: # cls loss (only if multiple classes)
  111. t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
  112. t[range(n), tcls[i]] = self.cp
  113. lcls += self.BCEcls(ps[:, 5:], t) # BCE
  114. # Append targets to text file
  115. # with open('targets.txt', 'a') as file:
  116. # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
  117. obji = self.BCEobj(pi[..., 4], tobj)
  118. lobj += obji * self.balance[i] # obj loss
  119. if self.autobalance:
  120. self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
  121. if self.autobalance:
  122. self.balance = [x / self.balance[self.ssi] for x in self.balance]
  123. lbox *= self.hyp['box']
  124. lobj *= self.hyp['obj']
  125. lcls *= self.hyp['cls']
  126. bs = tobj.shape[0] # batch size
  127. loss = lbox + lobj + lcls
  128. return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
  129. def build_targets(self, p, targets):
  130. # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
  131. na, nt = self.na, targets.shape[0] # number of anchors, targets
  132. tcls, tbox, indices, anch = [], [], [], []
  133. gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
  134. ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
  135. targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
  136. g = 0.5 # bias
  137. off = torch.tensor([[0, 0],
  138. [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
  139. # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
  140. ], device=targets.device).float() * g # offsets
  141. for i in range(self.nl):
  142. anchors = self.anchors[i]
  143. gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
  144. # Match targets to anchors
  145. t = targets * gain
  146. if nt:
  147. # Matches
  148. r = t[:, :, 4:6] / anchors[:, None] # wh ratio
  149. j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
  150. # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
  151. t = t[j] # filter
  152. # Offsets
  153. gxy = t[:, 2:4] # grid xy
  154. gxi = gain[[2, 3]] - gxy # inverse
  155. j, k = ((gxy % 1. < g) & (gxy > 1.)).T
  156. l, m = ((gxi % 1. < g) & (gxi > 1.)).T
  157. j = torch.stack((torch.ones_like(j), j, k, l, m))
  158. t = t.repeat((5, 1, 1))[j]
  159. offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
  160. else:
  161. t = targets[0]
  162. offsets = 0
  163. # Define
  164. b, c = t[:, :2].long().T # image, class
  165. gxy = t[:, 2:4] # grid xy
  166. gwh = t[:, 4:6] # grid wh
  167. gij = (gxy - offsets).long()
  168. gi, gj = gij.T # grid xy indices
  169. # Append
  170. a = t[:, 6].long() # anchor indices
  171. indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
  172. tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
  173. anch.append(anchors[a]) # anchors
  174. tcls.append(c) # class
  175. return tcls, tbox, indices, anch