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.

loss.py 9.4KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222
  1. # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
  2. """
  3. Loss functions
  4. """
  5. import torch
  6. import torch.nn as nn
  7. from utils.metrics import bbox_iou
  8. from utils.torch_utils import de_parallel
  9. def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
  10. # return positive, negative label smoothing BCE targets
  11. return 1.0 - 0.5 * eps, 0.5 * eps
  12. class BCEBlurWithLogitsLoss(nn.Module):
  13. # BCEwithLogitLoss() with reduced missing label effects.
  14. def __init__(self, alpha=0.05):
  15. super().__init__()
  16. self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss()
  17. self.alpha = alpha
  18. def forward(self, pred, true):
  19. loss = self.loss_fcn(pred, true)
  20. pred = torch.sigmoid(pred) # prob from logits
  21. dx = pred - true # reduce only missing label effects
  22. # dx = (pred - true).abs() # reduce missing label and false label effects
  23. alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
  24. loss *= alpha_factor
  25. return loss.mean()
  26. class FocalLoss(nn.Module):
  27. # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
  28. def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
  29. super().__init__()
  30. self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
  31. self.gamma = gamma
  32. self.alpha = alpha
  33. self.reduction = loss_fcn.reduction
  34. self.loss_fcn.reduction = 'none' # required to apply FL to each element
  35. def forward(self, pred, true):
  36. loss = self.loss_fcn(pred, true)
  37. # p_t = torch.exp(-loss)
  38. # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability
  39. # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
  40. pred_prob = torch.sigmoid(pred) # prob from logits
  41. p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
  42. alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
  43. modulating_factor = (1.0 - p_t) ** self.gamma
  44. loss *= alpha_factor * modulating_factor
  45. if self.reduction == 'mean':
  46. return loss.mean()
  47. elif self.reduction == 'sum':
  48. return loss.sum()
  49. else: # 'none'
  50. return loss
  51. class QFocalLoss(nn.Module):
  52. # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
  53. def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
  54. super().__init__()
  55. self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
  56. self.gamma = gamma
  57. self.alpha = alpha
  58. self.reduction = loss_fcn.reduction
  59. self.loss_fcn.reduction = 'none' # required to apply FL to each element
  60. def forward(self, pred, true):
  61. loss = self.loss_fcn(pred, true)
  62. pred_prob = torch.sigmoid(pred) # prob from logits
  63. alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
  64. modulating_factor = torch.abs(true - pred_prob) ** self.gamma
  65. loss *= alpha_factor * modulating_factor
  66. if self.reduction == 'mean':
  67. return loss.mean()
  68. elif self.reduction == 'sum':
  69. return loss.sum()
  70. else: # 'none'
  71. return loss
  72. class ComputeLoss:
  73. # Compute losses
  74. def __init__(self, model, autobalance=False):
  75. self.sort_obj_iou = False
  76. device = next(model.parameters()).device # get model device
  77. h = model.hyp # hyperparameters
  78. # Define criteria
  79. BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
  80. BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
  81. # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
  82. self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
  83. # Focal loss
  84. g = h['fl_gamma'] # focal loss gamma
  85. if g > 0:
  86. BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
  87. det = de_parallel(model).model[-1] # Detect() module
  88. self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, 0.02]) # P3-P7
  89. self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
  90. self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
  91. for k in 'na', 'nc', 'nl', 'anchors':
  92. setattr(self, k, getattr(det, k))
  93. def __call__(self, p, targets): # predictions, targets, model
  94. device = targets.device
  95. lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
  96. tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
  97. # Losses
  98. for i, pi in enumerate(p): # layer index, layer predictions
  99. b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
  100. tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
  101. n = b.shape[0] # number of targets
  102. if n:
  103. ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
  104. # Regression
  105. pxy = ps[:, :2].sigmoid() * 2 - 0.5
  106. pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
  107. pbox = torch.cat((pxy, pwh), 1) # predicted box
  108. iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target)
  109. lbox += (1.0 - iou).mean() # iou loss
  110. # Objectness
  111. score_iou = iou.detach().clamp(0).type(tobj.dtype)
  112. if self.sort_obj_iou:
  113. sort_id = torch.argsort(score_iou)
  114. b, a, gj, gi, score_iou = b[sort_id], a[sort_id], gj[sort_id], gi[sort_id], score_iou[sort_id]
  115. tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * score_iou # iou ratio
  116. # Classification
  117. if self.nc > 1: # cls loss (only if multiple classes)
  118. t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
  119. t[range(n), tcls[i]] = self.cp
  120. lcls += self.BCEcls(ps[:, 5:], t) # BCE
  121. # Append targets to text file
  122. # with open('targets.txt', 'a') as file:
  123. # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
  124. obji = self.BCEobj(pi[..., 4], tobj)
  125. lobj += obji * self.balance[i] # obj loss
  126. if self.autobalance:
  127. self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
  128. if self.autobalance:
  129. self.balance = [x / self.balance[self.ssi] for x in self.balance]
  130. lbox *= self.hyp['box']
  131. lobj *= self.hyp['obj']
  132. lcls *= self.hyp['cls']
  133. bs = tobj.shape[0] # batch size
  134. return (lbox + lobj + lcls) * bs, torch.cat((lbox, lobj, lcls)).detach()
  135. def build_targets(self, p, targets):
  136. # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
  137. na, nt = self.na, targets.shape[0] # number of anchors, targets
  138. tcls, tbox, indices, anch = [], [], [], []
  139. gain = torch.ones(7, device=targets.device) # normalized to gridspace gain
  140. ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
  141. targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
  142. g = 0.5 # bias
  143. off = torch.tensor([[0, 0],
  144. [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
  145. # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
  146. ], device=targets.device).float() * g # offsets
  147. for i in range(self.nl):
  148. anchors = self.anchors[i]
  149. gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
  150. # Match targets to anchors
  151. t = targets * gain
  152. if nt:
  153. # Matches
  154. r = t[:, :, 4:6] / anchors[:, None] # wh ratio
  155. j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t'] # compare
  156. # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
  157. t = t[j] # filter
  158. # Offsets
  159. gxy = t[:, 2:4] # grid xy
  160. gxi = gain[[2, 3]] - gxy # inverse
  161. j, k = ((gxy % 1 < g) & (gxy > 1)).T
  162. l, m = ((gxi % 1 < g) & (gxi > 1)).T
  163. j = torch.stack((torch.ones_like(j), j, k, l, m))
  164. t = t.repeat((5, 1, 1))[j]
  165. offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
  166. else:
  167. t = targets[0]
  168. offsets = 0
  169. # Define
  170. b, c = t[:, :2].long().T # image, class
  171. gxy = t[:, 2:4] # grid xy
  172. gwh = t[:, 4:6] # grid wh
  173. gij = (gxy - offsets).long()
  174. gi, gj = gij.T # grid xy indices
  175. # Append
  176. a = t[:, 6].long() # anchor indices
  177. indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
  178. tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
  179. anch.append(anchors[a]) # anchors
  180. tcls.append(c) # class
  181. return tcls, tbox, indices, anch