import torch import torch.nn.functional as F import torch.nn as nn from torch.autograd import Variable import numpy as np __all__ = ['SegmentationLosses', 'OhemCrossEntropy2d', 'OHEMSegmentationLosses'] class SegmentationLosses(nn.CrossEntropyLoss): """2D Cross Entropy Loss with Auxilary Loss""" def __init__(self, se_loss=False, se_weight=0.2, nclass=-1, aux=False, aux_weight=0.4, weight=None, ignore_index=-1): super(SegmentationLosses, self).__init__(weight, None, ignore_index) self.se_loss = se_loss self.aux = aux self.nclass = nclass self.se_weight = se_weight self.aux_weight = aux_weight self.bceloss = nn.BCELoss(weight) def forward(self, *inputs): if not self.se_loss and not self.aux: return super(SegmentationLosses, self).forward(*inputs) elif not self.se_loss: pred1, pred2, target = tuple(inputs) loss1 = super(SegmentationLosses, self).forward(pred1, target) loss2 = super(SegmentationLosses, self).forward(pred2, target) return loss1 + self.aux_weight * loss2 elif not self.aux: pred, se_pred, target = tuple(inputs) se_target = self._get_batch_label_vector(target, nclass=self.nclass).type_as(pred) loss1 = super(SegmentationLosses, self).forward(pred, target) loss2 = self.bceloss(torch.sigmoid(se_pred), se_target) return loss1 + self.se_weight * loss2 else: pred1, se_pred, pred2, target = tuple(inputs) se_target = self._get_batch_label_vector(target, nclass=self.nclass).type_as(pred1) loss1 = super(SegmentationLosses, self).forward(pred1, target) loss2 = super(SegmentationLosses, self).forward(pred2, target) loss3 = self.bceloss(torch.sigmoid(se_pred), se_target) return loss1 + self.aux_weight * loss2 + self.se_weight * loss3 @staticmethod def _get_batch_label_vector(target, nclass): # target is a 3D Variable BxHxW, output is 2D BxnClass batch = target.size(0) tvect = Variable(torch.zeros(batch, nclass)) for i in range(batch): hist = torch.histc(target[i].cpu().data.float(), bins=nclass, min=0, max=nclass-1) vect = hist>0 tvect[i] = vect return tvect # adapted from https://github.com/PkuRainBow/OCNet/blob/master/utils/loss.py class OhemCrossEntropy2d(nn.Module): def __init__(self, ignore_label=-1, thresh=0.7, min_kept=100000, use_weight=True): super(OhemCrossEntropy2d, self).__init__() self.ignore_label = ignore_label self.thresh = float(thresh) self.min_kept = int(min_kept) if use_weight: print("w/ class balance") weight = torch.FloatTensor([0.8373, 0.918, 0.866, 1.0345, 1.0166, 0.9969, 0.9754, 1.0489, 0.8786, 1.0023, 0.9539, 0.9843, 1.1116, 0.9037, 1.0865, 1.0955, 1.0865, 1.1529, 1.0507]) self.criterion = torch.nn.CrossEntropyLoss(weight=weight, ignore_index=ignore_label) else: print("w/o class balance") self.criterion = torch.nn.CrossEntropyLoss(ignore_index=ignore_label) def forward(self, predict, target, weight=None): """ Args: predict:(n, c, h, w) target:(n, h, w) weight (Tensor, optional): a manual rescaling weight given to each class. If given, has to be a Tensor of size "nclasses" """ assert not target.requires_grad assert predict.dim() == 4 assert target.dim() == 3 assert predict.size(0) == target.size(0), "{0} vs {1} ".format(predict.size(0), target.size(0)) assert predict.size(2) == target.size(1), "{0} vs {1} ".format(predict.size(2), target.size(1)) assert predict.size(3) == target.size(2), "{0} vs {1} ".format(predict.size(3), target.size(3)) n, c, h, w = predict.size() input_label = target.data.cpu().numpy().ravel().astype(np.int32) x = np.rollaxis(predict.data.cpu().numpy(), 1).reshape((c, -1)) input_prob = np.exp(x - x.max(axis=0).reshape((1, -1))) input_prob /= input_prob.sum(axis=0).reshape((1, -1)) valid_flag = input_label != self.ignore_label valid_inds = np.where(valid_flag)[0] label = input_label[valid_flag] num_valid = valid_flag.sum() if self.min_kept >= num_valid: print('Labels: {}'.format(num_valid)) elif num_valid > 0: prob = input_prob[:,valid_flag] pred = prob[label, np.arange(len(label), dtype=np.int32)] threshold = self.thresh if self.min_kept > 0: index = pred.argsort() threshold_index = index[ min(len(index), self.min_kept) - 1 ] if pred[threshold_index] > self.thresh: threshold = pred[threshold_index] kept_flag = pred <= threshold valid_inds = valid_inds[kept_flag] label = input_label[valid_inds].copy() input_label.fill(self.ignore_label) input_label[valid_inds] = label valid_flag_new = input_label != self.ignore_label # print(np.sum(valid_flag_new)) target = Variable(torch.from_numpy(input_label.reshape(target.size())).long().cuda()) return self.criterion(predict, target) class OHEMSegmentationLosses(OhemCrossEntropy2d): """2D Cross Entropy Loss with Auxilary Loss""" def __init__(self, se_loss=False, se_weight=0.2, nclass=-1, aux=False, aux_weight=0.4, weight=None, ignore_index=-1): super(OHEMSegmentationLosses, self).__init__(ignore_index) self.se_loss = se_loss self.aux = aux self.nclass = nclass self.se_weight = se_weight self.aux_weight = aux_weight self.bceloss = nn.BCELoss(weight) def forward(self, *inputs): if not self.se_loss and not self.aux: return super(OHEMSegmentationLosses, self).forward(*inputs) elif not self.se_loss: pred1, pred2, target = tuple(inputs) loss1 = super(OHEMSegmentationLosses, self).forward(pred1, target) loss2 = super(OHEMSegmentationLosses, self).forward(pred2, target) return loss1 + self.aux_weight * loss2 elif not self.aux: pred, se_pred, target = tuple(inputs) se_target = self._get_batch_label_vector(target, nclass=self.nclass).type_as(pred) loss1 = super(OHEMSegmentationLosses, self).forward(pred, target) loss2 = self.bceloss(torch.sigmoid(se_pred), se_target) return loss1 + self.se_weight * loss2 else: pred1, se_pred, pred2, target = tuple(inputs) se_target = self._get_batch_label_vector(target, nclass=self.nclass).type_as(pred1) loss1 = super(OHEMSegmentationLosses, self).forward(pred1, target) loss2 = super(OHEMSegmentationLosses, self).forward(pred2, target) loss3 = self.bceloss(torch.sigmoid(se_pred), se_target) return loss1 + self.aux_weight * loss2 + self.se_weight * loss3 @staticmethod def _get_batch_label_vector(target, nclass): # target is a 3D Variable BxHxW, output is 2D BxnClass batch = target.size(0) tvect = Variable(torch.zeros(batch, nclass)) for i in range(batch): hist = torch.histc(target[i].cpu().data.float(), bins=nclass, min=0, max=nclass-1) vect = hist>0 tvect[i] = vect return tvect