def clas_L(pred_batch, *gt_batch, lambda_clas=1, clas_weights=None, gaf):
'''
class loss
若某anchor对某object负责，则应训练其classification靠近该object的类别。
'''
loss = 0
cnt = 0
for pred_clas, gt_bb_or_lpts, gt_clas in zip(pred_batch[2], *gt_batch):
if gaf.LblPts:
keep = get_y(gt_bb_or_lpts)
else:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_clas = gt_clas[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
gt_clas = gt_clas - 1 # the databunch add a 'background' class to classes[0], but we don't want that,so gt_clas-1
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
pred_clas = pred_clas[idx]
loss += F.cross_entropy(pred_clas,
gt_clas,
weight=clas_weights,
reduction='sum')
cnt += gt_clas.shape[0]
return lambda_clas * loss / cnt
def cent_d(pred_batch, *gt_batch, gaf):
'''
bbox center difference
'''
dif = tensor(0.)
cnt = tensor(0.)
for pred_txy, gt_bb_or_lpts, _ in zip(pred_batch[0], *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
#pred_t = torch.cat([pred_txy,pred_thw],dim=1)
pred_t = pred_txy
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
pred_t = pred_t[idx]
if not gaf.LblPts:
pred_c = bbox2c(gaf.c2b(pred_t, idx))[..., :2]
gt_c = bbox2c(gt_bb_or_lpts)[..., :2]
else:
pred_c = gaf.c2b(pred_t, idx)
gt_c = gt_bb_or_lpts
tmp = (gt_c - pred_c).abs().sum()
dif += tmp
cnt += len(idx)
return dif / cnt / 2
def clas_acc(pred_batch, *gt_batch, gaf):
'''
classification accuracy
'''
posCnt = tensor(0.)
totCnt = tensor(0.)
for pred_clas, gt_bb_or_lpts, gt_clas in zip(pred_batch[2], *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_clas = gt_clas[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
gt_clas = gt_clas - 1 # the databunch add a 'background' class to classes[0], but we don't want that,so gt_clas-1
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
pred_clas = pred_clas[idx]
pred_clas = pred_clas.max(1)[1]
posCnt += (pred_clas == gt_clas).sum().item()
totCnt += gt_clas.shape[0]
return posCnt / totCnt
def nConf_L(pred_batch, *gt_batch, gaf, conf_th=0.5, lambda_nconf=1):
'''
negative confidence loss
若某 anchor 不对任何 object 负责，且它与任何 object 的 匹配得分 都差于 threshold，则训练其 conf_score 靠近 0。
'''
loss = 0
cnt = 0
for pred_conf, gt_bb_or_lpts, _ in zip(pred_batch[1], *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
#positive
tmp = (hits * 1).max(dim=0)[0]
#取得命中的anchor周围的anchor的下标立标
discards = []
for hidx in idx:
neibs = find_neibs(hidx, gaf.grids[0], dis=1)
for i in neibs:
discards += [i]
#把周围的邻居加进来，剩下的就是negative了
tmp[discards] = 1
neg_idx = torch.where(tmp == 0)[0] # 如果没有，该anchor是negative anchor
conf_neg = pred_conf[neg_idx]
# conf_neg = torch.sigmoid(conf_neg)
# loss += conf_neg.abs().sum()
tmp = F.binary_cross_entropy_with_logits(conf_neg,
torch.zeros_like(conf_neg),
reduction='sum')
loss += tmp
cnt += len(neg_idx)
return lambda_nconf * loss / cnt
def pConf_L(pred_batch, *gt_batch, lambda_pconf=1, clas_weights=None, gaf):
'''
positive confidence loss
若某 anchor 为某 object 负责，则训练其 conf_score 靠近 1。
'''
loss = 0
cnt = 0
for pred_conf, gt_bb_or_lpts, gt_clas in zip(pred_batch[1], *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_clas = gt_clas[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
gt_clas = gt_clas - 1
if clas_weights is not None: ws = clas_weights[gt_clas]
else: ws = None
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
conf_pos = pred_conf[idx]
# conf_pos = torch.sigmoid(conf_pos)
# tmp = (1-conf_pos).abs().sum()
if ws is not None:
tmp = F.binary_cross_entropy_with_logits(conf_pos,
torch.ones_like(conf_pos),
weight=ws[..., None],
reduction='sum')
else:
tmp = F.binary_cross_entropy_with_logits(conf_pos,
torch.ones_like(conf_pos),
reduction='sum')
loss += tmp
cnt += len(idx)
return lambda_pconf * loss / cnt
def cent_L(pred_batch, *gt_batch, lambda_cent=1, clas_weights=None, gaf):
'''
bbox center loss
若某 anchor 对某 object 负责，则应训练其预测之 中心 靠近该 object box 之 中心。
'''
loss = 0
cnt = 0
for pred_txy, gt_bb_or_lpts, gt_clas in zip(pred_batch[0], *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0: continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_clas = gt_clas[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
gt_clas = gt_clas - 1
if clas_weights is not None: ws = clas_weights[gt_clas]
else: ws = None
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
gt_t = gaf.b2c(gt_bb_or_lpts, idx, eps=1)
pred_txy = pred_txy[idx]
if ws is not None:
tmp = ((gt_t[..., :2] - pred_txy) * ws[..., None]).abs().sum()
else:
tmp = (gt_t[..., :2] - pred_txy).abs().sum()
loss += tmp
cnt += len(idx)
return lambda_cent * loss / cnt
def yolo_L(pred_batch,
*gt_batch,
conf_th=0.5,
lambda_cent=1,
lambda_pconf=1,
lambda_nconf=1,
lambda_clas=1,
clas_weights=None,
gaf):
'''
与detect_symbol里面的yolo_L相比的区别是：
不计算宽高方面的损失
neg_idx要去掉find_neibs返回的discard列表
clas_weights:
为了解决数据集的imbalance问题，一种方法是在dataloader中使用WeightedRandomSampler，但是这种方法不适用于目标检测问题。
因为，（1）目标检测的label不是一个简单的数值（2）目标检测问题的一张图片可能包括不同类别的多个目标。
所以为了解决目标检测问题中的imbalance问题，我们的方法是在损失函数中使用权重。
为各类别分配权重，各目标对应的损失乘以该目标所属类别的权重。
默认为None，即不使用权重。
若设置非None，则clas_weights应该是一个一维tensor，其长度等于数据集的类别数。
若设置为全1，则相当于不使用权重。
合理的设置应保证所有元素之和等于数据集的类别数，否则相当于对损失函数的整体做了缩放。
'''
clas_loss = 0
cent_loss = 0
pconf_loss = 0
nconf_loss = 0
pos_cnt = 0
neg_cnt = 0
for pred_txy, pred_conf, pred_clas, gt_bb_or_lpts, gt_clas in zip(
*pred_batch, *gt_batch):
if not gaf.LblPts:
keep = anchors_loss_metrics_detsym.get_y(gt_bb_or_lpts)
else:
keep = get_y(gt_bb_or_lpts)
if keep.numel() == 0:
#这时候所有anchor都是negative的。所以空白的也要贡献自己的loss
conf_neg = pred_conf #所有anchor的
nconf_loss += F.binary_cross_entropy_with_logits(
conf_neg, torch.zeros_like(conf_neg), reduction='sum')
neg_cnt += len(pred_conf)
continue
gt_bb_or_lpts = gt_bb_or_lpts[keep]
gt_clas = gt_clas[keep]
gt_bb_or_lpts = (gt_bb_or_lpts + 1) / 2
gt_clas = gt_clas - 1 # the databunch add a 'background' class to classes[0], but we don't want that,so gt_clas-1
if clas_weights is not None: ws = clas_weights[gt_clas]
else: ws = None
hits = gaf.get_hits(gt_bb_or_lpts)
idx = idx_fromHits(hits)
# classification loss
pred_clas = pred_clas[idx]
clas_loss += F.cross_entropy(pred_clas,
gt_clas,
weight=clas_weights,
reduction='sum')
# bbox center loss
gt_t = gaf.b2c(gt_bb_or_lpts, idx, eps=1)
pred_txy = pred_txy[idx]
if ws is not None:
cent_loss += ((gt_t[..., :2] - pred_txy) *
ws[..., None]).abs().sum()
else:
cent_loss += (gt_t[..., :2] - pred_txy).abs().sum()
# positive confidence loss
conf_pos = pred_conf[idx]
if ws is not None:
pconf_loss += F.binary_cross_entropy_with_logits(
conf_pos,
torch.ones_like(conf_pos),
weight=ws[..., None],
reduction='sum')
else:
pconf_loss += F.binary_cross_entropy_with_logits(
conf_pos, torch.ones_like(conf_pos), reduction='sum')
#positive
tmp = (hits * 1).max(dim=0)[0]
#取得命中的anchor周围的anchor的下标立标
discards = []
for hidx in idx:
neibs = find_neibs(hidx, gaf.grids[0], dis=1)
for i in neibs:
discards += [i]
#把周围的邻居加进来，剩下的就是negative了
tmp[discards] = 1
neg_idx = torch.where(tmp == 0)[0] # 如果没有，该anchor是negative anchor
conf_neg = pred_conf[neg_idx]
nconf_loss += F.binary_cross_entropy_with_logits(
conf_neg, torch.zeros_like(conf_neg), reduction='sum')
pos_cnt += len(idx)
neg_cnt += len(neg_idx)
if pos_cnt > 0: #测试的极端情况碰到都是空白的。只有nconf_loss在前面计算了。
clas_loss = lambda_clas * clas_loss / pos_cnt
cent_loss = lambda_cent * cent_loss / pos_cnt
pconf_loss = lambda_pconf * pconf_loss / pos_cnt
nconf_loss = lambda_nconf * nconf_loss / neg_cnt
return clas_loss + cent_loss + pconf_loss + nconf_loss