ultralytics 8.0.167 Tuner updates and HUB Pose and Classify fixes (#4656)
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
Glenn Jocher
GitHub
parent
8596ee241f
commit
d2cf7acce0
21 changed files with 174 additions and 144 deletions
|
|
@ -84,35 +84,36 @@ class DETRLoss(nn.Module):
|
|||
loss[name_giou] = self.loss_gain['giou'] * loss[name_giou]
|
||||
return {k: v.squeeze() for k, v in loss.items()}
|
||||
|
||||
def _get_loss_mask(self, masks, gt_mask, match_indices, postfix=''):
|
||||
# masks: [b, query, h, w], gt_mask: list[[n, H, W]]
|
||||
name_mask = f'loss_mask{postfix}'
|
||||
name_dice = f'loss_dice{postfix}'
|
||||
# This function is for future RT-DETR Segment models
|
||||
# def _get_loss_mask(self, masks, gt_mask, match_indices, postfix=''):
|
||||
# # masks: [b, query, h, w], gt_mask: list[[n, H, W]]
|
||||
# name_mask = f'loss_mask{postfix}'
|
||||
# name_dice = f'loss_dice{postfix}'
|
||||
#
|
||||
# loss = {}
|
||||
# if sum(len(a) for a in gt_mask) == 0:
|
||||
# loss[name_mask] = torch.tensor(0., device=self.device)
|
||||
# loss[name_dice] = torch.tensor(0., device=self.device)
|
||||
# return loss
|
||||
#
|
||||
# num_gts = len(gt_mask)
|
||||
# src_masks, target_masks = self._get_assigned_bboxes(masks, gt_mask, match_indices)
|
||||
# src_masks = F.interpolate(src_masks.unsqueeze(0), size=target_masks.shape[-2:], mode='bilinear')[0]
|
||||
# # TODO: torch does not have `sigmoid_focal_loss`, but it's not urgent since we don't use mask branch for now.
|
||||
# loss[name_mask] = self.loss_gain['mask'] * F.sigmoid_focal_loss(src_masks, target_masks,
|
||||
# torch.tensor([num_gts], dtype=torch.float32))
|
||||
# loss[name_dice] = self.loss_gain['dice'] * self._dice_loss(src_masks, target_masks, num_gts)
|
||||
# return loss
|
||||
|
||||
loss = {}
|
||||
if sum(len(a) for a in gt_mask) == 0:
|
||||
loss[name_mask] = torch.tensor(0., device=self.device)
|
||||
loss[name_dice] = torch.tensor(0., device=self.device)
|
||||
return loss
|
||||
|
||||
num_gts = len(gt_mask)
|
||||
src_masks, target_masks = self._get_assigned_bboxes(masks, gt_mask, match_indices)
|
||||
src_masks = F.interpolate(src_masks.unsqueeze(0), size=target_masks.shape[-2:], mode='bilinear')[0]
|
||||
# TODO: torch does not have `sigmoid_focal_loss`, but it's not urgent since we don't use mask branch for now.
|
||||
loss[name_mask] = self.loss_gain['mask'] * F.sigmoid_focal_loss(src_masks, target_masks,
|
||||
torch.tensor([num_gts], dtype=torch.float32))
|
||||
loss[name_dice] = self.loss_gain['dice'] * self._dice_loss(src_masks, target_masks, num_gts)
|
||||
return loss
|
||||
|
||||
@staticmethod
|
||||
def _dice_loss(inputs, targets, num_gts):
|
||||
inputs = F.sigmoid(inputs)
|
||||
inputs = inputs.flatten(1)
|
||||
targets = targets.flatten(1)
|
||||
numerator = 2 * (inputs * targets).sum(1)
|
||||
denominator = inputs.sum(-1) + targets.sum(-1)
|
||||
loss = 1 - (numerator + 1) / (denominator + 1)
|
||||
return loss.sum() / num_gts
|
||||
# This function is for future RT-DETR Segment models
|
||||
# @staticmethod
|
||||
# def _dice_loss(inputs, targets, num_gts):
|
||||
# inputs = F.sigmoid(inputs).flatten(1)
|
||||
# targets = targets.flatten(1)
|
||||
# numerator = 2 * (inputs * targets).sum(1)
|
||||
# denominator = inputs.sum(-1) + targets.sum(-1)
|
||||
# loss = 1 - (numerator + 1) / (denominator + 1)
|
||||
# return loss.sum() / num_gts
|
||||
|
||||
def _get_loss_aux(self,
|
||||
pred_bboxes,
|
||||
|
|
|
|||
|
|
@ -110,34 +110,35 @@ class HungarianMatcher(nn.Module):
|
|||
return [(torch.tensor(i, dtype=torch.long), torch.tensor(j, dtype=torch.long) + gt_groups[k])
|
||||
for k, (i, j) in enumerate(indices)]
|
||||
|
||||
def _cost_mask(self, bs, num_gts, masks=None, gt_mask=None):
|
||||
assert masks is not None and gt_mask is not None, 'Make sure the input has `mask` and `gt_mask`'
|
||||
# all masks share the same set of points for efficient matching
|
||||
sample_points = torch.rand([bs, 1, self.num_sample_points, 2])
|
||||
sample_points = 2.0 * sample_points - 1.0
|
||||
|
||||
out_mask = F.grid_sample(masks.detach(), sample_points, align_corners=False).squeeze(-2)
|
||||
out_mask = out_mask.flatten(0, 1)
|
||||
|
||||
tgt_mask = torch.cat(gt_mask).unsqueeze(1)
|
||||
sample_points = torch.cat([a.repeat(b, 1, 1, 1) for a, b in zip(sample_points, num_gts) if b > 0])
|
||||
tgt_mask = F.grid_sample(tgt_mask, sample_points, align_corners=False).squeeze([1, 2])
|
||||
|
||||
with torch.cuda.amp.autocast(False):
|
||||
# binary cross entropy cost
|
||||
pos_cost_mask = F.binary_cross_entropy_with_logits(out_mask, torch.ones_like(out_mask), reduction='none')
|
||||
neg_cost_mask = F.binary_cross_entropy_with_logits(out_mask, torch.zeros_like(out_mask), reduction='none')
|
||||
cost_mask = torch.matmul(pos_cost_mask, tgt_mask.T) + torch.matmul(neg_cost_mask, 1 - tgt_mask.T)
|
||||
cost_mask /= self.num_sample_points
|
||||
|
||||
# dice cost
|
||||
out_mask = F.sigmoid(out_mask)
|
||||
numerator = 2 * torch.matmul(out_mask, tgt_mask.T)
|
||||
denominator = out_mask.sum(-1, keepdim=True) + tgt_mask.sum(-1).unsqueeze(0)
|
||||
cost_dice = 1 - (numerator + 1) / (denominator + 1)
|
||||
|
||||
C = self.cost_gain['mask'] * cost_mask + self.cost_gain['dice'] * cost_dice
|
||||
return C
|
||||
# This function is for future RT-DETR Segment models
|
||||
# def _cost_mask(self, bs, num_gts, masks=None, gt_mask=None):
|
||||
# assert masks is not None and gt_mask is not None, 'Make sure the input has `mask` and `gt_mask`'
|
||||
# # all masks share the same set of points for efficient matching
|
||||
# sample_points = torch.rand([bs, 1, self.num_sample_points, 2])
|
||||
# sample_points = 2.0 * sample_points - 1.0
|
||||
#
|
||||
# out_mask = F.grid_sample(masks.detach(), sample_points, align_corners=False).squeeze(-2)
|
||||
# out_mask = out_mask.flatten(0, 1)
|
||||
#
|
||||
# tgt_mask = torch.cat(gt_mask).unsqueeze(1)
|
||||
# sample_points = torch.cat([a.repeat(b, 1, 1, 1) for a, b in zip(sample_points, num_gts) if b > 0])
|
||||
# tgt_mask = F.grid_sample(tgt_mask, sample_points, align_corners=False).squeeze([1, 2])
|
||||
#
|
||||
# with torch.cuda.amp.autocast(False):
|
||||
# # binary cross entropy cost
|
||||
# pos_cost_mask = F.binary_cross_entropy_with_logits(out_mask, torch.ones_like(out_mask), reduction='none')
|
||||
# neg_cost_mask = F.binary_cross_entropy_with_logits(out_mask, torch.zeros_like(out_mask), reduction='none')
|
||||
# cost_mask = torch.matmul(pos_cost_mask, tgt_mask.T) + torch.matmul(neg_cost_mask, 1 - tgt_mask.T)
|
||||
# cost_mask /= self.num_sample_points
|
||||
#
|
||||
# # dice cost
|
||||
# out_mask = F.sigmoid(out_mask)
|
||||
# numerator = 2 * torch.matmul(out_mask, tgt_mask.T)
|
||||
# denominator = out_mask.sum(-1, keepdim=True) + tgt_mask.sum(-1).unsqueeze(0)
|
||||
# cost_dice = 1 - (numerator + 1) / (denominator + 1)
|
||||
#
|
||||
# C = self.cost_gain['mask'] * cost_mask + self.cost_gain['dice'] * cost_dice
|
||||
# return C
|
||||
|
||||
|
||||
def get_cdn_group(batch,
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue