Bienqq
/
car-detection-bayes
2
1
Fork 0
Code Issues Pull Requests Releases Wiki Activity

ng dependence removed from build_targets()

This commit is contained in:
Glenn Jocher 2020-04-13 21:25:03 -07:00
parent 0dd5f8eee8
commit 76fb8d48d4
2 changed files with 13 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
models.py
View File

@ -148,7 +148,7 @@ class YOLOLayer(nn.Module):
def create_grids(self, ng=(13, 13), device='cpu'): def create_grids(self, ng=(13, 13), device='cpu'):
self.nx, self.ny = ng # x and y grid size self.nx, self.ny = ng # x and y grid size
self.ng = torch.tensor(ng, device=device) self.ng = torch.tensor(ng)
# build xy offsets # build xy offsets
if not self.training: if not self.training:

25
utils/utils.py
View File

@ -366,7 +366,7 @@ def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#iss
def compute_loss(p, targets, model): # predictions, targets, model def compute_loss(p, targets, model): # predictions, targets, model
ft = torch.cuda.FloatTensor if p[0].is_cuda else torch.Tensor ft = torch.cuda.FloatTensor if p[0].is_cuda else torch.Tensor
lcls, lbox, lobj = ft([0]), ft([0]), ft([0]) lcls, lbox, lobj = ft([0]), ft([0]), ft([0])
tcls, tbox, indices, anchor_vec = build_targets(model, targets) tcls, tbox, indices, anchor_vec = build_targets(p, targets, model)
h = model.hyp # hyperparameters h = model.hyp # hyperparameters
red = 'mean' # Loss reduction (sum or mean) red = 'mean' # Loss reduction (sum or mean)
@ -430,42 +430,41 @@ def compute_loss(p, targets, model): # predictions, targets, model
return loss, torch.cat((lbox, lobj, lcls, loss)).detach() return loss, torch.cat((lbox, lobj, lcls, loss)).detach()
def build_targets(model, targets): def build_targets(p, targets, model):
# targets = [image, class, x, y, w, h] # targets = [image, class, x, y, w, h]
nt = targets.shape[0] nt = targets.shape[0]
tcls, tbox, indices, av = [], [], [], [] tcls, tbox, indices, av = [], [], [], []
multi_gpu = type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) multi_gpu = type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
reject, use_all_anchors = True, True reject, use_all_anchors = True, True
gain = torch.ones(6, device=targets.device) # normalized to gridspace gain
for i in model.yolo_layers: for i in model.yolo_layers:
# get number of grid points and anchor vec for this yolo layer # get number of grid points and anchor vec for this yolo layer
if multi_gpu: anchor_vec = model.module.module_list[i].anchor_vec if multi_gpu else model.module_list[i].anchor_vec
ng, anchor_vec = model.module.module_list[i].ng, model.module.module_list[i].anchor_vec
else:
ng, anchor_vec = model.module_list[i].ng, model.module_list[i].anchor_vec
# iou of targets-anchors # iou of targets-anchors
t, a = targets, [] gain[2:] = torch.tensor(p[0].shape)[[2, 3, 2, 3]] # xyxy gain
gwh = t[:, 4:6] * ng t, a = targets * gain, []
gwh = t[:, 4:6]
if nt: if nt:
iou = wh_iou(anchor_vec, gwh) # iou(3,n) = wh_iou(anchor_vec(3,2), gwh(n,2)) iou = wh_iou(anchor_vec, gwh) # iou(3,n) = wh_iou(anchor_vec(3,2), gwh(n,2))
if use_all_anchors: if use_all_anchors:
na = anchor_vec.shape[0] # number of anchors na = anchor_vec.shape[0] # number of anchors
a = torch.arange(na).view((-1, 1)).repeat([1, nt]).view(-1) a = torch.arange(na).view(-1, 1).repeat(1, nt).view(-1)
t = targets.repeat([na, 1]) t = targets.repeat(na, 1)
gwh = gwh.repeat([na, 1])
else: # use best anchor only else: # use best anchor only
iou, a = iou.max(0) # best iou and anchor iou, a = iou.max(0) # best iou and anchor
# reject anchors below iou_thres (OPTIONAL, increases P, lowers R) # reject anchors below iou_thres (OPTIONAL, increases P, lowers R)
if reject: if reject:
j = iou.view(-1) > model.hyp['iou_t'] # iou threshold hyperparameter j = iou.view(-1) > model.hyp['iou_t'] # iou threshold hyperparameter
t, a, gwh = t[j], a[j], gwh[j] t, a = t[j], a[j]
# Indices # Indices
b, c = t[:, :2].long().t() # target image, class b, c = t[:, :2].long().t() # target image, class
gxy = t[:, 2:4] * ng # grid x, y gxy = t[:, 2:4] # grid x, y
gwh = t[:, 4:6] # grid w, h
gi, gj = gxy.long().t() # grid x, y indices gi, gj = gxy.long().t() # grid x, y indices
indices.append((b, a, gj, gi)) indices.append((b, a, gj, gi))