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

update to coco results82

This commit is contained in:
Glenn Jocher 2020-03-13 12:06:48 -07:00
parent 0de07da612
commit f30a8706e5
2 changed files with 16 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
train.py
View File

@ -97,7 +97,6 @@ def train():
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True) optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases) optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
optimizer.param_groups[2]['lr'] *= 2.0 # bias lr
del pg0, pg1, pg2 del pg0, pg1, pg2
start_epoch = 0 start_epoch = 0
@ -138,7 +137,7 @@ def train():
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0) model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://github.com/ultralytics/yolov3/issues/238 # Scheduler https://github.com/ultralytics/yolov3/issues/238
lf = lambda x: (1 + math.cos(x * math.pi / epochs)) / 2 * 0.99 + 0.01 # cosine https://arxiv.org/pdf/1812.01187.pdf lf = lambda x: (1 + math.cos(x * math.pi / epochs)) / 2 # cosine https://arxiv.org/pdf/1812.01187.pdf
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1) scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1)
# scheduler = lr_scheduler.MultiStepLR(optimizer, [round(epochs * x) for x in [0.8, 0.9]], 0.1, start_epoch - 1) # scheduler = lr_scheduler.MultiStepLR(optimizer, [round(epochs * x) for x in [0.8, 0.9]], 0.1, start_epoch - 1)
@ -192,8 +191,8 @@ def train():
collate_fn=dataset.collate_fn) collate_fn=dataset.collate_fn)
# Start training # Start training
nb = len(dataloader) nb = len(dataloader) # number of batches
prebias = False # start_epoch == 0 prebias = start_epoch == 0
model.nc = nc # attach number of classes to model model.nc = nc # attach number of classes to model
model.arc = opt.arc # attach yolo architecture model.arc = opt.arc # attach yolo architecture
model.hyp = hyp # attach hyperparameters to model model.hyp = hyp # attach hyperparameters to model
@ -207,14 +206,15 @@ def train():
print('Starting training for %g epochs...' % epochs) print('Starting training for %g epochs...' % epochs)
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
model.train() model.train()
model.gr = 1 - (1 + math.cos(min(epoch * 1, epochs) * math.pi / epochs)) / 2 # GIoU <-> 1.0 loss ratio
# Prebias # Prebias
if prebias: if prebias:
ne = max(round(30 / nb), 3) # number of prebias epochs ne = 3 # number of prebias epochs
ps = np.interp(epoch, [0, ne], [0.1, hyp['lr0'] * 2]), \ ps = 0.1, 0.9 # prebias settings (lr=0.1, momentum=0.9)
np.interp(epoch, [0, ne], [0.9, hyp['momentum']]) # prebias settings (lr=0.1, momentum=0.9) model.gr = 0.0 # giou loss ratio (obj_loss = 1.0)
if epoch == ne: if epoch == ne:
ps = hyp['lr0'], hyp['momentum'] # normal training settings
model.gr = 1.0 # giou loss ratio (obj_loss = giou)
print_model_biases(model) print_model_biases(model)
prebias = False prebias = False
@ -240,13 +240,12 @@ def train():
# Hyperparameter Burn-in # Hyperparameter Burn-in
n_burn = 200 # number of burn-in batches n_burn = 200 # number of burn-in batches
if ni <= n_burn: if ni <= n_burn:
g = ni / n_burn # gain # g = ni / n_burn # gain
for x in model.named_modules(): for x in model.named_modules(): # initial stats may be poor, wait to track
if x[0].endswith('BatchNorm2d'): if x[0].endswith('BatchNorm2d'):
# x[1].momentum = 1 - 0.9 * g # momentum falls from 1 - 0.1
x[1].track_running_stats = ni == n_burn x[1].track_running_stats = ni == n_burn
for x in optimizer.param_groups: # for x in optimizer.param_groups:
x['lr'] = x['initial_lr'] * lf(epoch) * g # gain rises from 0 - 1 # x['lr'] = x['initial_lr'] * lf(epoch) * g # gain rises from 0 - 1
# Plot images with bounding boxes # Plot images with bounding boxes
if ni < 1: if ni < 1:
@ -308,7 +307,7 @@ def train():
batch_size=batch_size * 2, batch_size=batch_size * 2,
img_size=img_size_test, img_size=img_size_test,
model=model, model=model,
conf_thres=0.001, # 0.001 if opt.evolve or (final_epoch and is_coco) else 0.01, conf_thres=0.001 if final_epoch else 0.01, # 0.001 for best mAP, 0.01 for speed
iou_thres=0.6, iou_thres=0.6,
save_json=final_epoch and is_coco, save_json=final_epoch and is_coco,
single_cls=opt.single_cls, single_cls=opt.single_cls,

6
utils/utils.py
View File

@ -406,7 +406,7 @@ def compute_loss(p, targets, model): # predictions, targets, model
pbox = torch.cat((pxy, pwh), 1) # predicted box pbox = torch.cat((pxy, pwh), 1) # predicted box
giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True) # giou computation giou = bbox_iou(pbox.t(), tbox[i], x1y1x2y2=False, GIoU=True) # giou computation
lbox += (1.0 - giou).sum() if red == 'sum' else (1.0 - giou).mean() # giou loss lbox += (1.0 - giou).sum() if red == 'sum' else (1.0 - giou).mean() # giou loss
tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().clamp(0).type(tobj.dtype) # giou ratio tobj[b, a, gj, gi] = (1.0 - model.gr) + model.gr * giou.detach().type(tobj.dtype) # giou ratio
if 'default' in arc and model.nc > 1: # cls loss (only if multiple classes) if 'default' in arc and model.nc > 1: # cls loss (only if multiple classes)
t = torch.zeros_like(ps[:, 5:]) + cn # targets t = torch.zeros_like(ps[:, 5:]) + cn # targets
@ -563,10 +563,10 @@ def non_max_suppression(prediction, conf_thres=0.1, iou_thres=0.6, multi_label=T
if batched: if batched:
c = pred[:, 5] * 0 if agnostic else pred[:, 5] # class-agnostic NMS c = pred[:, 5] * 0 if agnostic else pred[:, 5] # class-agnostic NMS
boxes, scores = pred[:, :4].clone(), pred[:, 4] boxes, scores = pred[:, :4].clone(), pred[:, 4]
boxes += c.view(-1, 1) * max_wh
if method == 'vision_batch': if method == 'vision_batch':
i = torchvision.ops.boxes.batched_nms(boxes, scores, c, iou_thres) i = torchvision.ops.boxes.nms(boxes, scores, iou_thres)
elif method == 'fast_batch': # FastNMS from https://github.com/dbolya/yolact elif method == 'fast_batch': # FastNMS from https://github.com/dbolya/yolact
boxes += c.view(-1, 1) * max_wh
iou = box_iou(boxes, boxes).triu_(diagonal=1) # upper triangular iou matrix iou = box_iou(boxes, boxes).triu_(diagonal=1) # upper triangular iou matrix
i = iou.max(dim=0)[0] < iou_thres i = iou.max(dim=0)[0] < iou_thres