def train():
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
print("dataset:", dataset[0])
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
# Parallel wraps the underlying module, but when saving and loading we don't want that
yolact_net = Yolact()
net = yolact_net
net.train()
if args.log:
log = Log(cfg.name,
args.log_folder,
dict(args._get_kwargs()),
overwrite=(args.resume is None),
log_gpu_stats=args.log_gpu)
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs, so disable it just to be safe.
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=cfg.ohem_negpos_ratio)
if args.batch_alloc is not None:
args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')]
if sum(args.batch_alloc) != args.batch_size:
print(
'Error: Batch allocation (%s) does not sum to batch size (%s).'
% (args.batch_alloc, args.batch_size))
exit(-1)
net = CustomDataParallel(NetLoss(net, criterion))
if args.cuda:
net = net.cuda()
# Initialize everything
if not cfg.freeze_bn:
yolact_net.freeze_bn() # Freeze bn so we don't kill our means
yolact_net(torch.zeros(1, 3, cfg.max_size, cfg.max_size).cuda())
if not cfg.freeze_bn: yolact_net.freeze_bn(True)
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
num_epochs = math.ceil(cfg.max_iter / epoch_size)
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
data_loader = data.DataLoader(dataset,
args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=detection_collate,
pin_memory=True)
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
global loss_types # Forms the print order
loss_avgs = {k: MovingAverage(100) for k in loss_types}
print('Begin training!')
print()
# try-except so you can use ctrl+c to save early and stop training
try:
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch + 1) * epoch_size < iteration:
continue
for datum in data_loader:
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch + 1) * epoch_size:
break
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [
x for x in cfg.delayed_settings if x[0] > iteration
]
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until:
set_lr(optimizer, (args.lr - cfg.lr_warmup_init) *
(iteration / cfg.lr_warmup_until) +
cfg.lr_warmup_init)
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while step_index < len(
cfg.lr_steps
) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer, args.lr * (args.gamma**step_index))
# Zero the grad to get ready to compute gradients
optimizer.zero_grad()
# Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss)
losses = net(datum)
losses = {k: (v).mean()
for k, v in losses.items()
} # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# no_inf_mean removes some components from the loss, so make sure to backward through all of it
# all_loss = sum([v.mean() for v in losses.values()])
# Backprop
loss.backward(
) # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(
datetime.timedelta(seconds=(cfg.max_iter - iteration) *
time_avg.get_avg())).split('.')[0]
total = sum([loss_avgs[k].get_avg() for k in losses])
loss_labels = sum([[k, loss_avgs[k].get_avg()]
for k in loss_types if k in losses], [])
print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) +
' T: %.3f || ETA: %s || timer: %.3f') %
tuple([epoch, iteration] + loss_labels +
[total, eta_str, elapsed]),
flush=True)
if args.log:
precision = 5
loss_info = {
k: round(losses[k].item(), precision)
for k in losses
}
loss_info['T'] = round(losses[k].item(), precision)
if args.log_gpu:
log.log_gpu_stats = (iteration % 10 == 0
) # nvidia-smi is sloooow
log.log('train',
loss=loss_info,
epoch=epoch,
iter=iteration,
lr=round(cur_lr, 10),
elapsed=elapsed)
log.log_gpu_stats = args.log_gpu
iteration += 1
if iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder,
cfg.name)
print('Saving state, iter:', iteration)
yolact_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
print('Deleting old save...')
os.remove(latest)
# This is done per epoch
if args.validation_epoch > 0:
if epoch % args.validation_epoch == 0 and epoch > 0:
compute_validation_map(epoch, iteration, yolact_net,
val_dataset,
log if args.log else None)
# Compute validation mAP after training is finished
compute_validation_map(epoch, iteration, yolact_net, val_dataset,
log if args.log else None)
except KeyboardInterrupt:
if args.interrupt:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
yolact_net.save_weights(
save_path(epoch,
repr(iteration) + '_interrupt'))
exit()
yolact_net.save_weights(save_path(epoch, iteration))
def train():
#1: train 결과를 저장할 폴더를 생성
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
#2: MSCOCO에서 제공하는 API를 통해 train dataset을 준비한다.
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
# 만약 train-validation기법을 사용한다면, eval dataset도 준비한다.
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
#3: 구현한 yolact() class의 객체를 만들고 train모드로 설정.
#주의 : net과 yolact_net은 메모리에 저장된 같은 객체를 공유한다.
# 다만 net은 이후에 yolact와 MultiBoxLoss가 결함되어 train을 위한
# 통합된 객체로 다시 정의되기 때문에 yolact넷 객체에만 따로 접근하기 위해
# yolact_net을 deep copy본으로 가지고 있는다.
yolact_net = Yolact()
net = yolact_net
net.train()
#######################################################################
#######RESUME 관련#####################################################
#4: args.log와 args.resume은 train도중 log를 남기는 것과, train이
# 불가피하게 중도에 정지되었을 경우, 중단 지점부터 재시작할 수 있도록
# 기능을 만든 것이므로 필요한 경우에만 더 자세히 보도록 하자.
if args.log:
log = Log(cfg.name,
args.log_folder,
dict(args._get_kwargs()),
overwrite=(args.resume is None),
log_gpu_stats=args.log_gpu)
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs, so disable it just to be safe.
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
#######END#############################################################
#######################################################################
#5: yolact의 optimizer와 loss함수를 설정한다.
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=cfg.ohem_negpos_ratio)
#6: 멀티 GPU를 사용하는 경우 각 GPU에 batch size를 분할해준다.
# 만약 총 Batch size가 맞지 않으면 뭔가 잘못된 것이므로 프로그램 종료.
if args.batch_alloc is not None:
args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')]
if sum(args.batch_alloc) != args.batch_size:
print(
'Error: Batch allocation (%s) does not sum to batch size (%s).'
% (args.batch_alloc, args.batch_size))
exit(-1)
#7: 현재까지 설정된 net과 loss 함수를 엮어 더 통합된 net으로 만듬.
# 이제 net을 호출하면, bbox를 detection하고, fast nms를 거쳐 한 번
# 필터링을 한 후, ground truth와 비교하여 loss를 계산하고, 이 과정을
# 멀티 GPU일 경우 알아서 각 device에 작업을 분할해준다.
# yolact_net은 net에 포함된 yolact()만을 가리킨다.
net = CustomDataParallel(NetLoss(net, criterion))
if args.cuda:
net = net.cuda()
#8: yolact_net의 batch_normalization layer를 모두 false로 만든 뒤에
# 0만을 가지고 있는 zero_tensor를 모델에 통과시켜, 파라미터를 초기화시켜준다.
# 그 후에 다시 batch_normalization layer를 train모드로 바꿔준다.
# 굳이 이런 과정을 거치는 이유는 저자가 batch_normalization에 미리 넣어놓은
# 평균/분산 값은 초기화하고 싶지 않기 때문이다.
if not cfg.freeze_bn:
yolact_net.freeze_bn() # Freeze bn so we don't kill our means
(torch.zeros(1, 3, cfg.max_size, cfg.max_size).cuda())
if not cfg.freeze_bn: yolact_net.freeze_bn(True)
#9: loss counters
# bbox의 위치에 대한 loss와, class confidence에 대한 loss 를 담을 변수를 생성하고,
# batch_size와 dataset의 크기에 맞는 1 epoch의 size와 몇 epoch를 돌려야하는지 구한다.
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0) #cw : 음수입력을 허용치 않기 위해... GOOD
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
num_epochs = math.ceil(cfg.max_iter / epoch_size)
#10:Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
# step_index는 learning rate decay를 위해 사용하는 index이다.
# data_loader는 train중에 순서대로 데이터셋을 준비해서 넘겨주는 class이다.
# 여기서 객체를 만들어 저장한다.
step_index = 0
data_loader = data.DataLoader(dataset,
args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=detection_collate,
pin_memory=True)
#11:특정 epoch와 iteration에 도달했을 때, 중간 과정을 save_path에 저장하기 위한
# 람다 함수를 정의하고, time_avg와 loss_avg는 MovingAverage 클래스의 객체로써
# 훈련 중간 과정의 loss를 이동평균 값으로 보여주기 위해 선언되는 객체이다.
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
global loss_types # Forms the print order
loss_avgs = {k: MovingAverage(100) for k in loss_types}
#12: main train이 시작되는 부분(#A ~ #F)
print('Begin training!')
print()
# A
# try-except를 사용하여 ctrl+c(keyboardInterrupt)를 통해
# 훈련을 중단하고 진행내용은 저장할 수 있다.
# 중단지점부터 재시작하고 싶으면 train.py실행 시 --resume인자를 사용한다.
try:
#9에서 계산된 num_epochs만큼 반복.
for epoch in range(num_epochs):
# B
# --resume을 이용해 시작했다면, 재시작 iter에 도달할 때까지 continue,
# 또한 data_loader에서 data를 불러오며 loss를 계산하는데,
# 도중에 목표 iteration에 도달했으면 break하여 1 epoch를 종료한다.
if (epoch + 1) * epoch_size < iteration:
continue
for datum in data_loader:
# 목표한만큼 훈련이 되었다면, 종료한다.
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch + 1) * epoch_size:
break
# 목표로 설정된 반복횟수가 max_iter보다 크면 max_iter에서 훈련을 마친다.
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
# 특정 iteration에 config값이 바뀌도록 할 경우의 작업을 수행한다.
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [
x for x in cfg.delayed_settings if x[0] > iteration
]
# C
# [learning rate 조정]
# train시작한지 얼마 안되었을 경우(lr_warmup_until기준) 훈련을 조금 가속시키기 위해 조정.
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until:
set_lr(optimizer, (args.lr - cfg.lr_warmup_init) *
(iteration / cfg.lr_warmup_until) +
cfg.lr_warmup_init)
# 특정 iteration에 도달할 때마다 learning rate decay수행.
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while step_index < len(
cfg.lr_steps
) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer, args.lr * (args.gamma**step_index))
# D
# loss 함수 계산.
# Zero the grad to get ready to compute gradients
optimizer.zero_grad()
# Forward Propagation을 수행하고 수행 결과로 loss 함수를 통해 1 iteration의 loss를 계산한다.
# 구체적인 동작은 Backbone.py의 resnet101, yolact.py의 yolact, MultiBoxLoss.py의 MultiBoxLoss 클래스를 모두 보아야 한다.
# (see CustomDataParallel and NetLoss)
losses = net(datum)
losses = {k: (v).mean()
for k, v in losses.items()
} # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# no_inf_mean removes some components from the loss, so make sure to backward through all of it
# all_loss = sum([v.mean() for v in losses.values()])
# E
# Backward Propagation을 수행하고,
# 계산가능한 값일 경우, optimizer.step()을 통해 parameters에 적용
# Backprop
loss.backward()
# Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# F
# train진행 과정에서 소요 시간과, 중간 loss값을 출력하여 중간 성과를
# 파악 할 수 있도록 해주는 파트.
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(
datetime.timedelta(seconds=(cfg.max_iter - iteration) *
time_avg.get_avg())).split('.')[0]
total = sum([loss_avgs[k].get_avg() for k in losses])
loss_labels = sum([[k, loss_avgs[k].get_avg()]
for k in loss_types if k in losses], [])
print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) +
' T: %.3f || ETA: %s || timer: %.3f') %
tuple([epoch, iteration] + loss_labels +
[total, eta_str, elapsed]),
flush=True)
# log를 파일로 기록
if args.log:
precision = 5
loss_info = {
k: round(losses[k].item(), precision)
for k in losses
}
loss_info['T'] = round(loss.item(), precision)
if args.log_gpu:
log.log_gpu_stats = (iteration % 10 == 0
) # nvidia-smi is sloooow
log.log('train',
loss=loss_info,
epoch=epoch,
iter=iteration,
lr=round(cur_lr, 10),
elapsed=elapsed)
log.log_gpu_stats = args.log_gpu
# ~F
# 1번 반복하면, 1 iter증가.
iteration += 1
# 주기마다 진행과정을 저장하는 작업 수행.
if iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder,
cfg.name)
print('Saving state, iter:', iteration)
yolact_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
print('Deleting old save...')
os.remove(latest)
# train-validation으로 작업을 수행하는 경우,
# 1 epoch를 돌렸을 때 validation 주기에 도달한 epoch였으면 validate 1회 진행하여 mAP측정.
if args.validation_epoch > 0:
if epoch % args.validation_epoch == 0 and epoch > 0:
compute_validation_map(epoch, iteration, yolact_net,
val_dataset,
log if args.log else None)
# Compute validation mAP after training is finished
compute_validation_map(epoch, iteration, yolact_net, val_dataset,
log if args.log else None)
#13: Ctrl + c를 이용하여 훈련을 중단했을 경우, save_foler에 weights를 저장하고 중단하여
# 다음에 다시 재시작할 수 있도록 한다.
except KeyboardInterrupt:
if args.interrupt:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
yolact_net.save_weights(
save_path(epoch,
repr(iteration) + '_interrupt'))
exit()
yolact_net.save_weights(save_path(epoch, iteration))
print(make_row([iou_type] + ['%.2f' % x if x < 100 else '%.1f' % x for x in all_maps[iou_type].values()]))
print(make_sep(len(all_maps['box']) + 1))
print()
if __name__ == '__main__':
parse_args()
if args.config is not None:
set_cfg(args.config)
if args.trained_model == 'interrupt':
args.trained_model = SavePath.get_interrupt('weights/')
elif args.trained_model == 'latest':
args.trained_model = SavePath.get_latest('weights/', cfg.name)
if args.config is None:
model_path = SavePath.from_str(args.trained_model)
# TODO: Bad practice? Probably want to do a name lookup instead.
args.config = model_path.model_name + '_config'
print('Config not specified. Parsed %s from the file name.\n' % args.config)
set_cfg(args.config)
if args.detect:
cfg.eval_mask_branch = False
if args.dataset is not None:
set_dataset(args.dataset)
with torch.no_grad():
def train(rank, args):
if args.num_gpus > 1:
multi_gpu_rescale(args)
if rank == 0:
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
# set up logger
setup_logger(output=os.path.join(args.log_folder, cfg.name),
distributed_rank=rank)
logger = logging.getLogger("yolact.train")
w = SummaryHelper(distributed_rank=rank,
log_dir=os.path.join(args.log_folder, cfg.name))
w.add_text("argv", " ".join(sys.argv))
logger.info("Args: {}".format(" ".join(sys.argv)))
import git
with git.Repo(search_parent_directories=True) as repo:
w.add_text("git_hash", repo.head.object.hexsha)
logger.info("git hash: {}".format(repo.head.object.hexsha))
try:
logger.info("Initializing torch.distributed backend...")
dist.init_process_group(backend='nccl',
init_method=args.dist_url,
world_size=args.num_gpus,
rank=rank)
except Exception as e:
logger.error("Process group URL: {}".format(args.dist_url))
raise e
dist.barrier()
if torch.cuda.device_count() > 1:
logger.info('Multiple GPUs detected! Turning off JIT.')
collate_fn = detection_collate
if cfg.dataset.name == 'YouTube VIS':
dataset = YoutubeVIS(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
configs=cfg.dataset,
transform=SSDAugmentationVideo(MEANS))
if cfg.dataset.joint == 'coco':
joint_dataset = COCODetection(
image_path=cfg.joint_dataset.train_images,
info_file=cfg.joint_dataset.train_info,
transform=SSDAugmentation(MEANS))
joint_collate_fn = detection_collate
if args.validation_epoch > 0:
setup_eval()
val_dataset = YoutubeVIS(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
configs=cfg.dataset,
transform=BaseTransformVideo(MEANS))
collate_fn = collate_fn_youtube_vis
elif cfg.dataset.name == 'FlyingChairs':
dataset = FlyingChairs(image_path=cfg.dataset.trainval_images,
info_file=cfg.dataset.trainval_info)
collate_fn = collate_fn_flying_chairs
else:
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
# Set cuda device early to avoid duplicate model in master GPU
if args.cuda:
torch.cuda.set_device(rank)
# Parallel wraps the underlying module, but when saving and loading we don't want that
yolact_net = Yolact()
net = yolact_net
net.train()
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs.
# use timer for experiments
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
logger.info('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume, args=args)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
logger.info('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
if cfg.flow.train_flow:
criterion = OpticalFlowLoss()
else:
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=3)
if args.cuda:
cudnn.benchmark = True
net.cuda(rank)
criterion.cuda(rank)
net = nn.parallel.DistributedDataParallel(net,
device_ids=[rank],
output_device=rank,
broadcast_buffers=False,
find_unused_parameters=True)
# net = nn.DataParallel(net).cuda()
# criterion = nn.DataParallel(criterion).cuda()
optimizer = optim.SGD(filter(lambda x: x.requires_grad, net.parameters()),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
w.set_step(iteration)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size // args.num_gpus
num_epochs = math.ceil(cfg.max_iter / epoch_size)
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
from data.sampler_utils import InfiniteSampler, build_batch_data_sampler
infinite_sampler = InfiniteSampler(dataset,
seed=args.random_seed,
num_replicas=args.num_gpus,
rank=rank,
shuffle=True)
train_sampler = build_batch_data_sampler(infinite_sampler,
images_per_batch=args.batch_size)
data_loader = data.DataLoader(
dataset,
num_workers=args.num_workers,
collate_fn=collate_fn,
multiprocessing_context="fork" if args.num_workers > 1 else None,
batch_sampler=train_sampler)
data_loader_iter = iter(data_loader)
if cfg.dataset.joint:
joint_infinite_sampler = InfiniteSampler(joint_dataset,
seed=args.random_seed,
num_replicas=args.num_gpus,
rank=rank,
shuffle=True)
joint_train_sampler = build_batch_data_sampler(
joint_infinite_sampler, images_per_batch=args.batch_size)
joint_data_loader = data.DataLoader(
joint_dataset,
num_workers=args.num_workers,
collate_fn=joint_collate_fn,
multiprocessing_context="fork" if args.num_workers > 1 else None,
batch_sampler=joint_train_sampler)
joint_data_loader_iter = iter(joint_data_loader)
dist.barrier()
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
data_time_avg = MovingAverage(10)
global loss_types # Forms the print order
loss_avgs = {k: MovingAverage(100) for k in loss_types}
def backward_and_log(prefix,
net_outs,
targets,
masks,
num_crowds,
extra_loss=None):
optimizer.zero_grad()
out = net_outs["pred_outs"]
wrapper = ScatterWrapper(targets, masks, num_crowds)
losses = criterion(out, wrapper, wrapper.make_mask())
losses = {k: v.mean()
for k, v in losses.items()} # Mean here because Dataparallel
if extra_loss is not None:
assert type(extra_loss) == dict
losses.update(extra_loss)
loss = sum([losses[k] for k in losses])
# Backprop
loss.backward() # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
w.add_scalar('{prefix}/{key}'.format(prefix=prefix, key=k),
losses[k].item())
return losses
logger.info('Begin training!')
# try-except so you can use ctrl+c to save early and stop training
try:
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch + 1) * epoch_size < iteration:
continue
while True:
data_start_time = time.perf_counter()
datum = next(data_loader_iter)
dist.barrier()
data_end_time = time.perf_counter()
data_time = data_end_time - data_start_time
if iteration != args.start_iter:
data_time_avg.add(data_time)
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch + 1) * epoch_size:
break
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [
x for x in cfg.delayed_settings if x[0] > iteration
]
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until and cfg.lr_warmup_init < args.lr:
set_lr(optimizer, (args.lr - cfg.lr_warmup_init) *
(iteration / cfg.lr_warmup_until) +
cfg.lr_warmup_init)
elif cfg.lr_schedule == 'cosine':
set_lr(
optimizer,
args.lr *
((math.cos(math.pi * iteration / cfg.max_iter) + 1.) *
.5))
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while cfg.lr_schedule == 'step' and step_index < len(
cfg.lr_steps
) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer, args.lr * (args.gamma**step_index))
global lr
w.add_scalar('meta/lr', lr)
if cfg.dataset.name == "FlyingChairs":
imgs_1, imgs_2, flows = prepare_flow_data(datum)
net_outs = net(None, extras=(imgs_1, imgs_2))
# Compute Loss
optimizer.zero_grad()
losses = criterion(net_outs, flows)
losses = {k: v.mean()
for k, v in losses.items()
} # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# Backprop
loss.backward(
) # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
w.add_scalar('loss/%s' % k, losses[k].item())
elif cfg.dataset.joint or not cfg.dataset.is_video:
if cfg.dataset.joint:
joint_datum = next(joint_data_loader_iter)
dist.barrier()
# Load training data
# Note, for training on multiple gpus this will use the custom replicate and gather I wrote up there
images, targets, masks, num_crowds = prepare_data(
joint_datum)
else:
images, targets, masks, num_crowds = prepare_data(
datum)
extras = {
"backbone": "full",
"interrupt": False,
"moving_statistics": {
"aligned_feats": []
}
}
net_outs = net(images, extras=extras)
out = net_outs["pred_outs"]
# Compute Loss
optimizer.zero_grad()
wrapper = ScatterWrapper(targets, masks, num_crowds)
losses = criterion(out, wrapper, wrapper.make_mask())
losses = {k: v.mean()
for k, v in losses.items()
} # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# Backprop
loss.backward(
) # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
w.add_scalar('joint/%s' % k, losses[k].item())
# Forward Pass
if cfg.dataset.is_video:
# reference frames
references = []
moving_statistics = {"aligned_feats": [], "conf_hist": []}
for idx, frame in enumerate(datum[:0:-1]):
images, annots = frame
extras = {
"backbone": "full",
"interrupt": True,
"keep_statistics": True,
"moving_statistics": moving_statistics
}
with torch.no_grad():
net_outs = net(images, extras=extras)
moving_statistics["feats"] = net_outs["feats"]
moving_statistics["lateral"] = net_outs["lateral"]
keys_to_save = ("outs_phase_1", "outs_phase_2")
for key in set(net_outs.keys()) - set(keys_to_save):
del net_outs[key]
references.append(net_outs)
# key frame with annotation, but not compute full backbone
frame = datum[0]
images, annots = frame
frame = (
images,
annots,
)
images, targets, masks, num_crowds = prepare_data(frame)
extras = {
"backbone": "full",
"interrupt": not cfg.flow.base_backward,
"moving_statistics": moving_statistics
}
gt_net_outs = net(images, extras=extras)
if cfg.flow.base_backward:
losses = backward_and_log("compute", gt_net_outs,
targets, masks, num_crowds)
keys_to_save = ("outs_phase_1", "outs_phase_2")
for key in set(gt_net_outs.keys()) - set(keys_to_save):
del gt_net_outs[key]
# now do the warp
if len(references) > 0:
reference_frame = references[0]
extras = {
"backbone": "partial",
"moving_statistics": moving_statistics
}
net_outs = net(images, extras=extras)
extra_loss = yolact_net.extra_loss(
net_outs, gt_net_outs)
losses = backward_and_log("warp",
net_outs,
targets,
masks,
num_crowds,
extra_loss=extra_loss)
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
w.add_scalar('meta/data_time', data_time)
w.add_scalar('meta/iter_time', elapsed)
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(
datetime.timedelta(seconds=(cfg.max_iter - iteration) *
time_avg.get_avg())).split('.')[0]
if torch.cuda.is_available():
max_mem_mb = torch.cuda.max_memory_allocated(
) / 1024.0 / 1024.0
# torch.cuda.reset_max_memory_allocated()
else:
max_mem_mb = None
logger.info("""\
eta: {eta} epoch: {epoch} iter: {iter} \
{losses} {loss_total} \
time: {time} data_time: {data_time} lr: {lr} {memory}\
""".format(eta=eta_str,
epoch=epoch,
iter=iteration,
losses=" ".join([
"{}: {:.3f}".format(k, loss_avgs[k].get_avg()) for k in losses
]),
loss_total="T: {:.3f}".format(
sum([loss_avgs[k].get_avg() for k in losses])),
data_time="{:.3f}".format(data_time_avg.get_avg()),
time="{:.3f}".format(elapsed),
lr="{:.6f}".format(lr),
memory="max_mem: {:.0f}M".format(max_mem_mb)))
if rank == 0 and iteration % 100 == 0:
if cfg.flow.train_flow:
import flowiz as fz
from layers.warp_utils import deform_op
tgt_size = (64, 64)
flow_size = flows.size()[2:]
vis_data = []
for pred_flow in net_outs:
vis_data.append(pred_flow)
deform_gt = deform_op(imgs_2, flows)
flows_pred = [
F.interpolate(x,
size=flow_size,
mode='bilinear',
align_corners=False)
for x in net_outs
]
deform_preds = [
deform_op(imgs_2, x) for x in flows_pred
]
vis_data.append(
F.interpolate(flows, size=tgt_size, mode='area'))
vis_data = [
F.interpolate(flow[:1], size=tgt_size)
for flow in vis_data
]
vis_data = [
fz.convert_from_flow(
flow[0].data.cpu().numpy().transpose(
1, 2, 0)).transpose(
2, 0, 1).astype('float32') / 255
for flow in vis_data
]
def convert_image(image):
image = F.interpolate(image,
size=tgt_size,
mode='area')
image = image[0]
image = image.data.cpu().numpy()
image = image[::-1]
image = image.transpose(1, 2, 0)
image = image * np.array(STD) + np.array(MEANS)
image = image.transpose(2, 0, 1)
image = image / 255
image = np.clip(image, -1, 1)
image = image[::-1]
return image
vis_data.append(convert_image(imgs_1))
vis_data.append(convert_image(imgs_2))
vis_data.append(convert_image(deform_gt))
vis_data.extend(
[convert_image(x) for x in deform_preds])
vis_data_stack = np.stack(vis_data, axis=0)
w.add_images("preds_flow", vis_data_stack)
elif cfg.flow.warp_mode == "flow":
import flowiz as fz
tgt_size = (64, 64)
vis_data = []
for pred_flow, _, _ in net_outs["preds_flow"]:
vis_data.append(pred_flow)
vis_data = [
F.interpolate(flow[:1], size=tgt_size)
for flow in vis_data
]
vis_data = [
fz.convert_from_flow(
flow[0].data.cpu().numpy().transpose(
1, 2, 0)).transpose(
2, 0, 1).astype('float32') / 255
for flow in vis_data
]
input_image = F.interpolate(images,
size=tgt_size,
mode='area')
input_image = input_image[0]
input_image = input_image.data.cpu().numpy()
input_image = input_image.transpose(1, 2, 0)
input_image = input_image * np.array(
STD[::-1]) + np.array(MEANS[::-1])
input_image = input_image.transpose(2, 0, 1)
input_image = input_image / 255
input_image = np.clip(input_image, -1, 1)
vis_data.append(input_image)
vis_data_stack = np.stack(vis_data, axis=0)
w.add_images("preds_flow", vis_data_stack)
iteration += 1
w.set_step(iteration)
if rank == 0 and iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder,
cfg.name)
logger.info('Saving state, iter: {}'.format(iteration))
yolact_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
logger.info('Deleting old save...')
os.remove(latest)
# This is done per epoch
if args.validation_epoch > 0:
if epoch % args.validation_epoch == 0 and epoch > 0:
if rank == 0:
compute_validation_map(yolact_net, val_dataset)
dist.barrier()
except KeyboardInterrupt:
if args.interrupt_no_save:
logger.info('No save on interrupt, just exiting...')
elif rank == 0:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
yolact_net.save_weights(
save_path(epoch,
repr(iteration) + '_interrupt'))
return
if rank == 0:
yolact_net.save_weights(save_path(epoch, iteration))
def train():
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
# Parallel wraps the underlying module, but when saving and loading we don't want that
yolact_net = Yolact()
net = yolact_net
net.train()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=3)
if args.cuda:
cudnn.benchmark = True
net = nn.DataParallel(net).cuda()
criterion = nn.DataParallel(criterion).cuda()
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
num_epochs = math.ceil(cfg.max_iter / epoch_size)
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
data_loader = data.DataLoader(dataset,
args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=detection_collate,
pin_memory=True)
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
global loss_types # Forms the print order
loss_avgs = {k: MovingAverage(100) for k in loss_types}
print('Begin training!')
print()
# try-except so you can use ctrl+c to save early and stop training
try:
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch + 1) * epoch_size < iteration:
continue
for datum in data_loader:
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch + 1) * epoch_size:
break
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [
x for x in cfg.delayed_settings if x[0] > iteration
]
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until:
set_lr(optimizer, (args.lr - cfg.lr_warmup_init) *
(iteration / cfg.lr_warmup_until) +
cfg.lr_warmup_init)
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while step_index < len(
cfg.lr_steps
) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer, args.lr * (args.gamma**step_index))
# Load training data
# Note, for training on multiple gpus this will use the custom replicate and gather I wrote up there
images, targets, masks, num_crowds = prepare_data(datum)
# Forward Pass
out = net(images)
# Compute Loss
optimizer.zero_grad()
wrapper = ScatterWrapper(targets, masks, num_crowds)
losses = criterion(out, wrapper, wrapper.make_mask())
losses = {k: v.mean()
for k, v in losses.items()
} # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# Backprop
loss.backward(
) # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer.step()
# Add the loss to the moving average for bookkeeping
for k in losses:
loss_avgs[k].add(losses[k].item())
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(
datetime.timedelta(seconds=(cfg.max_iter - iteration) *
time_avg.get_avg())).split('.')[0]
total = sum([loss_avgs[k].get_avg() for k in losses])
loss_labels = sum([[k, loss_avgs[k].get_avg()]
for k in loss_types if k in losses], [])
print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) +
' T: %.3f || ETA: %s || timer: %.3f') %
tuple([epoch, iteration] + loss_labels +
[total, eta_str, elapsed]),
flush=True)
iteration += 1
if iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder,
cfg.name)
print('Saving state, iter:', iteration)
yolact_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
print('Deleting old save...')
os.remove(latest)
# This is done per epoch
if args.validation_epoch > 0:
if epoch % args.validation_epoch == 0 and epoch > 0:
compute_validation_map(yolact_net, val_dataset)
except KeyboardInterrupt:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
yolact_net.save_weights(
save_path(epoch,
repr(iteration) + '_interrupt'))
exit()
yolact_net.save_weights(save_path(epoch, iteration))
def train():
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
# Parallel wraps the underlying module, but when saving and loading we don't want that
yolact_net = Yolact()
net = yolact_net
net.train()
print('\n--- Generator created! ---')
# NOTE
# I maunally set the original image size and seg size as 138
# might change in the future, for example 550
if cfg.pred_seg:
dis_size = 138
dis_net = Discriminator_Wgan(i_size = dis_size, s_size = dis_size)
# Change the initialization inside the dis_net class inside
# set the dis net's initial parameter values
# dis_net.apply(gan_init)
dis_net.train()
print('--- Discriminator created! ---\n')
if args.log:
log = Log(cfg.name, args.log_folder, dict(args._get_kwargs()),
overwrite=(args.resume is None), log_gpu_stats=args.log_gpu)
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs, so disable it just to be safe.
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path)
# optimizer_gen = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum,
# weight_decay=args.decay)
# if cfg.pred_seg:
# optimizer_dis = optim.SGD(dis_net.parameters(), lr=cfg.dis_lr, momentum=args.momentum,
# weight_decay=args.decay)
# schedule_dis = ReduceLROnPlateau(optimizer_dis, mode = 'min', patience=6, min_lr=1E-6)
# NOTE: Using the Ranger Optimizer for the generator
optimizer_gen = Ranger(net.parameters(), lr = args.lr, weight_decay=args.decay)
# optimizer_gen = optim.RMSprop(net.parameters(), lr = args.lr)
# FIXME: Might need to modify the lr in the optimizer carefually
# check this
# def make_D_optimizer(cfg, model):
# params = []
# for key, value in model.named_parameters():
# if not value.requires_grad:
# continue
# lr = cfg.SOLVER.BASE_LR/5.0
# weight_decay = cfg.SOLVER.WEIGHT_DECAY
# if "bias" in key:
# lr = cfg.SOLVER.BASE_LR * cfg.SOLVER.BIAS_LR_FACTOR/5.0
# weight_decay = cfg.SOLVER.WEIGHT_DECAY_BIAS
# params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]
# optimizer = torch.optim.SGD(params, lr, momentum=cfg.SOLVER.MOMENTUM)
# return optimizer
if cfg.pred_seg:
optimizer_dis = optim.SGD(dis_net.parameters(), lr=cfg.dis_lr)
# optimizer_dis = optim.RMSprop(dis_net.parameters(), lr = cfg.dis_lr)
schedule_dis = ReduceLROnPlateau(optimizer_dis, mode = 'min', patience=6, min_lr=1E-6)
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=cfg.ohem_negpos_ratio, pred_seg=cfg.pred_seg)
# criterion_dis = nn.BCELoss()
# Take the advice from WGAN
criterion_dis = DiscriminatorLoss_Maskrcnn()
criterion_gen = GeneratorLoss_Maskrcnn()
if args.batch_alloc is not None:
# e.g. args.batch_alloc: 24,24
args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')]
if sum(args.batch_alloc) != args.batch_size:
print('Error: Batch allocation (%s) does not sum to batch size (%s).' % (args.batch_alloc, args.batch_size))
exit(-1)
net = CustomDataParallel(NetLoss(net, criterion, pred_seg=cfg.pred_seg))
if args.cuda:
net = net.cuda()
# NOTE
if cfg.pred_seg:
dis_net = nn.DataParallel(dis_net)
dis_net = dis_net.cuda()
# Initialize everything
if not cfg.freeze_bn: yolact_net.freeze_bn() # Freeze bn so we don't kill our means
yolact_net(torch.zeros(1, 3, cfg.max_size, cfg.max_size).cuda())
if not cfg.freeze_bn: yolact_net.freeze_bn(True)
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
num_epochs = math.ceil(cfg.max_iter / epoch_size)
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
data_loader = data.DataLoader(dataset, args.batch_size,
num_workers=args.num_workers,
shuffle=True, collate_fn=detection_collate,
pin_memory=True)
# NOTE
val_loader = data.DataLoader(val_dataset, args.batch_size,
num_workers=args.num_workers*2,
shuffle=True, collate_fn=detection_collate,
pin_memory=True)
save_path = lambda epoch, iteration: SavePath(cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
global loss_types # Forms the print order
# TODO: global command can modify global variable inside of the function.
loss_avgs = { k: MovingAverage(100) for k in loss_types }
# NOTE
# Enable AMP
amp_enable = cfg.amp
scaler = torch.cuda.amp.GradScaler(enabled=amp_enable)
print('Begin training!')
print()
# try-except so you can use ctrl+c to save early and stop training
try:
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch+1)*epoch_size < iteration:
continue
for datum in data_loader:
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch+1)*epoch_size:
break
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [x for x in cfg.delayed_settings if x[0] > iteration]
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until:
set_lr(optimizer_gen, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init)
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while step_index < len(cfg.lr_steps) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer_gen, args.lr * (args.gamma ** step_index))
# NOTE
if cfg.pred_seg:
# ====== GAN Train ======
# train the gen and dis in different iteration
# it_alter_period = iteration % (cfg.gen_iter + cfg.dis_iter)
# FIXME:
# present_time = time.time()
for _ in range(cfg.dis_iter):
# freeze_pretrain(yolact_net, freeze=False)
# freeze_pretrain(net, freeze=False)
# freeze_pretrain(dis_net, freeze=False)
# if it_alter_period == 0:
# print('--- Generator freeze ---')
# print('--- Discriminator training ---')
if cfg.amp:
with torch.cuda.amp.autocast():
# ----- Discriminator part -----
# seg_list is the prediction mask
# can be regarded as generated images from YOLACT
# pred_list is the prediction label
# seg_list dim: list of (138,138,instances)
# pred_list dim: list of (instances)
losses, seg_list, pred_list = net(datum)
seg_clas, mask_clas, b, seg_size = seg_mask_clas(seg_list, pred_list, datum)
# input image size is [b, 3, 550, 550]
# downsample to [b, 3, seg_h, seg_w]
image_list = [img.to(cuda0) for img in datum[0]]
image = interpolate(torch.stack(image_list), size = seg_size,
mode='bilinear',align_corners=False)
# Because in the discriminator training, we do not
# want the gradient flow back to the generator part
# we detach seg_clas (mask_clas come the data, does not have grad)
output_pred = dis_net(img = image.detach(), seg = seg_clas.detach())
output_grou = dis_net(img = image.detach(), seg = mask_clas.detach())
# p = elem_mul_p.squeeze().permute(1,2,0).cpu().detach().numpy()
# g = elem_mul_g.squeeze().permute(1,2,0).cpu().detach().numpy()
# image = image.squeeze().permute(1,2,0).cpu().detach().numpy()
# from PIL import Image
# seg_PIL = Image.fromarray(p, 'RGB')
# mask_PIL = Image.fromarray(g, 'RGB')
# seg_PIL.save('mul_seg.png')
# mask_PIL.save('mul_mask.png')
# raise RuntimeError
# from matplotlib import pyplot as plt
# fig, (ax1, ax2) = plt.subplots(1,2)
# ax1.imshow(mask_show)
# ax2.imshow(seg_show)
# plt.show(block=False)
# plt.pause(2)
# plt.close()
# if iteration % (cfg.gen_iter + cfg.dis_iter) == 0:
# print(f'Probability of fake is fake: {output_pred.mean().item():.2f}')
# print(f'Probability of real is real: {output_grou.mean().item():.2f}')
# 0 for Fake/Generated
# 1 for True/Ground Truth
# fake_label = torch.zeros(b)
# real_label = torch.ones(b)
# Advice of practical implementation
# from https://arxiv.org/abs/1611.08408
# loss_pred = -criterion_dis(output_pred,target=real_label)
# loss_pred = criterion_dis(output_pred,target=fake_label)
# loss_grou = criterion_dis(output_grou,target=real_label)
# loss_dis = loss_pred + loss_grou
# Wasserstein Distance (Earth-Mover)
loss_dis = criterion_dis(input=output_grou,target=output_pred)
# Backprop the discriminator
# Scales loss. Calls backward() on scaled loss to create scaled gradients.
scaler.scale(loss_dis).backward()
scaler.step(optimizer_dis)
scaler.update()
optimizer_dis.zero_grad()
# clip the updated parameters
_ = [par.data.clamp_(-cfg.clip_value, cfg.clip_value) for par in dis_net.parameters()]
# ----- Generator part -----
# freeze_pretrain(yolact_net, freeze=False)
# freeze_pretrain(net, freeze=False)
# freeze_pretrain(dis_net, freeze=False)
# if it_alter_period == (cfg.dis_iter+1):
# print('--- Generator training ---')
# print('--- Discriminator freeze ---')
# FIXME:
# print(f'dis time pass: {time.time()-present_time:.2f}')
# FIXME:
# present_time = time.time()
with torch.cuda.amp.autocast():
losses, seg_list, pred_list = net(datum)
seg_clas, mask_clas, b, seg_size = seg_mask_clas(seg_list, pred_list, datum)
image_list = [img.to(cuda0) for img in datum[0]]
image = interpolate(torch.stack(image_list), size = seg_size,
mode='bilinear',align_corners=False)
# Perform forward pass of all-fake batch through D
# NOTE this seg_clas CANNOT detach, in order to flow the
# gradient back to the generator
# output = dis_net(img = image, seg = seg_clas)
# Since the log(1-D(G(x))) not provide sufficient gradients
# We want log(D(G(x)) instead, this can be achieve by
# use the real_label as target.
# This step is crucial for the information of discriminator
# to go into the generator.
# Calculate G's loss based on this output
# real_label = torch.ones(b)
# loss_gen = criterion_dis(output,target=real_label)
# GAN MaskRCNN
output_pred = dis_net(img = image, seg = seg_clas)
output_grou = dis_net(img = image, seg = mask_clas)
# Advice from WGAN
# loss_gen = -torch.mean(output)
loss_gen = criterion_gen(input=output_grou,target=output_pred)
# since the dis is already freeze, the gradients will only
# record the YOLACT
losses = { k: (v).mean() for k,v in losses.items() } # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
loss += loss_gen
# Generator backprop
scaler.scale(loss).backward()
scaler.step(optimizer_gen)
scaler.update()
optimizer_gen.zero_grad()
# FIXME:
# print(f'gen time pass: {time.time()-present_time:.2f}')
# print('GAN part over')
else:
losses, seg_list, pred_list = net(datum)
seg_clas, mask_clas, b, seg_size = seg_mask_clas(seg_list, pred_list, datum)
image_list = [img.to(cuda0) for img in datum[0]]
image = interpolate(torch.stack(image_list), size = seg_size,
mode='bilinear',align_corners=False)
output_pred = dis_net(img = image.detach(), seg = seg_clas.detach())
output_grou = dis_net(img = image.detach(), seg = mask_clas.detach())
loss_dis = criterion_dis(input=output_grou,target=output_pred)
loss_dis.backward()
optimizer_dis.step()
optimizer_dis.zero_grad()
_ = [par.data.clamp_(-cfg.clip_value, cfg.clip_value) for par in dis_net.parameters()]
# ----- Generator part -----
# FIXME:
# print(f'dis time pass: {time.time()-present_time:.2f}')
# FIXME:
# present_time = time.time()
losses, seg_list, pred_list = net(datum)
seg_clas, mask_clas, b, seg_size = seg_mask_clas(seg_list, pred_list, datum)
image_list = [img.to(cuda0) for img in datum[0]]
image = interpolate(torch.stack(image_list), size = seg_size,
mode='bilinear',align_corners=False)
# GAN MaskRCNN
output_pred = dis_net(img = image, seg = seg_clas)
output_grou = dis_net(img = image, seg = mask_clas)
loss_gen = criterion_gen(input=output_grou,target=output_pred)
# since the dis is already freeze, the gradients will only
# record the YOLACT
losses = { k: (v).mean() for k,v in losses.items() } # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
loss += loss_gen
loss.backward()
# Do this to free up vram even if loss is not finite
optimizer_gen.zero_grad()
if torch.isfinite(loss).item():
# since the optimizer_gen is for YOLACT only
# only the gen will be updated
optimizer_gen.step()
# FIXME:
# print(f'gen time pass: {time.time()-present_time:.2f}')
# print('GAN part over')
else:
# ====== Normal YOLACT Train ======
# Zero the grad to get ready to compute gradients
optimizer_gen.zero_grad()
# Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss)
losses = net(datum)
losses = { k: (v).mean() for k,v in losses.items() } # Mean here because Dataparallel
loss = sum([losses[k] for k in losses])
# no_inf_mean removes some components from the loss, so make sure to backward through all of it
# all_loss = sum([v.mean() for v in losses.values()])
# Backprop
loss.backward() # Do this to free up vram even if loss is not finite
if torch.isfinite(loss).item():
optimizer_gen.step()
# Add the loss to the moving average for bookkeeping
_ = [loss_avgs[k].add(losses[k].item()) for k in losses]
# for k in losses:
# loss_avgs[k].add(losses[k].item())
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(datetime.timedelta(seconds=(cfg.max_iter-iteration) * time_avg.get_avg())).split('.')[0]
total = sum([loss_avgs[k].get_avg() for k in losses])
loss_labels = sum([[k, loss_avgs[k].get_avg()] for k in loss_types if k in losses], [])
if cfg.pred_seg:
print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) + ' T: %.3f || ETA: %s || timer: %.3f')
% tuple([epoch, iteration] + loss_labels + [total, eta_str, elapsed]), flush=True)
# print(f'Generator loss: {loss_gen:.2f} | Discriminator loss: {loss_dis:.2f}')
# Loss Key:
# - B: Box Localization Loss
# - C: Class Confidence Loss
# - M: Mask Loss
# - P: Prototype Loss
# - D: Coefficient Diversity Loss
# - E: Class Existence Loss
# - S: Semantic Segmentation Loss
# - T: Total loss
if args.log:
precision = 5
loss_info = {k: round(losses[k].item(), precision) for k in losses}
loss_info['T'] = round(loss.item(), precision)
if args.log_gpu:
log.log_gpu_stats = (iteration % 10 == 0) # nvidia-smi is sloooow
log.log('train', loss=loss_info, epoch=epoch, iter=iteration,
lr=round(cur_lr, 10), elapsed=elapsed)
log.log_gpu_stats = args.log_gpu
iteration += 1
if iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder, cfg.name)
print('Saving state, iter:', iteration)
yolact_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
print('Deleting old save...')
os.remove(latest)
# This is done per epoch
if args.validation_epoch > 0:
# NOTE: Validation loss
# if cfg.pred_seg:
# net.eval()
# dis_net.eval()
# cfg.gan_eval = True
# with torch.no_grad():
# for datum in tqdm(val_loader, desc='GAN Validation'):
# losses, seg_list, pred_list = net(datum)
# losses, seg_list, pred_list = net(datum)
# # TODO: warp below as a function
# seg_list = [v.permute(2,1,0).contiguous() for v in seg_list]
# b = len(seg_list) # batch size
# _, seg_h, seg_w = seg_list[0].size()
# seg_clas = torch.zeros(b, cfg.num_classes-1, seg_h, seg_w)
# mask_clas = torch.zeros(b, cfg.num_classes-1, seg_h, seg_w)
# target_list = [target for target in datum[1][0]]
# mask_list = [interpolate(mask.unsqueeze(0), size = (seg_h,seg_w),mode='bilinear', \
# align_corners=False).squeeze() for mask in datum[1][1]]
# for idx in range(b):
# for i, (pred, i_target) in enumerate(zip(pred_list[idx], target_list[idx])):
# seg_clas[idx, pred, ...] += seg_list[idx][i,...]
# mask_clas[idx, i_target[-1].long(), ...] += mask_list[idx][i,...]
# seg_clas = torch.clamp(seg_clas, 0, 1)
# image = interpolate(torch.stack(datum[0]), size = (seg_h,seg_w),
# mode='bilinear',align_corners=False)
# real_label = torch.ones(b)
# output_pred = dis_net(img = image, seg = seg_clas)
# output_grou = dis_net(img = image, seg = mask_clas)
# loss_pred = -criterion_dis(output_pred,target=real_label)
# loss_grou = criterion_dis(output_grou,target=real_label)
# loss_dis = loss_pred + loss_grou
# losses = { k: (v).mean() for k,v in losses.items() }
# loss = sum([losses[k] for k in losses])
# val_loss = loss - cfg.lambda_dis*loss_dis
# schedule_dis.step(loss_dis)
# lr = [group['lr'] for group in optimizer_dis.param_groups]
# print(f'Discriminator lr: {lr[0]}')
# net.train()
if epoch % args.validation_epoch == 0 and epoch > 0:
cfg.gan_eval = False
dis_net.eval()
compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None)
# Compute validation mAP after training is finished
compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None)
except KeyboardInterrupt:
if args.interrupt:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
yolact_net.save_weights(save_path(epoch, repr(iteration) + '_interrupt'))
exit()
yolact_net.save_weights(save_path(epoch, iteration))
def interpret():
if not os.path.exists(args.save_folder):
os.mkdir(args.save_folder)
dataset = COCODetection(image_path=cfg.dataset.train_images,
info_file=cfg.dataset.train_info,
transform=SSDAugmentation(MEANS))
if args.validation_epoch > 0:
setup_eval()
val_dataset = COCODetection(image_path=cfg.dataset.valid_images,
info_file=cfg.dataset.valid_info,
transform=BaseTransform(MEANS))
# Parallel wraps the underlying module, but when saving and loading we don't want that
yolact_net = Yolact()
net = yolact_net
net.train()
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs.
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
yolact_net.load_weights(args.resume)
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
yolact_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
optimizer = optim.SGD(net.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
criterion = MultiBoxLoss(num_classes=cfg.num_classes,
pos_threshold=cfg.positive_iou_threshold,
neg_threshold=cfg.negative_iou_threshold,
negpos_ratio=3)
if args.cuda:
cudnn.benchmark = True
net = nn.DataParallel(net).cuda()
criterion = nn.DataParallel(criterion).cuda()
# net = net.cuda()
# criterion = criterion.cuda()
# criterion = criterion.cuda()
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
print("Dataset Size:")
print(len(dataset))
num_epochs = math.ceil(cfg.max_iter / epoch_size)
num_epochs = 1
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
data_loader = data.DataLoader(dataset,
args.batch_size,
num_workers=args.num_workers,
shuffle=True,
collate_fn=detection_collate,
pin_memory=True)
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
global loss_types # Forms the print order
loss_avgs = {k: MovingAverage(100) for k in loss_types}
print('Begin interpret!')
print()
# try-except so you can use ctrl+c to save early and stop training
try:
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch + 1) * epoch_size < iteration:
continue
count = 0
for datum in data_loader:
del datum
count += 1
if count % 10000 == 0:
print(count)
continue
except KeyboardInterrupt:
print('Stopping early. Saving network...')
print("Loaded Dataset Numbers")
print(count)
def train(args, cfg, option, DataSet):
if args.exp_name is not None:
args.save_folder = os.path.join(args.save_folder, args.exp_name)
args.log_folder = os.path.join(args.log_folder, args.exp_name)
if not os.path.exists(args.save_folder):
os.makedirs(args.save_folder, exist_ok=True)
if not os.path.exists(args.log_folder):
os.makedirs(args.log_folder, exist_ok=True)
if True:
dataset = DataSet(image_path=cfg.dataset.train_images,
mask_out_ch=cfg.gt_inst_ch,
info_file=cfg.dataset.train_info,
option=cfg.dataset,
transform=SSDAugmentation(cfg, MEANS),
running_mode='train')
else:
dataset = DataSet(image_path=cfg.dataset.valid_images,
mask_out_ch=cfg.gt_inst_ch,
info_file=cfg.dataset.valid_info,
option=cfg.dataset,
transform=SSDAugmentation(cfg, MEANS),
running_mode='train')
# Parallel wraps the underlying module, but when saving and loading we don't want that
dvis_net = DVIS(cfg)
net = dvis_net
net.train()
if args.log:
log = Log(cfg.name,
args.log_folder,
dict(args._get_kwargs()),
overwrite=(args.resume is None),
log_gpu_stats=args.log_gpu)
# I don't use the timer during training (I use a different timing method).
# Apparently there's a race condition with multiple GPUs, so disable it just to be safe.
timer.disable_all()
# Both of these can set args.resume to None, so do them before the check
if args.resume == 'interrupt':
args.resume = SavePath.get_interrupt(args.save_folder)
elif args.resume == 'latest':
args.resume = SavePath.get_latest(args.save_folder, cfg.name)
if args.resume is not None:
print('Resuming training, loading {}...'.format(args.resume))
dvis_net.load_weights(args.resume,
load_firstLayer=option['model_1stLayer_en'],
load_lastLayer=option['model_lastLayer_en'])
if args.start_iter == -1:
args.start_iter = SavePath.from_str(args.resume).iteration
else:
print('Initializing weights...')
dvis_net.init_weights(backbone_path=args.save_folder +
cfg.backbone.path)
#optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum,
# weight_decay=args.decay)
optimizer = optim.SGD([{
'params': net.backbone.parameters(),
'lr': args.lr * option['bkb_lr_alpha']
}, {
'params': net.fpn.parameters(),
'lr': args.lr * option['fpn_lr_alpha']
}, {
'params': net.proto_net.parameters(),
'lr': args.lr * option['proto_net_lr_alpha']
}],
lr=args.lr,
momentum=args.momentum,
weight_decay=args.decay)
criterion = LossEvaluate(option, class_weights=cfg.dataset.sem_weights)
if args.batch_alloc is not None:
args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')]
if sum(args.batch_alloc) != args.batch_size:
print(
'Error: Batch allocation (%s) does not sum to batch size (%s).'
% (args.batch_alloc, args.batch_size))
exit(-1)
net = NetLoss(net, criterion)
net = CustomDataParallel(net)
if args.cuda:
net = net.cuda()
# Initialize everything
if not cfg.freeze_bn:
dvis_net.freeze_bn() # Freeze bn so we don't kill our means
# loss counters
loc_loss = 0
conf_loss = 0
iteration = max(args.start_iter, 0)
last_time = time.time()
epoch_size = len(dataset) // args.batch_size
num_epochs = math.ceil(cfg.max_iter / epoch_size)
# Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index
step_index = 0
data_loader = data.DataLoader(dataset,
args.batch_size,
num_workers=args.num_workers,
shuffle=False,
collate_fn=detection_collate,
pin_memory=True)
writer = SummaryWriter(log_dir=args.log_folder)
save_path = lambda epoch, iteration: SavePath(
cfg.name, epoch, iteration).get_path(root=args.save_folder)
time_avg = MovingAverage()
loss_keys = [
'binary', 'pi', 'l1', 'regul', 'iou', 'classify', 'eval_prec',
'eval_rec', 'eval_acc'
]
vis_keys = ['preds', 'gts', 'rgb', 'wghts', 'grad']
loss_avgs = {k: MovingAverage(100) for k in loss_keys}
print('Begin training!')
# try-except so you can use ctrl+c to save early and stop training
try:
log_loss = dict()
for epoch in range(num_epochs):
# Resume from start_iter
if (epoch + 1) * epoch_size < iteration:
continue
for datum in data_loader:
# Stop if we've reached an epoch if we're resuming from start_iter
if iteration == (epoch + 1) * epoch_size:
break
# Stop at the configured number of iterations even if mid-epoch
if iteration == cfg.max_iter:
break
if iteration < 99:
iteration += 1
continue
# Change a config setting if we've reached the specified iteration
changed = False
for change in cfg.delayed_settings:
if iteration >= change[0]:
changed = True
cfg.replace(change[1])
# Reset the loss averages because things might have changed
for avg in loss_avgs:
avg.reset()
# If a config setting was changed, remove it from the list so we don't keep checking
if changed:
cfg.delayed_settings = [
x for x in cfg.delayed_settings if x[0] > iteration
]
# Warm up by linearly interpolating the learning rate from some smaller value
if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until:
set_lr(optimizer, (args.lr - cfg.lr_warmup_init) *
(iteration / cfg.lr_warmup_until) +
cfg.lr_warmup_init)
# Adjust the learning rate at the given iterations, but also if we resume from past that iteration
while step_index < len(
cfg.lr_steps
) and iteration >= cfg.lr_steps[step_index]:
step_index += 1
set_lr(optimizer, args.lr * (args.gamma**step_index))
# Zero the grad to get ready to compute gradients
optimizer.zero_grad()
# Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss0)
ret = net(datum)
# Mean here because Dataparallel and do Backprop
losses = {k: ret[k].mean() for k in loss_keys if k in ret}
det_loss_keys = [k for k in loss_keys if k in losses]
all_loss = sum([losses[k] for k in det_loss_keys])
for k in det_loss_keys:
loss_avgs[k].add(losses[k].item())
# backward and optimize
if args.show_gradients == True:
ret['preds_0'].retain_grad()
all_loss.backward(retain_graph=True)
ret['grad'] = ret['preds_0'].grad[:, 0, :, :]
else:
all_loss.backward(
) # Do this to free up vram even if loss is not finite
if torch.isfinite(all_loss).item():
optimizer.step()
ret['preds'] = torch.nn.ReLU()(ret['preds'])
vis_imgs = {k: ret[k] for k in vis_keys if k in ret}
cur_time = time.time()
elapsed = cur_time - last_time
last_time = cur_time
# Exclude graph setup from the timing information
if iteration != args.start_iter:
time_avg.add(elapsed)
if iteration % 10 == 0:
eta_str = str(
datetime.timedelta(seconds=(cfg.max_iter - iteration) *
time_avg.get_avg())).split('.')[0]
total = sum([
loss_avgs[k].get_avg() for k in det_loss_keys
if 'eval' not in k
])
loss_labels = sum(
[[k, loss_avgs[k].get_avg()]
for k in loss_keys if k in det_loss_keys], [])
print(('[%3d] %7d ||' +
(' %s: %.3f |' * len(det_loss_keys)) +
' T: %.3f || ETA: %s || timer: %.3f') %
tuple([epoch, iteration] + loss_labels +
[total, eta_str, elapsed]),
flush=True)
if args.log:
log_step = 50 // args.batch_size
for k in det_loss_keys:
if k not in log_loss:
log_loss[k] = loss_avgs[k].get_avg()
else:
log_loss[k] += loss_avgs[k].get_avg()
if iteration % log_step == log_step - 1:
for k in det_loss_keys:
writer.add_scalar(k + '_loss',
log_loss[k] / float(log_step),
iteration / log_step)
log_loss[k] = 0
log_fig_step = 100
if iteration % log_fig_step == log_fig_step - 1:
if 'davis' in args.dataset:
vis_imgs['rgb'] = vis_imgs['rgb'][:, :3, :, :]
fig = plot_tfboard_figure(
cfg, vis_imgs, show_grad=args.show_gradients)
writer.add_figure('prediction _ grad',
fig,
global_step=iteration / log_fig_step)
iteration += 1
if iteration % args.save_interval == 0 and iteration != args.start_iter:
if args.keep_latest:
latest = SavePath.get_latest(args.save_folder,
cfg.name)
print('Saving state, iter:', iteration)
dvis_net.save_weights(save_path(epoch, iteration))
if args.keep_latest and latest is not None:
if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval:
print('Deleting old save...')
os.remove(latest)
del ret, vis_imgs, losses
# end of batch run
# end of epoch
except KeyboardInterrupt:
if args.interrupt:
print('Stopping early. Saving network...')
# Delete previous copy of the interrupted network so we don't spam the weights folder
SavePath.remove_interrupt(args.save_folder)
writer.close()
dvis_net.save_weights(
save_path(epoch,
repr(iteration) + '_interrupt'))
exit()
writer.close()
dvis_net.save_weights(save_path(epoch, iteration))
