def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('fcis_psroi_align_resnet101', ),
default='fcis_psroi_align_resnet101')
parser.add_argument('--pretrained-model')
parser.add_argument('--iou-thresh', type=float, default=0.5)
parser.add_argument('--gpu', type=int, default=-1)
args = parser.parse_args()
if args.model == 'fcis_psroi_align_resnet101':
if args.pretrained_model:
model = FCISPSROIAlignResNet101(
n_fg_class=len(sbd_instance_segmentation_label_names),
pretrained_model=args.pretrained_model)
else:
model = FCISPSROIAlignResNet101(pretrained_model='sbd')
model.use_preset('evaluate')
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
dataset = SBDInstanceSegmentationDataset(split='val')
iterator = iterators.SerialIterator(dataset,
1,
repeat=False,
shuffle=False)
in_values, out_values, rest_values = apply_to_iterator(model.predict,
iterator,
hook=ProgressHook(
len(dataset)))
# delete unused iterators explicitly
del in_values
pred_masks, pred_labels, pred_scores = out_values
gt_masks, gt_labels = rest_values
result = eval_instance_segmentation_voc(pred_masks,
pred_labels,
pred_scores,
gt_masks,
gt_labels,
args.iou_thresh,
use_07_metric=True)
print('')
print('mAP: {:f}'.format(result['map']))
for l, name in enumerate(sbd_instance_segmentation_label_names):
if result['ap'][l]:
print('{:s}: {:f}'.format(name, result['ap'][l]))
else:
print('{:s}: -'.format(name))
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: FCIS')
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--out',
'-o',
default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--lr', '-l', type=float, default=0.0005)
parser.add_argument('--lr-cooldown-factor',
'-lcf',
type=float,
default=0.1)
parser.add_argument('--epoch', '-e', type=int, default=42)
parser.add_argument('--cooldown-epoch', '-ce', type=int, default=28)
args = parser.parse_args()
np.random.seed(args.seed)
# dataset
train_dataset = SBDInstanceSegmentationDataset(split='train')
test_dataset = SBDInstanceSegmentationDataset(split='val')
# model
fcis = FCISResNet101(n_fg_class=len(sbd_instance_segmentation_label_names),
pretrained_model='imagenet',
iter2=False)
fcis.use_preset('evaluate')
model = FCISTrainChain(fcis)
# gpu
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
# optimizer
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
optimizer.setup(model)
model.fcis.head.conv1.W.update_rule.add_hook(GradientScaling(3.0))
model.fcis.head.conv1.b.update_rule.add_hook(GradientScaling(3.0))
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
for param in model.params():
if param.name in ['beta', 'gamma']:
param.update_rule.enabled = False
model.fcis.extractor.conv1.disable_update()
model.fcis.extractor.res2.disable_update()
train_dataset = TransformDataset(train_dataset, Transform(model.fcis))
# iterator
train_iter = chainer.iterators.SerialIterator(train_dataset, batch_size=1)
test_iter = chainer.iterators.SerialIterator(test_dataset,
batch_size=1,
repeat=False,
shuffle=False)
updater = chainer.training.updater.StandardUpdater(
train_iter, optimizer, converter=concat_examples, device=args.gpu)
trainer = chainer.training.Trainer(updater, (args.epoch, 'epoch'),
out=args.out)
# lr scheduler
trainer.extend(chainer.training.extensions.ExponentialShift(
'lr', args.lr_cooldown_factor, init=args.lr),
trigger=(args.cooldown_epoch, 'epoch'))
# interval
log_interval = 100, 'iteration'
plot_interval = 3000, 'iteration'
print_interval = 20, 'iteration'
# training extensions
trainer.extend(extensions.snapshot_object(model.fcis,
filename='snapshot_model.npz'),
trigger=(args.epoch, 'epoch'))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(
extensions.LogReport(log_name='log.json', trigger=log_interval))
trainer.extend(extensions.PrintReport([
'iteration',
'epoch',
'elapsed_time',
'lr',
'main/loss',
'main/rpn_loc_loss',
'main/rpn_cls_loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/roi_mask_loss',
'validation/main/map',
]),
trigger=print_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
if extensions.PlotReport.available():
trainer.extend(extensions.PlotReport(['main/loss'],
file_name='loss.png',
trigger=plot_interval),
trigger=plot_interval)
trainer.extend(InstanceSegmentationVOCEvaluator(
test_iter,
model.fcis,
iou_thresh=0.5,
use_07_metric=True,
label_names=sbd_instance_segmentation_label_names),
trigger=ManualScheduleTrigger([
len(train_dataset) * args.cooldown_epoch,
len(train_dataset) * args.epoch
], 'iteration'))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
def setup(dataset, model_name, pretrained_model, batchsize):
cls, pretrained_models, default_batchsize = models[model_name]
dataset_name = dataset
if pretrained_model is None:
pretrained_model = pretrained_models.get(dataset_name, dataset_name)
if batchsize is None:
batchsize = default_batchsize
if dataset_name == 'sbd':
dataset = SBDInstanceSegmentationDataset(split='val')
label_names = sbd_instance_segmentation_label_names
model = cls(
n_fg_class=len(label_names), pretrained_model=pretrained_model)
model.use_preset('evaluate')
def eval_(out_values, rest_values):
pred_masks, pred_labels, pred_scores = out_values
gt_masks, gt_labels = rest_values
result = eval_instance_segmentation_voc(
pred_masks, pred_labels, pred_scores,
gt_masks, gt_labels, use_07_metric=True)
print('')
print('mAP: {:f}'.format(result['map']))
for l, name in enumerate(sbd_instance_segmentation_label_names):
if result['ap'][l]:
print('{:s}: {:f}'.format(name, result['ap'][l]))
else:
print('{:s}: -'.format(name))
elif dataset_name == 'coco':
dataset = COCOInstanceSegmentationDataset(
split='minival', year='2014',
use_crowded=True, return_crowded=True, return_area=True)
label_names = coco_instance_segmentation_label_names
if model_name == 'fcis_resnet101':
proposal_creator_params = cls.proposal_creator_params
proposal_creator_params['min_size'] = 2
model = cls(
n_fg_class=len(label_names),
anchor_scales=(4, 8, 16, 32),
pretrained_model=pretrained_model,
proposal_creator_params=proposal_creator_params)
model.use_preset('coco_evaluate')
else:
model = cls(
n_fg_class=len(label_names), pretrained_model=pretrained_model)
model.use_preset('evaluate')
def eval_(out_values, rest_values):
pred_masks, pred_labels, pred_scores = out_values
gt_masks, gt_labels, gt_areas, gt_crowdeds = rest_values
result = eval_instance_segmentation_coco(
pred_masks, pred_labels, pred_scores,
gt_masks, gt_labels, gt_areas, gt_crowdeds)
print()
for area in ('all', 'large', 'medium', 'small'):
print('mmAP ({}):'.format(area),
result['map/iou=0.50:0.95/area={}/max_dets=100'.format(
area)])
return dataset, label_names, eval_, model, batchsize
def setUp(self):
self.dataset = SBDInstanceSegmentationDataset(split=self.split)
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: FCIS')
parser.add_argument('--out',
'-o',
default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--lr',
'-l',
type=float,
default=None,
help='Learning rate for multi GPUs')
parser.add_argument('--batchsize', type=int, default=8)
parser.add_argument('--epoch', '-e', type=int, default=42)
parser.add_argument('--cooldown-epoch', '-ce', type=int, default=28)
args = parser.parse_args()
# https://docs.chainer.org/en/stable/chainermn/tutorial/tips_faqs.html#using-multiprocessiterator
if hasattr(multiprocessing, 'set_start_method'):
multiprocessing.set_start_method('forkserver')
p = multiprocessing.Process()
p.start()
p.join()
# chainermn
comm = chainermn.create_communicator('pure_nccl')
device = comm.intra_rank
np.random.seed(args.seed)
# model
fcis = FCISResNet101(n_fg_class=len(sbd_instance_segmentation_label_names),
pretrained_model='imagenet',
iter2=False)
fcis.use_preset('evaluate')
model = FCISTrainChain(fcis)
chainer.cuda.get_device_from_id(device).use()
model.to_gpu()
# dataset
train_dataset = TransformDataset(
SBDInstanceSegmentationDataset(split='train'),
('img', 'mask', 'label', 'bbox', 'scale'), Transform(model.fcis))
if comm.rank == 0:
indices = np.arange(len(train_dataset))
else:
indices = None
indices = chainermn.scatter_dataset(indices, comm, shuffle=True)
train_dataset = train_dataset.slice[indices]
train_iter = chainer.iterators.SerialIterator(train_dataset,
batch_size=args.batchsize //
comm.size)
if comm.rank == 0:
test_dataset = SBDInstanceSegmentationDataset(split='val')
test_iter = chainer.iterators.SerialIterator(test_dataset,
batch_size=1,
repeat=False,
shuffle=False)
# optimizer
optimizer = chainermn.create_multi_node_optimizer(
chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9), comm)
optimizer.setup(model)
model.fcis.head.conv1.W.update_rule.add_hook(GradientScaling(3.0))
model.fcis.head.conv1.b.update_rule.add_hook(GradientScaling(3.0))
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
for param in model.params():
if param.name in ['beta', 'gamma']:
param.update_rule.enabled = False
model.fcis.extractor.conv1.disable_update()
model.fcis.extractor.res2.disable_update()
updater = chainer.training.updater.StandardUpdater(
train_iter, optimizer, converter=concat_examples, device=device)
trainer = chainer.training.Trainer(updater, (args.epoch, 'epoch'),
out=args.out)
@make_shift('lr')
def lr_scheduler(trainer):
if args.lr is None:
base_lr = 0.0005 * args.batchsize
else:
base_lr = args.lr
epoch = trainer.updater.epoch
if epoch < args.cooldown_epoch:
rate = 1
else:
rate = 0.1
return rate * base_lr
trainer.extend(lr_scheduler)
if comm.rank == 0:
# interval
log_interval = 100, 'iteration'
plot_interval = 3000, 'iteration'
print_interval = 20, 'iteration'
# training extensions
trainer.extend(extensions.snapshot_object(
model.fcis, filename='snapshot_model.npz'),
trigger=(args.epoch, 'epoch'))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(
extensions.LogReport(log_name='log.json', trigger=log_interval))
trainer.extend(extensions.PrintReport([
'iteration',
'epoch',
'elapsed_time',
'lr',
'main/loss',
'main/rpn_loc_loss',
'main/rpn_cls_loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/roi_mask_loss',
'validation/main/map',
]),
trigger=print_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
if extensions.PlotReport.available():
trainer.extend(extensions.PlotReport(['main/loss'],
file_name='loss.png',
trigger=plot_interval),
trigger=plot_interval)
trainer.extend(InstanceSegmentationVOCEvaluator(
test_iter,
model.fcis,
iou_thresh=0.5,
use_07_metric=True,
label_names=sbd_instance_segmentation_label_names),
trigger=ManualScheduleTrigger([
len(train_dataset) * args.cooldown_epoch,
len(train_dataset) * args.epoch
], 'iteration'))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
