def setUp(self):
bboxes = [generate_random_bbox(5, (256, 324), 24, 120)
for _ in range(10)]
labels = np.ones((10, 5))
self.dataset = TupleDataset(
np.random.uniform(size=(10, 3, 32, 48)),
bboxes,
labels)
self.link = _DetectionStubLink(bboxes, labels)
self.iterator = SerialIterator(
self.dataset, 5, repeat=False, shuffle=False)
self.evaluator = DetectionVOCEvaluator(self.iterator, self.link)
self.expect_map = 1
class TestDetectionVOCEvaluator(unittest.TestCase):
def setUp(self):
bboxes = [
generate_random_bbox(5, (256, 324), 24, 120) for _ in range(10)
]
labels = np.ones((10, 5))
self.dataset = TupleDataset(np.random.uniform(size=(10, 3, 32, 48)),
bboxes, labels)
self.link = _DetectionStubLink(bboxes, labels)
self.iterator = SerialIterator(self.dataset,
5,
repeat=False,
shuffle=False)
self.evaluator = DetectionVOCEvaluator(self.iterator,
self.link,
label_names=('cls0', 'cls1',
'cls2'))
self.expected_ap = 1
def test_evaluate(self):
reporter = chainer.Reporter()
reporter.add_observer('target', self.link)
with reporter:
mean = self.evaluator.evaluate()
# No observation is reported to the current reporter. Instead the
# evaluator collect results in order to calculate their mean.
self.assertEqual(len(reporter.observation), 0)
np.testing.assert_equal(mean['target/map'], self.expected_ap)
np.testing.assert_equal(mean['target/ap/cls0'], np.nan)
np.testing.assert_equal(mean['target/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['target/ap/cls2'], np.nan)
def test_call(self):
mean = self.evaluator()
# main is used as default
np.testing.assert_equal(mean['main/map'], self.expected_ap)
np.testing.assert_equal(mean['main/ap/cls0'], np.nan)
np.testing.assert_equal(mean['main/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['main/ap/cls2'], np.nan)
def test_evaluator_name(self):
self.evaluator.name = 'eval'
mean = self.evaluator()
# name is used as a prefix
np.testing.assert_equal(mean['eval/main/map'], self.expected_ap)
np.testing.assert_equal(mean['eval/main/ap/cls0'], np.nan)
np.testing.assert_equal(mean['eval/main/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['eval/main/ap/cls2'], np.nan)
def test_current_report(self):
reporter = chainer.Reporter()
with reporter:
mean = self.evaluator()
# The result is reported to the current reporter.
self.assertEqual(reporter.observation, mean)
class TestDetectionVOCEvaluator(unittest.TestCase):
def setUp(self):
bboxes = [generate_random_bbox(5, (256, 324), 24, 120)
for _ in range(10)]
labels = np.ones((10, 5))
self.dataset = TupleDataset(
np.random.uniform(size=(10, 3, 32, 48)),
bboxes,
labels)
self.link = _DetectionStubLink(bboxes, labels)
self.iterator = SerialIterator(
self.dataset, 5, repeat=False, shuffle=False)
self.evaluator = DetectionVOCEvaluator(
self.iterator, self.link, label_names=('cls0', 'cls1', 'cls2'))
self.expected_ap = 1
def test_evaluate(self):
reporter = chainer.Reporter()
reporter.add_observer('target', self.link)
with reporter:
mean = self.evaluator.evaluate()
# No observation is reported to the current reporter. Instead the
# evaluator collect results in order to calculate their mean.
self.assertEqual(len(reporter.observation), 0)
np.testing.assert_equal(mean['target/map'], self.expected_ap)
np.testing.assert_equal(mean['target/ap/cls0'], np.nan)
np.testing.assert_equal(mean['target/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['target/ap/cls2'], np.nan)
def test_call(self):
mean = self.evaluator()
# main is used as default
np.testing.assert_equal(mean['main/map'], self.expected_ap)
np.testing.assert_equal(mean['main/ap/cls0'], np.nan)
np.testing.assert_equal(mean['main/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['main/ap/cls2'], np.nan)
def test_evaluator_name(self):
self.evaluator.name = 'eval'
mean = self.evaluator()
# name is used as a prefix
np.testing.assert_equal(mean['eval/main/map'], self.expected_ap)
np.testing.assert_equal(mean['eval/main/ap/cls0'], np.nan)
np.testing.assert_equal(mean['eval/main/ap/cls1'], self.expected_ap)
np.testing.assert_equal(mean['eval/main/ap/cls2'], np.nan)
def test_current_report(self):
reporter = chainer.Reporter()
with reporter:
mean = self.evaluator()
# The result is reported to the current reporter.
self.assertEqual(reporter.observation, mean)
def setUp(self):
bboxes = [generate_random_bbox(5, (256, 324), 24, 120)
for _ in range(10)]
labels = np.ones((10, 5))
self.dataset = TupleDataset(
np.random.uniform(size=(10, 3, 32, 48)),
bboxes,
labels)
self.link = _DetectionStubLink(bboxes, labels)
self.iterator = SerialIterator(
self.dataset, 5, repeat=False, shuffle=False)
self.evaluator = DetectionVOCEvaluator(
self.iterator, self.link, label_names=('cls0', 'cls1', 'cls2'))
self.expected_ap = 1
def test_consistency(self):
reporter = chainer.Reporter()
if self.comm.rank == 0:
multi_iterator = SerialIterator(self.dataset,
self.batchsize,
repeat=False,
shuffle=False)
else:
multi_iterator = None
multi_link = _DetectionStubLink(self.bboxes, self.labels,
self.initial_count)
multi_evaluator = DetectionVOCEvaluator(multi_iterator,
multi_link,
label_names=('cls0', 'cls1',
'cls2'),
comm=self.comm)
reporter.add_observer('target', multi_link)
with reporter:
multi_mean = multi_evaluator.evaluate()
if self.comm.rank != 0:
self.assertEqual(multi_mean, {})
return
single_iterator = SerialIterator(self.dataset,
self.batchsize,
repeat=False,
shuffle=False)
single_link = _DetectionStubLink(self.bboxes, self.labels)
single_evaluator = DetectionVOCEvaluator(single_iterator,
single_link,
label_names=('cls0', 'cls1',
'cls2'))
reporter.add_observer('target', single_link)
with reporter:
single_mean = single_evaluator.evaluate()
self.assertEqual(set(multi_mean.keys()), set(single_mean.keys()))
for key in multi_mean.keys():
np.testing.assert_equal(single_mean[key], multi_mean[key])
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: Faster R-CNN')
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--lr', '-l', type=float, default=1e-3)
parser.add_argument('--out', '-o', default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--step_size', '-ss', type=int, default=50000)
parser.add_argument('--iteration', '-i', type=int, default=70000)
parser.add_argument('--train_data_dir', '-t', default=WIDER_TRAIN_DIR,
help='Training dataset (WIDER_train)')
parser.add_argument('--train_annotation', '-ta', default=WIDER_TRAIN_ANNOTATION_MAT,
help='Annotation file (.mat) for training dataset')
parser.add_argument('--val_data_dir', '-v', default=WIDER_VAL_DIR,
help='Validation dataset (WIDER_train)')
parser.add_argument('--val_annotation', '-va', default=WIDER_VAL_ANNOTATION_MAT,
help='Annotation file (.mat) for validation dataset')
args = parser.parse_args()
np.random.seed(args.seed)
# for logging pocessed files
logger = logging.getLogger('logger')
logger.setLevel(logging.DEBUG)
handler = logging.FileHandler(filename='filelog.log')
handler.setLevel(logging.DEBUG)
logger.addHandler(handler)
blacklist = []
with open(BLACKLIST_FILE, 'r') as f:
for line in f:
l = line.strip()
if l:
blacklist.append(line.strip())
# train_data = VOCDetectionDataset(split='trainval', year='2007')
# test_data = VOCDetectionDataset(split='test', year='2007',
# use_difficult=True, return_difficult=True)
train_data = WIDERFACEDataset(args.train_data_dir, args.train_annotation,
logger=logger, exclude_file_list=blacklist)
test_data = WIDERFACEDataset(args.val_data_dir, args.val_annotation)
# faster_rcnn = FasterRCNNVGG16(n_fg_class=len(voc_detection_label_names),
# pretrained_model='imagenet')
faster_rcnn.use_preset('evaluate')
model = FasterRCNNTrainChain(faster_rcnn)
if args.gpu >= 0:
model.to_gpu(args.gpu)
chainer.cuda.get_device(args.gpu).use()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
train_data = TransformDataset(train_data, transform)
#import pdb; pdb.set_trace()
#train_iter = chainer.iterators.MultiprocessIterator(
# train_data, batch_size=1, n_processes=None, shared_mem=100000000)
train_iter = chainer.iterators.SerialIterator(
train_data, batch_size=1)
test_iter = chainer.iterators.SerialIterator(
test_data, batch_size=1, repeat=False, shuffle=False)
updater = chainer.training.updater.StandardUpdater(
train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(
updater, (args.iteration, 'iteration'), out=args.out)
trainer.extend(
extensions.snapshot_object(model.faster_rcnn, 'snapshot_model.npz'),
trigger=(args.iteration, 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(args.step_size, 'iteration'))
log_interval = 20, 'iteration'
plot_interval = 3000, 'iteration'
print_interval = 20, 'iteration'
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport(
['iteration', 'epoch', 'elapsed_time', 'lr',
'main/loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/rpn_loc_loss',
'main/rpn_cls_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(
DetectionVOCEvaluator(
test_iter, model.faster_rcnn, use_07_metric=True,
label_names=('face',)),
trigger=ManualScheduleTrigger(
[args.step_size, args.iteration], 'iteration'),
invoke_before_training=False)
trainer.extend(extensions.dump_graph('main/loss'))
#try:
# warnings.filterwarnings('error', category=RuntimeWarning)
trainer.run()
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: Faster R-CNN')
parser.add_argument('--dataset',
choices=('voc07', 'voc0712'),
help='The dataset to use: VOC07, VOC07+12',
default='voc07')
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--lr', '-l', type=float, default=1e-3)
parser.add_argument('--out',
'-o',
default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--step_size', '-ss', type=int, default=50000)
parser.add_argument('--iteration', '-i', type=int, default=70000)
args = parser.parse_args()
np.random.seed(args.seed)
if args.dataset == 'voc07':
train_data = VOCBboxDataset(split='trainval', year='2007')
elif args.dataset == 'voc0712':
train_data = ConcatenatedDataset(
VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval'))
test_data = VOCBboxDataset(split='test',
year='2007',
use_difficult=True,
return_difficult=True)
faster_rcnn = FasterRCNNVGG16(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
faster_rcnn.use_preset('evaluate')
model = FasterRCNNTrainChain(faster_rcnn)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer_hooks.WeightDecay(rate=0.0005))
train_data = TransformDataset(train_data, Transform(faster_rcnn))
train_iter = chainer.iterators.MultiprocessIterator(train_data,
batch_size=1,
n_processes=None,
shared_mem=100000000)
test_iter = chainer.iterators.SerialIterator(test_data,
batch_size=1,
repeat=False,
shuffle=False)
updater = chainer.training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
out=args.out)
trainer.extend(extensions.snapshot_object(model.faster_rcnn,
'snapshot_model.npz'),
trigger=(args.iteration, 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(args.step_size, 'iteration'))
log_interval = 20, 'iteration'
plot_interval = 3000, 'iteration'
print_interval = 20, 'iteration'
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'iteration',
'epoch',
'elapsed_time',
'lr',
'main/loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/rpn_loc_loss',
'main/rpn_cls_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(DetectionVOCEvaluator(test_iter,
model.faster_rcnn,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=ManualScheduleTrigger(
[args.step_size, args.iteration], 'iteration'))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
def handler(context):
dataset_alias = context.datasets
trainval_2007_dataset_id = dataset_alias['trainval2007']
trainval_2012_dataset_id = dataset_alias['trainval2012']
test_2007_dataset_id = dataset_alias['test2007']
trainval_2007_dataset = list(
load_dataset_from_api(trainval_2007_dataset_id))
trainval_2012_dataset = list(
load_dataset_from_api(trainval_2012_dataset_id))
test_2007_dataset = list(load_dataset_from_api(test_2007_dataset_id))
if network_model == 'ssd300':
model = SSD300(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
elif network_model == 'ssd512':
model = SSD512(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if USE_GPU >= 0:
chainer.cuda.get_device_from_id(USE_GPU).use()
model.to_gpu()
trainval_2007 = DetectionDatasetFromAPI(trainval_2007_dataset)
trainval_2012 = DetectionDatasetFromAPI(trainval_2012_dataset)
test_2007 = DetectionDatasetFromAPI(test_2007_dataset,
use_difficult=True,
return_difficult=True)
train = TransformDataset(ConcatenatedDataset(trainval_2007, trainval_2012),
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.SerialIterator(train, BATCHSIZE)
test_iter = chainer.iterators.SerialIterator(test_2007,
BATCHSIZE,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=USE_GPU)
trainer = training.Trainer(updater, (nb_iterations, 'iteration'),
out=ABEJA_TRAINING_RESULT_DIR)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger([80000, 100000],
'iteration'))
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=(10000, 'iteration'))
log_interval = 100, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
print_entries = [
'iteration', 'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map'
]
report_entries = [
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]
trainer.extend(Statistics(report_entries,
nb_iterations,
obs_key='iteration'),
trigger=log_interval)
trainer.extend(Tensorboard(report_entries, out_dir=log_path))
trainer.extend(extensions.PrintReport(print_entries), trigger=log_interval)
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(nb_iterations, 'iteration'))
trainer.run()
def main():
bbox_label_names = ('loop')
n_itrs = 70000
n_step = 50000
np.random.seed(0)
train_data = DefectDetectionDataset(split='train')
test_data = DefectDetectionDataset(split='test')
proposal_params = {'min_size': 8}
faster_rcnn = FasterRCNNVGG16(n_fg_class=1, pretrained_model='imagenet', ratios=[0.5, 1, 2],
anchor_scales=[1, 4, 8, 16], min_size=512, max_size=1024,
proposal_creator_params=proposal_params)
faster_rcnn.use_preset('evaluate')
model = FasterRCNNTrainChain(faster_rcnn)
chainer.cuda.get_device_from_id(0).use()
model.to_gpu()
optimizer = chainer.optimizers.MomentumSGD(lr=1e-3, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
train_data = TransformDataset(train_data, Transform(faster_rcnn))
train_iter = chainer.iterators.MultiprocessIterator(
train_data, batch_size=1, n_processes=None, shared_mem=100000000)
test_iter = chainer.iterators.SerialIterator(
test_data, batch_size=1, repeat=False, shuffle=False)
updater = chainer.training.updater.StandardUpdater(
train_iter, optimizer, device=0)
trainer = training.Trainer(
updater, (n_itrs, 'iteration'), out='result')
trainer.extend(
extensions.snapshot_object(model.faster_rcnn, 'snapshot_model_{.updater.iteration}.npz'),
trigger=(n_itrs/5, 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(n_step, 'iteration'))
log_interval = 50, 'iteration'
plot_interval = 100, 'iteration'
print_interval = 20, 'iteration'
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport(
['iteration', 'epoch', 'elapsed_time', 'lr',
'main/loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/rpn_loc_loss',
'main/rpn_cls_loss',
'validation/main/map',
]), trigger=print_interval)
trainer.extend(extensions.ProgressBar(update_interval=5))
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(
['main/loss'],
file_name='loss.png', trigger=plot_interval
),
trigger=plot_interval
)
trainer.extend(
DetectionVOCEvaluator(
test_iter, model.faster_rcnn, use_07_metric=True,
label_names=bbox_label_names),
trigger=ManualScheduleTrigger(
[100, 500, 1000, 5000, 10000, 20000, 40000, 60000, n_step, n_itrs], 'iteration'))
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--batchsize', type=int, default=4)
parser.add_argument('--epoch', type=int, default=10)
parser.add_argument('--mini', action="store_true")
args = parser.parse_args()
if hasattr(multiprocessing, 'set_start_method'):
multiprocessing.set_start_method('forkserver')
p = multiprocessing.Process()
p.start()
p.join()
comm = chainermn.create_communicator('pure_nccl')
print(comm.size)
device = comm.intra_rank
num_class = len(voc_bbox_label_names)
data_augmentation_transform = DataAugmentationTransform(512)
center_detection_transform = CenterDetectionTransform(512, num_class, 4)
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
data_augmentation_transform)
if comm.rank == 0:
train = TransformDataset(train, center_detection_transform)
if args.mini:
train = datasets.SubDataset(train, 0, 100)
else:
train = None
train = chainermn.scatter_dataset(train, comm, shuffle=True)
train_iter = chainer.iterators.MultiprocessIterator(train,
args.batchsize //
comm.size,
n_processes=2)
if comm.rank == 0:
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
if args.mini:
test = datasets.SubDataset(test, 0, 20)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
detector = CenterDetector(HourglassNet, 512, num_class)
train_chain = CenterDetectorTrain(detector, 1, 0.1, 1, comm=comm)
chainer.cuda.get_device_from_id(device).use()
train_chain.to_gpu()
optimizer = chainermn.create_multi_node_optimizer(Adam(amsgrad=True), comm)
optimizer.setup(train_chain)
updater = StandardUpdater(train_iter, optimizer, device=device)
trainer = Trainer(updater, (args.epoch, 'epoch'))
if comm.rank == 0:
log_interval = 1, 'epoch'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch',
'iteration',
'lr',
'main/loss',
'main/hm_loss',
'main/wh_loss',
'main/offset_loss',
'validation/main/map',
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(DetectionVOCEvaluator(test_iter,
detector,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=(1, 'epoch'))
trainer.extend(extensions.snapshot_object(
detector, 'detector{.updator.epoch:03}.npz'),
trigger=(1, 'epoch'))
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('dataset', help="path to train json file")
parser.add_argument('test_dataset', help="path to test dataset json file")
parser.add_argument(
'--dataset-root',
help=
"path to dataset root if dataset file is not already in root folder of dataset"
)
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd512')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--gpu', type=int, nargs='*', default=[])
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
parser.add_argument('--lr',
type=float,
default=0.001,
help="default learning rate")
parser.add_argument('--port',
type=int,
default=1337,
help="port for bbox sending")
parser.add_argument('--ip',
default='127.0.0.1',
help="destination ip for bbox sending")
parser.add_argument(
'--test-image',
help="path to test image that shall be displayed in bbox vis")
args = parser.parse_args()
if args.dataset_root is None:
args.dataset_root = os.path.dirname(args.dataset)
if args.model == 'ssd300':
model = SSD300(n_fg_class=1, pretrained_model='imagenet')
image_size = (300, 300)
elif args.model == 'ssd512':
model = SSD512(n_fg_class=1, pretrained_model='imagenet')
image_size = (512, 512)
else:
raise NotImplementedError("The model you want to train does not exist")
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
train = TransformDataset(
SheepDataset(args.dataset_root, args.dataset, image_size=image_size),
Transform(model.coder, model.insize, model.mean))
if len(args.gpu) > 1:
gpu_datasets = split_dataset_n_random(train, len(args.gpu))
if not len(gpu_datasets[0]) == len(gpu_datasets[-1]):
adapted_second_split = split_dataset(gpu_datasets[-1],
len(gpu_datasets[0]))[0]
gpu_datasets[-1] = adapted_second_split
else:
gpu_datasets = [train]
train_iter = [
ThreadIterator(gpu_dataset, args.batchsize)
for gpu_dataset in gpu_datasets
]
test = SheepDataset(args.dataset_root,
args.test_dataset,
image_size=image_size)
test_iter = chainer.iterators.MultithreadIterator(test,
args.batchsize,
repeat=False,
shuffle=False,
n_threads=2)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.Adam(alpha=args.lr)
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
if len(args.gpu) <= 1:
updater = training.updaters.StandardUpdater(
train_iter[0],
optimizer,
device=args.gpu[0] if len(args.gpu) > 0 else -1,
)
else:
updater = training.updaters.MultiprocessParallelUpdater(
train_iter, optimizer, devices=args.gpu)
updater.setup_workers()
if len(args.gpu) > 0 and args.gpu[0] >= 0:
chainer.backends.cuda.get_device_from_id(args.gpu[0]).use()
model.to_gpu()
trainer = training.Trainer(updater, (200, 'epoch'), args.out)
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=(1000, 'iteration'))
# build logger
# make sure to log all data necessary for prediction
log_interval = 100, 'iteration'
data_to_log = {
'image_size': image_size,
'model_type': args.model,
}
# add all command line arguments
for argument in filter(lambda x: not x.startswith('_'), dir(args)):
data_to_log[argument] = getattr(args, argument)
# create callback that logs all auxiliary data the first time things get logged
def backup_train_config(stats_cpu):
if stats_cpu['iteration'] == log_interval:
stats_cpu.update(data_to_log)
trainer.extend(
extensions.LogReport(trigger=log_interval,
postprocess=backup_train_config))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(5000, 'iteration'))
if args.test_image is not None:
plot_image = train._dataset.load_image(args.test_image,
resize_to=image_size)
else:
plot_image, _, _ = train.get_example(0)
plot_image += train._transform.mean
bbox_plotter = BBOXPlotter(
plot_image,
os.path.join(args.out, 'bboxes'),
send_bboxes=True,
upstream_port=args.port,
upstream_ip=args.ip,
)
trainer.extend(bbox_plotter, trigger=(10, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
trainer.run()
def handler(context):
dataset_alias = context.datasets
data = list(load_dataset_from_api(dataset_alias['train']))
np.random.seed(0)
data = np.random.permutation(data)
nb_data = len(data)
nb_train = int(7 * nb_data // 10)
train_data_raw = data[:nb_train]
test_data_raw = data[nb_train:]
premodel = SSD300(n_fg_class=20, pretrained_model='voc0712')
model = SSD300(n_fg_class=1)
copy_ssd(model, premodel)
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if USE_GPU >= 0:
chainer.cuda.get_device_from_id(USE_GPU).use()
model.to_gpu()
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
fix_ssd(train_chain)
train_data = DetectionDatasetFromAPI(train_data_raw)
test_data = DetectionDatasetFromAPI(test_data_raw,
use_difficult=True,
return_difficult=True)
train_data = TransformDataset(
train_data, Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.SerialIterator(train_data, BATCHSIZE)
test_iter = chainer.iterators.SerialIterator(test_data,
BATCHSIZE,
repeat=False,
shuffle=False)
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=USE_GPU)
trainer = training.Trainer(updater, (nb_epochs, 'epoch'),
out=ABEJA_TRAINING_RESULT_DIR)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger([1200, 1600],
'epoch'))
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=['cup']),
trigger=(1, 'epoch'))
log_interval = 1, 'epoch'
trainer.extend(extensions.LogReport(trigger=log_interval))
print_entries = [
'epoch', 'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map'
]
report_entries = [
'epoch', 'lr', 'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map'
]
trainer.extend(Statistics(report_entries, nb_epochs), trigger=log_interval)
trainer.extend(Tensorboard(report_entries, out_dir=log_path))
trainer.extend(extensions.PrintReport(print_entries), trigger=log_interval)
trainer.extend(extensions.snapshot_object(model,
'model_epoch_{.updater.epoch}'),
trigger=(nb_epochs, 'epoch'))
trainer.run()
else:
param.update_rule.add_hook(WeightDecay(0.0005))
# 最適化関数とトレーニングデータセットを入力
# 学習部分ににoptimizerを繋げる．
updater = training.updaters.StandardUpdater(train_iter, optimizer, device=gpu)
trainer = training.Trainer(updater, (iteration, 'iteration'), out)
# 指数関数的に学習率を変更できるように設定
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger(step, 'iteration'))
# 学習とともにAPやmAPを評価するように設定
trainer.extend(
DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=(voc_bbox_label_names +
('new_label', ))))
log_interval = 10, 'iteration'
# 評価結果の出力
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
# 結果の表示
def main():
rospack = rospkg.RosPack()
jsk_perception_datasets_path = osp.join(
rospack.get_path('jsk_perception'), 'learning_datasets')
parser = argparse.ArgumentParser()
# Dataset directory
parser.add_argument('--train-dataset-dir', type=str,
default=osp.join(jsk_perception_datasets_path,
'kitchen_dataset', 'train'))
parser.add_argument('--val-dataset-dir', type=str,
default=osp.join(jsk_perception_datasets_path,
'kitchen_dataset', 'test'))
parser.add_argument('--dataset-type', type=str,
default='instance')
parser.add_argument(
'--model-name', choices=('ssd300', 'ssd512'), default='ssd512')
parser.add_argument('--gpu', type=int, default=0)
parser.add_argument('--batch-size', type=int, default=8)
parser.add_argument('--max-epoch', type=int, default=100)
parser.add_argument('--out-dir', type=str, default=None)
args = parser.parse_args()
if (args.dataset_type == 'instance'):
train_dataset = DetectionDataset(args.train_dataset_dir)
elif (args.dataset_type == 'bbox'):
train_dataset = BboxDetectionDataset(args.train_dataset_dir)
else:
print('unsuppported dataset type')
return
fg_label_names = train_dataset.fg_class_names
if args.model_name == 'ssd300':
model = SSD300(
n_fg_class=len(fg_label_names),
pretrained_model='imagenet')
elif args.model_name == 'ssd512':
model = SSD512(
n_fg_class=len(fg_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
train = TransformDataset(
train_dataset,
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.MultiprocessIterator(train, args.batch_size)
if (args.dataset_type == 'instance'):
test_dataset = DetectionDataset(args.val_dataset_dir)
elif (args.dataset_type == 'bbox'):
test_dataset = BboxDetectionDataset(args.val_dataset_dir)
test_iter = chainer.iterators.SerialIterator(
test_dataset, args.batch_size, repeat=False, shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(
train_iter, optimizer, device=args.gpu)
now = datetime.datetime.now()
timestamp = now.strftime('%Y%m%d-%H%M%S')
if args.out_dir is None:
out_dir = osp.join(
rospkg.get_ros_home(), 'learning_logs', timestamp)
else:
out_dir = args.out_dir
step_epoch = [args.max_epoch * 2 // 3, args.max_epoch * 5 // 6]
trainer = training.Trainer(
updater, (args.max_epoch, 'epoch'), out_dir)
trainer.extend(
extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger(step_epoch, 'epoch'))
trainer.extend(
DetectionVOCEvaluator(
test_iter, model, use_07_metric=True,
label_names=fg_label_names),
trigger=triggers.ManualScheduleTrigger(
step_epoch + [args.max_epoch], 'epoch'))
log_interval = 10, 'iteration'
trainer.extend(
extensions.LogReport(log_name='log.json', trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'lr',
'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map']),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(
extensions.snapshot_object(
model, 'model_snapshot.npz'),
trigger=(args.max_epoch, 'epoch'))
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--data-dir', type=str,
default=os.path.join("RoadDamageDataset", "All"))
parser.add_argument('--batchsize', type=int, default=32,
help='Learning minibatch size')
parser.add_argument('--gpu', type=int, default=-1,
help='GPU ID (negative value indicates CPU')
parser.add_argument('--base-network', choices=('vgg16', 'resnet101'),
default='vgg16', help='Base network')
parser.add_argument('--pretrained-model', default=None,
help='Pretrained SSD model')
parser.add_argument('--pretrained-extractor', default='auto',
help='Pretrained CNN model to extract feature maps')
parser.add_argument('--out', default='result-detection',
help='Directory to output the result')
parser.add_argument('--resume', default=None,
help='Initialize the trainer from given file')
args = parser.parse_args()
print("Data directory : {}".format(args.data_dir))
print("Batchsize : {}".format(args.batchsize))
print("GPU ID : {}".format(args.gpu))
print("Base network : {}".format(args.base_network))
print("Pretrained extractor : {}".format(args.pretrained_extractor))
print("Pretrained model : {}".format(args.pretrained_model))
print("Output directory : {}".format(args.out))
print("Resume from : {}".format(args.resume))
if args.base_network == 'vgg16':
# pretrained_extractor is currently not available for this class
model = chainercv.links.SSD300(
n_fg_class=len(roaddamage_label_names),
pretrained_model=args.pretrained_model)
preprocessing = MeanSubtraction(model.mean)
elif args.base_network == 'resnet101':
model = ssd_resnet101.SSD224(
n_fg_class=len(roaddamage_label_names),
pretrained_extractor=args.pretrained_extractor,
pretrained_model=args.pretrained_model)
preprocessing = ResNetPreparation(model.insize)
else:
raise ValueError('Invalid base network')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
train = TransformDataset(
RoadDamageDataset(args.data_dir, split='train'),
Transform(model.coder, model.insize, model.mean)
)
train = TransformDataset(train, preprocessing)
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test = RoadDamageDataset(args.data_dir, split='val')
test_iter = chainer.iterators.SerialIterator(
test, args.batchsize, repeat=False, shuffle=False)
# initial lr is set to 3e-4 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (120000, 'iteration'), args.out)
trainer.extend(
extensions.ExponentialShift('lr', 0.1, init=3e-4),
trigger=triggers.ManualScheduleTrigger([80000, 100000], 'iteration'))
trainer.extend(
DetectionVOCEvaluator(
test_iter, model, use_07_metric=True,
label_names=roaddamage_label_names),
trigger=(4000, 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'lr',
'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map']),
trigger=log_interval)
# trainer.extend(extensions.ProgressBar())
trainer.extend(extensions.snapshot(), trigger=(4000, 'iteration'))
trainer.extend(
extensions.snapshot_object(model, 'model_iter_{.updater.iteration}'),
trigger=(4000, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
print("setup finished")
trainer.run()
model.to_cpu()
serializers.save_npz("model-detector.npz", model)
def main():
# cuDNNのautotuneを有効にする
chainer.cuda.set_max_workspace_size(512 * 1024 * 1024)
chainer.config.autotune = True
gpu_id = 0
batchsize = 6
out_num = 'results'
log_interval = 1, 'epoch'
epoch_max = 500
initial_lr = 0.0001
lr_decay_rate = 0.1
lr_decay_timing = [200, 300, 400]
# モデルの設定
model = SSD300(n_fg_class=len(voc_labels), pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
# GPUの設定
chainer.cuda.get_device_from_id(gpu_id).use()
model.to_gpu()
# データセットの設定
train_dataset = MyVoTTVOCDataset(
'C:\Python_Programs\chainer_practice\Telescope_corner', 'train')
valid_dataset = MyVoTTVOCDataset(
'C:\Python_Programs\chainer_practice\Telescope_corner', 'val')
# データ拡張
transformed_train_dataset = TransformDataset(
train_dataset, Transform(model.coder, model.insize, model.mean))
# イテレーターの設定
train_iter = chainer.iterators.MultiprocessIterator(
transformed_train_dataset, batchsize)
valid_iter = chainer.iterators.SerialIterator(valid_dataset,
batchsize,
repeat=False,
shuffle=False)
# オプティマイザーの設定
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
# アップデーターの設定
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=gpu_id)
# トレーナーの設定
trainer = training.Trainer(updater, (epoch_max, 'epoch'), out_num)
trainer.extend(extensions.ExponentialShift('lr',
lr_decay_rate,
init=initial_lr),
trigger=triggers.ManualScheduleTrigger(
lr_decay_timing, 'epoch'))
trainer.extend(DetectionVOCEvaluator(valid_iter,
model,
use_07_metric=False,
label_names=voc_labels),
trigger=(1, 'epoch'))
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map', 'elapsed_time'
]),
trigger=log_interval)
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(
['main/loss', 'main/loss/loc', 'main/loss/conf'],
'epoch',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport(['validation/main/map'],
'epoch',
file_name='accuracy.png'))
trainer.extend(
extensions.snapshot(filename='snapshot_epoch_{.updater.epoch}.npz'),
trigger=(10, 'epoch'))
# 途中で止めた学習を再開する場合は、trainerにスナップショットをロードして再開する
# serializers.load_npz('results/snapshot_epoch_100.npz', trainer)
# 学習実行
trainer.run()
# 学習データの保存
model.to_cpu()
serializers.save_npz('my_ssd_model.npz', model)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpus', type=str, default="-1")
parser.add_argument('--batchsize', type=int, default=2)
parser.add_argument('--epoch', type=int, default=10)
parser.add_argument('--mini', action="store_true")
args = parser.parse_args()
gpus = list(filter(lambda x: x >= 0, map(int, args.gpus.split(","))))
num_class = len(voc_bbox_label_names)
data_augmentation_transform = DataAugmentationTransform(512)
center_detection_transform = CenterDetectionTransform(512, num_class, 4)
train = TransformDataset(
ConcatenatedDataset(
VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')
),
data_augmentation_transform
)
train = TransformDataset(train, center_detection_transform)
if args.mini:
train = datasets.SubDataset(train, 0, 100)
train_iter = chainer.iterators.MultiprocessIterator(train, args.batchsize)
test = VOCBboxDataset(
year='2007', split='test',
use_difficult=True, return_difficult=True)
if args.mini:
test = datasets.SubDataset(test, 0, 20)
test_iter = chainer.iterators.SerialIterator(
test, args.batchsize // len(gpus), repeat=False, shuffle=False)
detector = CenterDetector(HourglassNet, 512, num_class)
train_chain = CenterDetectorTrain(detector, 1, 0.1, 1)
gpus.sort()
first_gpu = gpus[0]
remain_gpu = gpus[1:]
train_chain.to_gpu(first_gpu)
optimizer = Adam(amsgrad=True)
optimizer.setup(train_chain)
devices = {
"main": first_gpu
}
for i, gpu in enumerate(remain_gpu):
devices[f"{i + 2}"] = gpu
updater = training.updaters.ParallelUpdater(
train_iter,
optimizer,
devices=devices,
)
log_interval = 1, 'epoch'
trainer = Trainer(updater, (args.epoch, 'epoch'))
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
[
'epoch', 'iteration', 'lr',
'main/loss', 'main/hm_loss', 'main/wh_loss', 'main/offset_loss',
'validation/main/map',
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(
DetectionVOCEvaluator(
test_iter, detector, use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=(1, 'epoch'))
trainer.extend(
extensions.snapshot_object(detector, 'detector{.updater.epoch:03}.npz'),
trigger=(1, 'epoch')
)
trainer.run()
return loss
train_chain = MultiboxTrainChain(model)
optimizer = chainer.optimizers.MomentumSGD(1e-3)
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
evaluator = DetectionVOCEvaluator(
test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names,
)
class save_model(training.Extension): #ver 1.2
def __init__(self,
model,
save_name,
before_iter=0,
saved_dir='saved_model/',
save_after=0):
self.model = model
self.save_name = save_name
self.saved_dir = saved_dir
self.before_iter = before_iter
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd300')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--test-batchsize', type=int, default=16)
parser.add_argument('--iteration', type=int, default=120000)
parser.add_argument('--step', type=int, nargs='*', default=[80000, 100000])
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
parser.add_argument('--dtype',
type=str,
choices=dtypes.keys(),
default='float32',
help='Select the data type of the model')
parser.add_argument('--model-dir',
default=None,
type=str,
help='Where to store models')
parser.add_argument('--dataset-dir',
default=None,
type=str,
help='Where to store datasets')
parser.add_argument('--dynamic-interval',
default=None,
type=int,
help='Interval for dynamic loss scaling')
parser.add_argument('--init-scale',
default=1,
type=float,
help='Initial scale for ada loss')
parser.add_argument('--loss-scale-method',
default='approx_range',
type=str,
help='Method for adaptive loss scaling')
parser.add_argument('--scale-upper-bound',
default=32800,
type=float,
help='Hard upper bound for each scale factor')
parser.add_argument('--accum-upper-bound',
default=32800,
type=float,
help='Accumulated upper bound for all scale factors')
parser.add_argument('--update-per-n-iteration',
default=100,
type=int,
help='Update the loss scale value per n iteration')
parser.add_argument('--snapshot-per-n-iteration',
default=10000,
type=int,
help='The frequency of taking snapshots')
parser.add_argument('--n-uf', default=1e-3, type=float)
parser.add_argument('--nosanity-check', default=False, action='store_true')
parser.add_argument('--nouse-fp32-update',
default=False, action='store_true')
parser.add_argument('--profiling', default=False, action='store_true')
parser.add_argument('--verbose',
action='store_true',
default=False,
help='Verbose output')
args = parser.parse_args()
# Setting data types
if args.dtype != 'float32':
chainer.global_config.use_cudnn = 'never'
chainer.global_config.dtype = dtypes[args.dtype]
print('==> Setting the data type to {}'.format(args.dtype))
# Initialize model
if args.model == 'ssd300':
model = SSD300(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
# Apply adaptive loss scaling
recorder = AdaLossRecorder(sample_per_n_iter=100)
profiler = Profiler()
sanity_checker = SanityChecker(check_per_n_iter=100) if not args.nosanity_check else None
# Update the model to support AdaLoss
# TODO: refactorize
model_ = AdaLossScaled(
model,
init_scale=args.init_scale,
cfg={
'loss_scale_method': args.loss_scale_method,
'scale_upper_bound': args.scale_upper_bound,
'accum_upper_bound': args.accum_upper_bound,
'update_per_n_iteration': args.update_per_n_iteration,
'recorder': recorder,
'profiler': profiler,
'sanity_checker': sanity_checker,
'n_uf_threshold': args.n_uf,
},
transforms=[
AdaLossTransformLinear(),
AdaLossTransformConvolution2D(),
],
verbose=args.verbose)
# Finalize the model
train_chain = MultiboxTrainChain(model_)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
cp.random.seed(0)
# NOTE: we have to transfer modules explicitly to GPU
model.coder.to_gpu()
model.extractor.to_gpu()
model.multibox.to_gpu()
# Prepare dataset
if args.model_dir is not None:
chainer.dataset.set_dataset_root(args.model_dir)
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
Transform(model.coder, model.insize, model.mean, dtype=dtypes[args.dtype]))
# train_iter = chainer.iterators.MultiprocessIterator(
# train, args.batchsize) # , n_processes=8, n_prefetch=2)
train_iter = chainer.iterators.MultithreadIterator(train, args.batchsize)
# train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
args.test_batchsize,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
if args.dtype == 'mixed16':
if not args.nouse_fp32_update:
print('==> Using FP32 update for dtype=mixed16')
optimizer.use_fp32_update() # by default use fp32 update
# HACK: support skipping update by existing loss scaling functionality
if args.dynamic_interval is not None:
optimizer.loss_scaling(interval=args.dynamic_interval, scale=None)
else:
optimizer.loss_scaling(interval=float('inf'), scale=None)
optimizer._loss_scale_max = 1.0 # to prevent actual loss scaling
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
args.out)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger(
args.step, 'iteration'))
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.observe_value(
'loss_scale',
lambda trainer: trainer.updater.get_optimizer('main')._loss_scale),
trigger=log_interval)
metrics = [
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]
if args.dynamic_interval is not None:
metrics.insert(2, 'loss_scale')
trainer.extend(extensions.PrintReport(metrics), trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(args.iteration, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
hook = AdaLossMonitor(sample_per_n_iter=100,
verbose=args.verbose,
includes=['Grad', 'Deconvolution'])
recorder.trainer = trainer
hook.trainer = trainer
with ExitStack() as stack:
stack.enter_context(hook)
trainer.run()
recorder.export().to_csv(os.path.join(args.out, 'loss_scale.csv'))
profiler.export().to_csv(os.path.join(args.out, 'profile.csv'))
if sanity_checker:
sanity_checker.export().to_csv(os.path.join(args.out, 'sanity_check.csv'))
hook.export_history().to_csv(os.path.join(args.out, 'grad_stats.csv'))
def run(input_dir, test_dir, output, batch_size,
iterator='SerialIterator',
device=-1, pretrained_model='', save_trigger=10000,
test_trigger=1000,
parser_module='XMLParser',
train_module='MultiboxTrainChain',
model_module='chainercv.links.SSD300'):
pretrained_model = join(PROJECT_DIR, pretrained_model)
if pretrained_model and os.path.isfile(pretrained_model):
print('Pretrained model {} loaded.'.format(pretrained_model))
else:
print('Pretrained model file not found, ' +
'using imagenet as default.')
pretrained_model = 'imagenet'
parser = _import_module('multibuildingdetector.parsers.{}', parser_module)
model = _import_class(model_module)(n_fg_class=len(parser.LABEL_NAMES),
pretrained_model=pretrained_model)
model.use_preset('evaluate')
train_chain = _import_class('multibuildingdetector.trainchains.{}'
.format(train_module))(model)
if device >= 0:
chainer.cuda.get_device_from_id(device).use()
model.to_gpu()
train, test = load_train_test_set(input_dir, test_dir, parser)
augmented_train = TransformDataset(
train,
ImageAugmentation(model.coder, model.insize, model.mean))
train_iter = getattr(chainer.iterators, iterator)(augmented_train,
batch_size)
optimizer = chainer.optimizers.Adam()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
if device is None:
updater = chainer.training.StandardUpdater(train_iter, optimizer)
else:
updater = chainer.training.StandardUpdater(train_iter, optimizer,
device=device)
trainer = chainer.training.Trainer(updater, (120000, 'iteration'), output)
log_fields = ['main/' + x for x in train_chain.loss_labels]
if train_module == 'MultiboxTrainChain':
test_iter = chainer.iterators.SerialIterator(
test,
batch_size, repeat=False, shuffle=False)
trainer.extend(
DetectionVOCEvaluator(
test_iter, model, use_07_metric=True,
label_names=parser.LABEL_NAMES),
trigger=(test_trigger, 'iteration'))
log_fields.append('validation/main/map')
else:
triplet_test = TransformDataset(
test,
ImageAugmentation(model.coder, model.insize, model.mean,
augment=False))
test_iter = chainer.iterators.SerialIterator(
triplet_test,
batch_size, repeat=False, shuffle=False)
trainer.extend(
TripletEvaluator(
test_iter, model,
label_names=parser.LABEL_NAMES,
save_plt=True,
save_path=output),
trigger=(test_trigger, 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'lr',
*log_fields]))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(save_trigger, 'iteration'))
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.run()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr)
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.iteration, "iteration"),
args.out)
val_interval = args.val_iter, "iteration"
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=label_names),
trigger=val_interval)
log_interval = args.log_iter, "iteration"
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(), trigger=val_interval)
trainer.extend(extensions.snapshot_object(
chainer.cuda.get_device_from_id(0).use()
model.to_gpu()
# Data Augmentation
train = TransformDataset(
ConcatenatedDataset(
VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')
),
Transform(model.coder, model.insize, model.mean)
)
test = VOCBboxDataset(year='2007', split='test', use_difficult=True, return_difficult=True)
train_iter = iterators.SerialIterator(train, args.batchsize)
test_iter = iterators.SerialIterator(test, args.batchsize, repeat=False, shuffle=False)
# Evaluator
evaluator = DetectionVOCEvaluator(test_iter, net.predictor, use_07_metric=True, label_names=voc_bbox_label_names)
trainer = create_trainer(train_iter, net, args.gpu_id, args.initial_lr,
args.weight_decay, args.freeze_layer, args.small_lr_layers,
args.small_initial_lr, args.num_epochs_or_iter,
args.epoch_or_iter, args.save_dir)
if args.load_path:
chainer.serializers.load_npz(args.load_path, trainer)
trainer_extend(trainer, net, evaluator, args.small_lr_layers,
args.lr_decay_rate, args.lr_decay_epoch,
args.epoch_or_iter, args.save_trainer_interval)
trainer.run()
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(
train_iter, optimizer, device=gpu_id)
trainer = training.Trainer(
updater,
(training_epoch, 'epoch'), out)
trainer.extend(
extensions.ExponentialShift('lr', lr_decay_rate, init=initial_lr),
trigger=triggers.ManualScheduleTrigger(lr_decay_timing, 'epoch'))
trainer.extend(
DetectionVOCEvaluator(
valid_iter, model, use_07_metric=False,
label_names=bball_labels),
trigger=(1, 'epoch'))
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'lr',
'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map', 'elapsed_time']),
trigger=log_interval)
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(
['main/loss', 'main/loss/loc', 'main/loss/conf'],
'epoch', file_name='loss.png'))
trainer.extend(
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd300')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--np', type=int, default=8)
parser.add_argument('--test-batchsize', type=int, default=16)
parser.add_argument('--iteration', type=int, default=120000)
parser.add_argument('--step', type=int, nargs='*', default=[80000, 100000])
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
parser.add_argument('--dtype',
type=str,
choices=dtypes.keys(),
default='float32',
help='Select the data type of the model')
parser.add_argument('--model-dir',
default=None,
type=str,
help='Where to store models')
parser.add_argument('--dataset-dir',
default=None,
type=str,
help='Where to store datasets')
parser.add_argument('--dynamic-interval',
default=None,
type=int,
help='Interval for dynamic loss scaling')
parser.add_argument('--init-scale',
default=1,
type=float,
help='Initial scale for ada loss')
parser.add_argument('--loss-scale-method',
default='approx_range',
type=str,
help='Method for adaptive loss scaling')
parser.add_argument('--scale-upper-bound',
default=16,
type=float,
help='Hard upper bound for each scale factor')
parser.add_argument('--accum-upper-bound',
default=1024,
type=float,
help='Accumulated upper bound for all scale factors')
parser.add_argument('--update-per-n-iteration',
default=1,
type=int,
help='Update the loss scale value per n iteration')
parser.add_argument('--snapshot-per-n-iteration',
default=10000,
type=int,
help='The frequency of taking snapshots')
parser.add_argument('--n-uf', default=1e-3, type=float)
parser.add_argument('--nosanity-check', default=False, action='store_true')
parser.add_argument('--nouse-fp32-update',
default=False,
action='store_true')
parser.add_argument('--profiling', default=False, action='store_true')
parser.add_argument('--verbose',
action='store_true',
default=False,
help='Verbose output')
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()
comm = chainermn.create_communicator('pure_nccl')
device = comm.intra_rank
# Set up workspace
# 12 GB GPU RAM for workspace
chainer.cuda.set_max_workspace_size(16 * 1024 * 1024 * 1024)
chainer.global_config.cv_resize_backend = 'cv2'
# Setup the data type
# when initializing models as follows, their data types will be casted.
# Weethave to forbid the usage of cudnn
if args.dtype != 'float32':
chainer.global_config.use_cudnn = 'never'
chainer.global_config.dtype = dtypes[args.dtype]
print('==> Setting the data type to {}'.format(args.dtype))
if args.model_dir is not None:
chainer.dataset.set_dataset_root(args.model_dir)
if args.model == 'ssd300':
model = SSD300(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
######################################
# Setup model
#######################################
# Apply ada loss transform
recorder = AdaLossRecorder(sample_per_n_iter=100)
profiler = Profiler()
sanity_checker = SanityChecker(
check_per_n_iter=100) if not args.nosanity_check else None
# Update the model to support AdaLoss
# TODO: refactorize
model_ = AdaLossScaled(
model,
init_scale=args.init_scale,
cfg={
'loss_scale_method': args.loss_scale_method,
'scale_upper_bound': args.scale_upper_bound,
'accum_upper_bound': args.accum_upper_bound,
'update_per_n_iteration': args.update_per_n_iteration,
'recorder': recorder,
'profiler': profiler,
'sanity_checker': sanity_checker,
'n_uf_threshold': args.n_uf,
# 'power_of_two': False,
},
transforms=[
AdaLossTransformLinear(),
AdaLossTransformConvolution2D(),
],
verbose=args.verbose)
if comm.rank == 0:
print(model)
train_chain = MultiboxTrainChain(model_, comm=comm)
chainer.cuda.get_device_from_id(device).use()
# to GPU
model.coder.to_gpu()
model.extractor.to_gpu()
model.multibox.to_gpu()
shared_mem = 100 * 1000 * 1000 * 4
if args.dataset_dir is not None:
chainer.dataset.set_dataset_root(args.dataset_dir)
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
('img', 'mb_loc', 'mb_label'),
Transform(model.coder,
model.insize,
model.mean,
dtype=dtypes[args.dtype]))
if comm.rank == 0:
indices = np.arange(len(train))
else:
indices = None
indices = chainermn.scatter_dataset(indices, comm, shuffle=True)
train = train.slice[indices]
train_iter = chainer.iterators.MultiprocessIterator(train,
args.batchsize //
comm.size,
n_processes=8,
n_prefetch=2,
shared_mem=shared_mem)
if comm.rank == 0: # NOTE: only performed on the first device
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
args.test_batchsize,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainermn.create_multi_node_optimizer(
chainer.optimizers.MomentumSGD(), comm)
if args.dtype == 'mixed16':
if not args.nouse_fp32_update:
print('==> Using FP32 update for dtype=mixed16')
optimizer.use_fp32_update() # by default use fp32 update
# HACK: support skipping update by existing loss scaling functionality
if args.dynamic_interval is not None:
optimizer.loss_scaling(interval=args.dynamic_interval, scale=None)
else:
optimizer.loss_scaling(interval=float('inf'), scale=None)
optimizer._loss_scale_max = 1.0 # to prevent actual loss scaling
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=device)
# if args.dtype == 'mixed16':
# updater.loss_scale = 8
iteration_interval = (args.iteration, 'iteration')
trainer = training.Trainer(updater, iteration_interval, args.out)
# trainer.extend(extensions.ExponentialShift('lr', 0.1, init=args.lr),
# trigger=triggers.ManualScheduleTrigger(
# args.step, 'iteration'))
if args.batchsize != 32:
warmup_attr_ratio = 0.1
# NOTE: this is confusing but it means n_iter
warmup_n_epoch = 1000
lr_shift = chainerlp.extensions.ExponentialShift(
'lr',
0.1,
init=args.lr * warmup_attr_ratio,
warmup_attr_ratio=warmup_attr_ratio,
warmup_n_epoch=warmup_n_epoch,
schedule=args.step)
trainer.extend(lr_shift, trigger=(1, 'iteration'))
if comm.rank == 0:
if not args.profiling:
trainer.extend(DetectionVOCEvaluator(
test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.observe_value(
'loss_scale',
lambda trainer: trainer.updater.get_optimizer('main')._loss_scale),
trigger=log_interval)
metrics = [
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]
if args.dynamic_interval is not None:
metrics.insert(2, 'loss_scale')
trainer.extend(extensions.PrintReport(metrics), trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(),
trigger=(args.snapshot_per_n_iteration, 'iteration'))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(args.iteration, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
hook = AdaLossMonitor(sample_per_n_iter=100,
verbose=args.verbose,
includes=['Grad', 'Deconvolution'])
recorder.trainer = trainer
hook.trainer = trainer
with ExitStack() as stack:
if comm.rank == 0:
stack.enter_context(hook)
trainer.run()
# store recorded results
if comm.rank == 0: # NOTE: only export in the first rank
recorder.export().to_csv(os.path.join(args.out, 'loss_scale.csv'))
profiler.export().to_csv(os.path.join(args.out, 'profile.csv'))
if sanity_checker:
sanity_checker.export().to_csv(
os.path.join(args.out, 'sanity_check.csv'))
hook.export_history().to_csv(os.path.join(args.out, 'grad_stats.csv'))
def main():
args = parse_args()
res = Resource(args, train=True)
train, test, train_gt, test_gt = load_train_test(
train_dir=const.PREPROCESSED_TRAIN_DIR, gt_dir=const.XML_DIR)
res.log_info(f'Train: {len(train)}, test: {len(test)}')
model = ARCHS[args.model](n_fg_class=len(const.LABELS),
pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
train_dataset = TransformDataset(
ISIC2018Task1Dataset(train, train_gt),
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.MultithreadIterator(
train_dataset, args.batchsize, n_threads=args.loaderjob)
test_dataset = TransformDataset(
ISIC2018Task1Dataset(test, test_gt),
Transform(model.coder, model.insize, model.mean))
test_iter = chainer.iterators.MultithreadIterator(test_dataset,
args.batchsize,
shuffle=False,
repeat=False,
n_threads=args.loaderjob)
optimizer = chainer.optimizers.Adam()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.out)
trainer.extend(
DetectionVOCEvaluator(test_iter,
model,
use_07_metric=False,
label_names=const.LABELS))
trainer.extend(extensions.LogReport())
trainer.extend(extensions.observe_lr())
trainer.extend(
extensions.PrintReport([
'epoch', 'iteration', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]))
trainer.extend(extensions.ProgressBar(update_interval=10))
snapshot_trigger = triggers.MaxValueTrigger(key='validation/main/map')
snapshot_object_trigger = triggers.MaxValueTrigger(
key='validation/main/map')
trainer.extend(extensions.snapshot(filename='snapshot_best.npz'),
trigger=snapshot_trigger)
trainer.extend(extensions.snapshot_object(model, 'model_best.npz'),
trigger=snapshot_object_trigger)
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
# save last model
chainer.serializers.save_npz(os.path.join(args.out, 'snapshot_last.npz'),
trainer)
chainer.serializers.save_npz(os.path.join(args.out, 'model_last.npz'),
model)
def train(train_data, val_data, label_names,
iteration, lr, step_points,
batchsize, gpu, out, val_iteration,
log_iteration, loaderjob,
resume):
model = SSD300(n_fg_class=len(label_names), pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model) # alpha and k?
if gpu >= 0:
chainer.cuda.get_device_from_id(gpu).use()
model.to_gpu()
train_data = TransformDataset(
train_data,
Transform(model.coder, model.insize, model.mean))
if loaderjob <= 0:
train_iter = chainer.iterators.SerialIterator(train_data, batchsize)
else:
train_iter = chainer.iterators.MultiprocessIterator(train_data, batchsize) # , n_processes=min((loaderjob, batchsize)))
val_iter = chainer.iterators.SerialIterator(val_data, batchsize, repeat=False, shuffle=False) # think about repeat
# initial lr is set to 1e-4 (default run_train) by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.StandardUpdater(train_iter, optimizer, device=gpu)
trainer = training.Trainer(updater, (iteration, 'iteration'), out)
trainer.extend(
extensions.ExponentialShift('lr', 0.1, init=lr),
trigger=triggers.ManualScheduleTrigger(step_points, 'iteration'))
val_interval = (val_iteration, 'iteration')
trainer.extend(
DetectionVOCEvaluator(
val_iter, model, use_07_metric=True,
label_names=label_names),
trigger=val_interval)
###
# Logging
log_interval = log_iteration, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport(
['epoch', 'iteration', 'lr',
'main/loss', 'main/loss/loc', 'main/loss/conf',
'validation/main/map']),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(), trigger=val_interval)
trainer.extend(
extensions.snapshot_object(model, 'model_iter_{.updater.iteration}'),
trigger=val_interval)
###
if resume:
serializers.load_npz(resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd300')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--iteration', type=int, default=120000)
parser.add_argument('--step', type=int, nargs='*', default=[80000, 100000])
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
args = parser.parse_args()
if args.model == 'ssd300':
model = SSD300(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.MultiprocessIterator(train, args.batchsize)
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
args.out)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger(
args.step, 'iteration'))
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(args.iteration, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd300')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--labelnum', type=int, default=50)
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
parser.add_argument('--image_label',
'-il',
help='Path to training image-label list file')
parser.add_argument('--bbox', help='Path to training bbox list file')
parser.add_argument('--image_label_test',
'-ilt',
help='Path to training image-label list file')
parser.add_argument('--bbox_test', help='Path to training bbox list file')
parser.add_argument('--image_root',
'-TR',
default='.',
help='Root directory path of image files')
args = parser.parse_args()
comm = chainermn.create_communicator('naive')
if comm.mpi_comm.rank == 0:
print('==========================================')
print('Num process (COMM_WORLD): {}'.format(MPI.COMM_WORLD.Get_size()))
if args.model == 'ssd300':
model = SSD300(n_fg_class=args.labelnum, pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(n_fg_class=args.labelnum, pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
from test_datasets import DeepFashionBboxDataset
if comm.rank == 0:
train = DeepFashionBboxDataset(args.bbox, args.image_label,
args.image_root)
test = DeepFashionBboxDataset(args.bbox_test, args.image_label_test,
args.image_root)
train = TransformDataset(
train, Transform(model.coder, model.insize, model.mean))
else:
train, test = None, None
train = chainermn.scatter_dataset(train, comm, shuffle=True)
test = chainermn.scatter_dataset(test, comm, shuffle=True)
train_iter = chainer.iterators.MultiprocessIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainer.optimizers.MomentumSGD()
optimizer = chainermn.create_multi_node_optimizer(optimizer, comm)
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (120000, 'iteration'), args.out)
checkpoint_interval = (1000, 'iteration')
checkpointer = chainermn.create_multi_node_checkpointer(
name='imagenet-example', comm=comm)
checkpointer.maybe_load(trainer, optimizer)
trainer.extend(checkpointer, trigger=checkpoint_interval)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger([80000, 100000],
'iteration'))
evaluator = DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names)
evaluator = chainermn.create_multi_node_evaluator(evaluator, comm)
trainer.extend(evaluator, trigger=(10000, 'iteration'))
if comm.rank == 0:
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(), trigger=(10000, 'iteration'))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(120000, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: Faster R-CNN')
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--lr', '-l', type=float, default=1e-3)
parser.add_argument('--out',
'-o',
default='result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--step_size', '-ss', type=int, default=50000)
parser.add_argument('--iteration', '-i', type=int, default=70000)
args = parser.parse_args()
np.random.seed(args.seed)
train_data = VOCDetectionDataset(split='trainval', year='2007')
test_data = VOCDetectionDataset(split='test',
year='2007',
use_difficult=True,
return_difficult=True)
faster_rcnn = FasterRCNNVGG16(n_fg_class=len(voc_detection_label_names),
pretrained_model='imagenet')
faster_rcnn.use_preset('evaluate')
model = FasterRCNNTrainChain(faster_rcnn)
if args.gpu >= 0:
model.to_gpu(args.gpu)
chainer.cuda.get_device(args.gpu).use()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(rate=0.0005))
def transform(in_data):
img, bbox, label = in_data
_, H, W = img.shape
img = faster_rcnn.prepare(img)
_, o_H, o_W = img.shape
scale = o_H / H
bbox = transforms.resize_bbox(bbox, (H, W), (o_H, o_W))
# horizontally flip
img, params = transforms.random_flip(img,
x_random=True,
return_param=True)
bbox = transforms.flip_bbox(bbox, (o_H, o_W), x_flip=params['x_flip'])
return img, bbox, label, scale
train_data = TransformDataset(train_data, transform)
train_iter = chainer.iterators.MultiprocessIterator(train_data,
batch_size=1,
n_processes=None,
shared_mem=100000000)
test_iter = chainer.iterators.SerialIterator(test_data,
batch_size=1,
repeat=False,
shuffle=False)
updater = chainer.training.updater.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
out=args.out)
trainer.extend(extensions.snapshot_object(model.faster_rcnn,
'snapshot_model.npz'),
trigger=(args.iteration, 'iteration'))
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(args.step_size, 'iteration'))
log_interval = 20, 'iteration'
plot_interval = 3000, 'iteration'
print_interval = 20, 'iteration'
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'iteration',
'epoch',
'elapsed_time',
'lr',
'main/loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/rpn_loc_loss',
'main/rpn_cls_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(
DetectionVOCEvaluator(test_iter,
model.faster_rcnn,
use_07_metric=True,
label_names=voc_detection_label_names),
trigger=ManualScheduleTrigger([args.step_size, args.iteration],
'iteration'),
invoke_before_training=False)
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--batchsize', type=int, default=2)
parser.add_argument('--epoch', type=int, default=10)
parser.add_argument('--mini', action="store_true")
parser.add_argument('--input_size', type=int, default=512)
args = parser.parse_args()
dtype = np.float32
num_class = len(voc_bbox_label_names)
data_augmentation_transform = DataAugmentationTransform(args.input_size)
center_detection_transform = CenterDetectionTransform(args.input_size,
num_class,
4,
dtype=dtype)
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
data_augmentation_transform)
train = TransformDataset(train, center_detection_transform)
if args.mini:
train = datasets.SubDataset(train, 0, 100)
train_iter = chainer.iterators.MultiprocessIterator(train, args.batchsize)
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
if args.mini:
test = datasets.SubDataset(test, 0, 20)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
detector = CenterDetector(HourglassNet,
args.input_size,
num_class,
dtype=dtype)
#detector = CenterDetector(SimpleCNN, args.input_size, num_class)
train_chain = CenterDetectorTrain(detector, 1, 0.1, 1)
#train_chain = CenterDetectorTrain(detector, 1, 0, 0)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
train_chain.to_gpu(args.gpu)
optimizer = Adam(alpha=1.25e-4)
#optimizer = SGD()
optimizer.setup(train_chain)
updater = StandardUpdater(train_iter, optimizer, device=args.gpu)
log_interval = 1, 'epoch'
log_interval_mini = 500, 'iteration'
trainer = Trainer(updater, (args.epoch, 'epoch'), out=f"result{args.gpu}")
trainer.extend(extensions.LogReport(trigger=log_interval_mini))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch',
'iteration',
'lr',
'main/loss',
'main/hm_loss',
'main/wh_loss',
'main/offset_loss',
'main/hm_mae',
'main/hm_pos_loss',
'main/hm_neg_loss',
'validation/main/map',
]),
trigger=log_interval_mini)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(DetectionVOCEvaluator(test_iter,
detector,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=(1, 'epoch'))
trainer.extend(extensions.snapshot_object(
detector, 'detector{.updater.epoch:03}.npz'),
trigger=(1, 'epoch'))
trainer.run()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model',
choices=('ssd300', 'ssd512'),
default='ssd300')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--test-batchsize', type=int, default=16)
parser.add_argument('--iteration', type=int, default=120000)
parser.add_argument('--step', type=int, nargs='*', default=[80000, 100000])
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
args = parser.parse_args()
comm = chainermn.create_communicator()
device = comm.intra_rank
if args.model == 'ssd300':
model = SSD300(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(n_fg_class=len(voc_bbox_label_names),
pretrained_model='imagenet')
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
chainer.cuda.get_device_from_id(device).use()
model.to_gpu()
train = TransformDataset(
ConcatenatedDataset(VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval')),
('img', 'mb_loc', 'mb_label'),
Transform(model.coder, model.insize, model.mean))
if comm.rank == 0:
indices = np.arange(len(train))
else:
indices = None
indices = chainermn.scatter_dataset(indices, comm, shuffle=True)
train = train.slice[indices]
# http://chainermn.readthedocs.io/en/latest/tutorial/tips_faqs.html#using-multiprocessiterator
if hasattr(multiprocessing, 'set_start_method'):
multiprocessing.set_start_method('forkserver')
train_iter = chainer.iterators.MultiprocessIterator(train,
args.batchsize //
comm.size,
n_processes=2)
if comm.rank == 0:
test = VOCBboxDataset(year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
args.test_batchsize,
repeat=False,
shuffle=False)
# initial lr is set to 1e-3 by ExponentialShift
optimizer = chainermn.create_multi_node_optimizer(
chainer.optimizers.MomentumSGD(), comm)
optimizer.setup(train_chain)
for param in train_chain.params():
if param.name == 'b':
param.update_rule.add_hook(GradientScaling(2))
else:
param.update_rule.add_hook(WeightDecay(0.0005))
updater = training.updaters.StandardUpdater(train_iter,
optimizer,
device=device)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
args.out)
trainer.extend(extensions.ExponentialShift('lr', 0.1, init=1e-3),
trigger=triggers.ManualScheduleTrigger(
args.step, 'iteration'))
if comm.rank == 0:
trainer.extend(DetectionVOCEvaluator(test_iter,
model,
use_07_metric=True,
label_names=voc_bbox_label_names),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
log_interval = 10, 'iteration'
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'main/loss/loc',
'main/loss/conf', 'validation/main/map'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.extend(extensions.snapshot(),
trigger=triggers.ManualScheduleTrigger(
args.step + [args.iteration], 'iteration'))
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=(args.iteration, 'iteration'))
if args.resume:
serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
parser = argparse.ArgumentParser(
description='ChainerCV training example: Faster R-CNN')
parser.add_argument(
'--dataset_path',
'-path',
type=str,
default="/home/takagi.kazunari/projects/datasets/SUNRGBD_2DBB_fixed")
parser.add_argument('--gpu', '-g', type=int, default=-1)
parser.add_argument('--lr', '-l', type=float, default=1e-3)
parser.add_argument('--out',
'-o',
default='sunrgbd_result',
help='Output directory')
parser.add_argument('--seed', '-s', type=int, default=0)
parser.add_argument('--step_size', '-ss', type=int, default=50000)
parser.add_argument('--iteration', '-i', type=int, default=70000)
args = parser.parse_args()
np.random.seed(args.seed)
train_data = SUNRGBDDataset(args.dataset_path, mode="train")
test_data = SUNRGBDDataset(args.dataset_path, mode="test")
sunrgbd_bbox_label_names = train_data.get_dataset_label()
faster_rcnn = FasterRCNNVGG16(n_fg_class=len(sunrgbd_bbox_label_names),
pretrained_model='imagenet')
faster_rcnn.use_preset('evaluate')
model = FasterRCNNTrainChain(faster_rcnn)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
optimizer = chainer.optimizers.MomentumSGD(lr=args.lr, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer_hooks.WeightDecay(rate=0.0005))
train_data = TransformDataset(train_data, Transform(faster_rcnn))
train_iter = chainer.iterators.MultiprocessIterator(train_data,
batch_size=1,
n_processes=None,
shared_mem=100000000)
test_iter = chainer.iterators.SerialIterator(test_data,
batch_size=1,
repeat=False,
shuffle=False)
updater = chainer.training.updaters.StandardUpdater(train_iter,
optimizer,
device=args.gpu)
now_time = str(datetime.datetime.today()).replace(" ", "_")
save_dir = osp.join(args.out, now_time)
trainer = training.Trainer(updater, (args.iteration, 'iteration'),
out=save_dir)
#save_iteration = [i for i in range(100, args.iteration, args.step_size)]
weight_save_interval = 5000, 'iteration'
evaluation_interval = 10000, 'iteration'
trainer.extend(extensions.snapshot_object(
model.faster_rcnn, 'sunrgbd_model_{.updater.iteration}.npz'),
trigger=weight_save_interval)
trainer.extend(extensions.ExponentialShift('lr', 0.1),
trigger=(args.step_size, 'iteration'))
log_interval = 20, 'iteration'
plot_interval = 10, 'iteration'
print_interval = 20, 'iteration'
trainer.extend(chainer.training.extensions.observe_lr(),
trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'iteration',
'epoch',
'elapsed_time',
'lr',
'main/loss',
'main/roi_loc_loss',
'main/roi_cls_loss',
'main/rpn_loc_loss',
'main/rpn_cls_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)
#do_evaluation_iteration = [i for i in range(0, args.iteration, 500)]
trainer.extend(DetectionVOCEvaluator(test_iter,
model.faster_rcnn,
use_07_metric=True,
label_names=sunrgbd_bbox_label_names),
trigger=evaluation_interval)
trainer.extend(extensions.dump_graph('main/loss'))
trainer.run()
