def setup_dataset(cfg, split):
if split == 'train':
dataset_type = cfg.dataset.train
elif split == 'eval':
dataset_type = cfg.dataset.eval
else:
raise ValueError()
if dataset_type == 'COCO':
if split == 'train':
dataset = COCOBboxDataset(split='train', year='2017')
elif split == 'eval':
dataset = COCOBboxDataset(split='val',
year='2017',
use_crowded=True,
return_area=True,
return_crowded=True)
else:
raise ValueError()
elif dataset_type == 'VOC':
if split == 'train':
dataset = ConcatenatedDataset(
VOCBboxDataset(year='2007', split='trainval'),
VOCBboxDataset(year='2012', split='trainval'))
elif split == 'eval':
dataset = VOCBboxDataset(split='test',
year='2007',
use_difficult=True,
return_difficult=True)
else:
raise ValueError()
else:
raise ValueError()
return dataset
def _wrapper(self, to_double=False, bit=10, mulaw=True):
speaker_size = 4
model = _create_model(
local_size=0,
speaker_size=speaker_size,
)
datasets = [
SignWaveDataset(
sampling_rate=sampling_rate,
sampling_length=sampling_length,
to_double=to_double,
bit=bit,
mulaw=mulaw,
frequency=(i + 1) * 110,
) for i in range(speaker_size)
]
dataset = SpeakerWavesDataset(
wave_dataset=ConcatenatedDataset(*datasets),
speaker_nums=list(
chain.from_iterable([i] * len(d)
for i, d in enumerate(datasets))),
)
updater, reporter = setup_support(batch_size, gpu, model, dataset)
trained_nll = _get_trained_nll()
def _first_hook(o):
self.assertTrue(o['main/nll_coarse'].data > trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data > trained_nll)
def _last_hook(o):
self.assertTrue(o['main/nll_coarse'].data < trained_nll)
if to_double:
self.assertTrue(o['main/nll_fine'].data < trained_nll)
train_support(iteration, reporter, updater, _first_hook, _last_hook)
# save model
serializers.save_npz(
'/tmp/'
f'test_training_wavernn'
f'-to_double={to_double}'
f'-bit={bit}'
f'-mulaw={mulaw}'
f'-speaker_size={speaker_size}'
f'-iteration={iteration}.npz',
model.predictor,
)
def load_dataset(config):
train_config = config['train']
cl = get_class(train_config['module'])
train_loader = getattr(cl, train_config['name'])
train_data = train_loader(**train_config['args'])
if parse_dict(config, 'train2', None):
train_config = config['train2']
cl = get_class(train_config['module'])
train_loader = getattr(cl, train_config['name'])
train_data2 = train_loader(**train_config['args'])
train_data = ConcatenatedDataset(train_data, train_data2)
test_data = None
if parse_dict(config, 'valid', None):
test_config = config['valid']
cl = get_class(test_config['module'])
test_loader = getattr(cl, test_config['name'])
test_data = test_loader(**test_config['args'])
return train_data, test_data
def Dataset(fns, for_evaluate=False):
inputs = [
LazyInput(
path_wave=wave_paths[fn],
path_silence=silence_paths[fn],
path_local=local_paths[fn],
) for fn in fns
]
if not for_evaluate:
dataset = WavesDataset(
inputs=inputs,
sampling_length=config.sampling_length,
to_double=not config.only_coarse,
bit=config.bit_size,
mulaw=config.mulaw,
local_padding_size=config.local_padding_size,
gaussian_noise_sigma=config.gaussian_noise_sigma,
)
else:
dataset = NonEncodeWavesDataset(
inputs=inputs,
time_length=config.time_length_evaluate,
local_padding_time_length=config.
local_padding_time_length_evaluate,
)
if speaker_nums is not None:
dataset = SpeakerWavesDataset(
wave_dataset=dataset,
speaker_nums=[speaker_nums[fn] for fn in fns],
)
if for_evaluate:
dataset = ConcatenatedDataset(*([dataset] *
config.num_times_evaluate))
return dataset
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 setUp(self):
self.concatenated_dataset = ConcatenatedDataset(*self.datasets)
self.expected_dataset = [
sample for dataset in self.datasets for sample in dataset
]
train07 = VOCBboxDataset(data_dir='auto',
year='2007',
split='trainval',
use_difficult=True,
return_difficult=False)
train12 = VOCBboxDataset(data_dir='auto',
year='2012',
split='trainval',
use_difficult=True,
return_difficult=False)
train07_right = train07[len(train07) // 2:]
train12_right = train12[len(train12) // 2:]
train_right = ConcatenatedDataset(train07_right, train12_right)
train = TransformDataset(train_right,
_Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.SerialIterator(train,
batch_size,
repeat=True,
shuffle=True)
test = VOCBboxDataset(data_dir='auto',
year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
batch_size,
repeat=False,
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 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()
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
'--model', choices=('fpn', 'ssd300', 'ssd512'), default='fpn')
parser.add_argument('--batchsize', type=int, default=32)
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--out', default='result')
parser.add_argument('--data_dir', type=str, default='auto')
parser.add_argument('--dataset', choices=['voc', 'coco'], default='voc')
parser.add_argument('--lr', type=float, default=1e-3)
parser.add_argument('--init_scale', type=float, default=1e-2)
parser.add_argument('--resume')
args = parser.parse_args()
if args.dataset == 'voc':
train = ConcatenatedDataset(
VOCBboxDataset(
year='2007',
split='trainval',
data_dir=join(args.data_dir, 'VOCdevkit/VOC2007')
if args.data_dir != 'auto' else args.data_dir),
VOCBboxDataset(
year='2012',
split='trainval',
data_dir=join(args.data_dir, 'VOCdevkit/VOC2012')
if args.data_dir != 'auto' else args.data_dir))
test = VOCBboxDataset(
year='2007',
split='test',
use_difficult=True,
return_difficult=True,
data_dir=join(args.data_dir, 'VOCdevkit/VOC2007')
if args.data_dir != 'auto' else args.data_dir)
label_names = voc_bbox_label_names
elif args.dataset == 'coco':
# todo: use train+valminusminival(=coco2017train)
# https://github.com/chainer/chainercv/issues/651
train = COCOBboxDataset(data_dir=args.data_dir,
split='train')
test = COCOBboxDataset(data_dir=args.data_dir, split='val')
label_names = coco_bbox_label_names
if args.model == 'ssd300':
model = SSD300(
n_fg_class=len(label_names), pretrained_model='imagenet')
elif args.model == 'ssd512':
model = SSD512(
n_fg_class=len(label_names), pretrained_model='imagenet')
elif args.model == 'fpn':
model = FPNSSD(
n_fg_class=len(label_names), pretrained_model='imagenet', init_scale=args.init_scale)
model.use_preset('evaluate')
train_chain = MultiboxTrainChain(model)
if args.gpu >= 0:
chainer.backends.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
train = TransformDataset(
train,
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.MultithreadIterator(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.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=args.lr),
trigger=triggers.ManualScheduleTrigger([80000, 100000], 'iteration'))
trainer.extend(
DetectionVOCEvaluator(
test_iter,
model,
use_07_metric=True,
label_names=label_names),
trigger=(10000, 'iteration'))
log_interval = 100, '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()
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()
return bboxes, labels, scores
model.predict_each_box = predict_each_box.__get__(model)
train07 = VOCBboxDataset(data_dir='auto',
year='2007',
split='trainval',
use_difficult=True,
return_difficult=False)
train12 = VOCBboxDataset(data_dir='auto',
year='2012',
split='trainval',
use_difficult=True,
return_difficult=False)
train = ConcatenatedDataset(train07, train12)
test = VOCBboxDataset(data_dir='auto',
year='2007',
split='test',
use_difficult=True,
return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,
batch_size,
repeat=False,
shuffle=False)
print('image prepare')
if subset == 'left':
train07_left = train07[:len(train07) // 2]
test_left = test[:len(test) // 2]
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
'--model', choices=('ssd300', 'ssd512'), default='ssd300')
parser.add_argument('--batchsize', type=int, default=1)
parser.add_argument('--gpu', type=int, default=-1)
parser.add_argument('--out', default='result')
parser.add_argument('--resume')
args = parser.parse_args()
model = SSD300(
n_fg_class=len(inria_bbox_label_names),
pretrained_model='./ssd_vgg16_imagenet_2017_06_09.npz')
print("###n_fg_class= ", len(inria_bbox_label_names))
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(
INRIABboxDataset(data_dir='../INRIAPerson', split='Train')
),
Transform(model.coder, model.insize, model.mean))
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test = INRIABboxDataset(data_dir='../INRIAPerson', split='Test')
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.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (120000, 'iteration'), args.out)
trainer.extend(
extensions.ExponentialShift('lr', 0.1, init=1e-4),
trigger=triggers.ManualScheduleTrigger([80000, 100000], 'iteration'))
trainer.extend(
DetectionINRIAEvaluator(
test_iter, model, use_07_metric=False,
label_names=inria_bbox_label_names),
trigger=(1, 'iteration'))
log_interval = 1, '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=1))
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()
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()
args = parser.parse_args()
# TODO: small_lr_layers
chainer.cuda.set_max_workspace_size(512 * 1024 * 1024)
chainer.config.autotune = True
chainer.config.cudnn_fast_batch_normalization = True
model = SSD300(n_fg_class=20, pretrained_model='imagenet')
model.use_preset('evaluate')
net = SSDClassifier(model, alpha=1, k=3)
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)
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():
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()
batch_size = 8
model = SSD300(
n_fg_class=21,
pretrained_model='imagenet')
model.to_gpu(gpu_id)
train07 = VOCBboxDataset(data_dir='auto',year='2007', split='trainval',use_difficult=True,return_difficult=False)
train12 = VOCBboxDataset(data_dir='auto',year='2012', split='trainval',use_difficult=True,return_difficult=False)
train07_left = train07[:len(train07)//2]
train12_left = train12[:len(train12)//2]
train_left = ConcatenatedDataset(train07_left,train12_left)
train = TransformDataset(train_left,_Transform(model.coder,model.insize,model.mean))
train_iter = chainer.iterators.SerialIterator(train,batch_size,repeat=True,shuffle=True)
test = VOCBboxDataset(data_dir='auto',year='2007', split='test',use_difficult=True,return_difficult=True)
test_iter = chainer.iterators.SerialIterator(test,batch_size,repeat=False,shuffle=False)
class MultiboxTrainChain(chainer.Chain):
def __init__(self, model, alpha=1, k=3):
super(MultiboxTrainChain, self).__init__()
with self.init_scope():
self.model = model
self.alpha = alpha
self.k = k
