class EvalDetectionCLI(scfg.Config):
default = {
'true': scfg.Path(None, help='true coco dataset'),
'pred': scfg.Path(None, help='predicted coco dataset'),
'out_dpath': scfg.Path('./out', help='output directory')
}
pass
class SegmentationConfig(scfg.Config):
"""
Default configuration for setting up a training session
"""
default = {
'name': scfg.Value('untitled', help='A human readable tag that is "nice" for humans'),
'workdir': scfg.Path('~/work/sseg', help='Dump all results in your workdir'),
'workers': scfg.Value(0, help='number of parallel dataloading jobs'),
'xpu': scfg.Value('argv', help='See netharn.XPU for details. can be cpu/gpu/cuda0/0,1,2,3)'),
'augmenter': scfg.Value('simple', help='type of training dataset augmentation'),
'class_weights': scfg.Value('log-median-idf', help='how to weight inbalanced classes'),
# 'class_weights': scfg.Value(None, help='how to weight inbalanced classes'),
'datasets': scfg.Value('special:shapes256', help='Either a special key or a coco file'),
'train_dataset': scfg.Value(None),
'vali_dataset': scfg.Value(None),
'test_dataset': scfg.Value(None),
'arch': scfg.Value('psp', help='Network architecture code'),
'optim': scfg.Value('adam', help='Weight optimizer. Can be SGD, ADAM, ADAMW, etc..'),
'backend': scfg.Value('npy', help='fast lookup backnd. may be npy or cog'),
'input_dims': scfg.Value((224, 224), help='Window size to input to the network'),
'input_overlap': scfg.Value(0.25, help='amount of overlap when creating a sliding window dataset'),
'normalize_inputs': scfg.Value(True, help='if True, precompute training mean and std for data whitening'),
'batch_size': scfg.Value(4, help='number of items per batch'),
'bstep': scfg.Value(4, help='number of batches before a gradient descent step'),
'max_epoch': scfg.Value(140, help='Maximum number of epochs'),
'patience': scfg.Value(140, help='Maximum "bad" validation epochs before early stopping'),
'lr': scfg.Value(1e-4, help='Base learning rate'),
'decay': scfg.Value(1e-5, help='Base weight decay'),
'focus': scfg.Value(0.0, help='focus for focal loss'),
'schedule': scfg.Value('step90', help=('Special coercable netharn code. Eg: onecycle50, step50, gamma')),
'init': scfg.Value('kaiming_normal', help='How to initialized weights. (can be a path to a pretrained model)'),
'pretrained': scfg.Path(help=('alternative way to specify a path to a pretrained model')),
}
def normalize(self):
if self['pretrained'] in ['null', 'None']:
self['pretrained'] = None
if self['pretrained'] is not None:
self['init'] = 'pretrained'
class MyConfig(scfg.Config):
default = {
'option1': scfg.Value('bar', help='an option'),
'option2': scfg.Value((1, 2, 3), tuple, help='another option'),
'option3': None,
'option4': 'foo',
'discrete': scfg.Value(None, choices=['a', 'b', 'c']),
'apath': scfg.Path(help='a path'),
}
class ImageClfConfig(scfg.Config):
"""
Default configuration for setting up a training session
"""
default = {
'nice': scfg.Path('untitled', help='A human readable tag that is "nice" for humans'),
'workdir': scfg.Path('~/work/tiny-imagenet', help='Dump all results in your workdir'),
'workers': scfg.Value(0, help='number of parallel dataloading jobs'),
'xpu': scfg.Value('argv', help='See netharn.XPU for details. can be cpu/gpu/cuda0/0,1,2,3)'),
'augmenter': scfg.Value(True, help='type of training dataset augmentation'),
# 'datasets': scfg.Value('special:tiny_imgnet', help='Eventually you may be able to sepcify a coco file'),
'train_dataset': scfg.Value(None),
'vali_dataset': scfg.Value(None),
'test_dataset': scfg.Value(None),
'arch': scfg.Value('resnet50', help='Network architecture code'),
'optim': scfg.Value('adam', help='Weight optimizer. Can be SGD, ADAM, ADAMW, etc..'),
'input_dims': scfg.Value((224, 224), help='Window size to input to the network'),
'batch_size': scfg.Value(6, help='number of items per batch'),
'max_epoch': scfg.Value(100, help='Maximum number of epochs'),
'patience': scfg.Value(100, help='Maximum "bad" validation epochs before early stopping'),
'lr': scfg.Value(1e-3, help='Base learning rate'),
'decay': scfg.Value(1e-5, help='Base weight decay'),
'schedule': scfg.Value('onecycle71', help=('Special coercable netharn code. Eg: onecycle50, step50, gamma')),
'init': scfg.Value('kaiming_normal', help='How to initialized weights. (can be a path to a pretrained model)'),
'pretrained': scfg.Path(help=('alternative way to specify a path to a pretrained model')),
}
def normalize(self):
if self['pretrained'] in ['null', 'None']:
self['pretrained'] = None
if self['pretrained'] is not None:
self['init'] = 'pretrained'
class DetectFitConfig(scfg.Config):
default = {
# Personal Preference
'nice':
scfg.Value(
'untitled',
help=
('a human readable tag for your experiment (we also keep a '
'failsafe computer readable tag in case you update hyperparams, '
'but forget to update this flag)')),
# System Options
'workdir':
scfg.Path('~/work/detect',
help='path where this script can dump stuff'),
'workers':
scfg.Value(0, help='number of DataLoader processes'),
'xpu':
scfg.Value('argv', help='a CUDA device or a CPU'),
# Data (the hardest part of machine learning)
'datasets':
scfg.Value('special:shapes1024', help='special dataset key'),
'train_dataset':
scfg.Value(None, help='override train with a custom coco dataset'),
'vali_dataset':
scfg.Value(None, help='override vali with a custom coco dataset'),
'test_dataset':
scfg.Value(None, help='override test with a custom coco dataset'),
# Dataset options
'multiscale':
False,
'visible_thresh':
scfg.Value(
0.5,
help=
'percentage of a box that must be visible to be included in truth'
),
'input_dims':
scfg.Value((256, 256), help='size to '),
'normalize_inputs':
scfg.Value(
False,
help='if True, precompute training mean and std for data whitening'
),
'augment':
scfg.Value('simple',
help='key indicating augmentation strategy',
choices=['complex', 'simple', None]),
'ovthresh':
0.5,
# High level options
'arch':
scfg.Value('yolo2', help='network toplogy', choices=['yolo2']),
'optim':
scfg.Value('adam',
help='torch optimizer',
choices=['sgd', 'adam', 'adamw']),
'batch_size':
scfg.Value(
4, help='number of images that run through the network at a time'),
'bstep':
scfg.Value(8, help='num batches before stepping'),
'lr':
scfg.Value(1e-3, help='learning rate'), # 1e-4,
'decay':
scfg.Value(1e-5, help='weight decay'),
'schedule':
scfg.Value('step90', help='learning rate / momentum scheduler'),
'max_epoch':
scfg.Value(140, help='Maximum number of epochs'),
'patience':
scfg.Value(
140,
help='Maximum number of bad epochs on validation before stopping'),
# Initialization
'init':
scfg.Value('imagenet', help='initialization strategy'),
'pretrained':
scfg.Path(help='path to a netharn deploy file'),
# Loss Terms
'focus':
scfg.Value(0.0, help='focus for Focal Loss'),
}
def normalize(self):
if self['pretrained'] in ['null', 'None']:
self['pretrained'] = None
if self['datasets'] == 'special:voc':
self['train_dataset'] = ub.expandpath(
'~/data/VOC/voc-trainval.mscoco.json')
self['vali_dataset'] = ub.expandpath(
'~/data/VOC/voc-test-2007.mscoco.json')
key = self.get('pretrained', None) or self.get('init', None)
if key == 'imagenet':
self['pretrained'] = yolo2.initial_imagenet_weights()
elif key == 'lightnet':
self['pretrained'] = yolo2.demo_voc_weights()
if self['pretrained'] is not None:
self['init'] = 'pretrained'
class ClfConfig(scfg.Config):
"""
This is the default configuration for running the classification example.
Instances of this class behave like a dictionary. However, they can also be
specified on the command line, via kwargs, or by pointing to a YAML/json
file. See :module:``scriptconfig`` for details of how to use
:class:`scriptconfig.Config` objects.
"""
default = {
'name':
scfg.Value('clf_example',
help='A human readable tag that is "name" for humans'),
'workdir':
scfg.Path('~/work/netharn', help='Dump all results in your workdir'),
'workers':
scfg.Value(2, help='number of parallel dataloading jobs'),
'xpu':
scfg.Value(
'auto',
help=
'See netharn.XPU for details. can be auto/cpu/xpu/cuda0/0,1,2,3)'),
'datasets':
scfg.Value('special:shapes256',
help='Either a special key or a coco file'),
'train_dataset':
scfg.Value(None),
'vali_dataset':
scfg.Value(None),
'test_dataset':
scfg.Value(None),
'sampler_backend':
scfg.Value(None, help='ndsampler backend'),
'channels':
scfg.Value('rgb', help='special channel code. See ChannelSpec'),
'arch':
scfg.Value('resnet50', help='Network architecture code'),
'optim':
scfg.Value('adam',
help='Weight optimizer. Can be SGD, ADAM, ADAMW, etc..'),
'input_dims':
scfg.Value((224, 224), help='Window size to input to the network'),
'normalize_inputs':
scfg.Value(
True,
help=(
'if True, precompute training mean and std for data whitening'
)),
'balance':
scfg.Value(None, help='balance strategy. Can be category or None'),
'augmenter':
scfg.Value('simple', help='type of training dataset augmentation'),
'batch_size':
scfg.Value(3, help='number of items per batch'),
'num_batches':
scfg.Value(
'auto',
help=
'Number of batches per epoch (mainly for balanced batch sampling)'
),
'max_epoch':
scfg.Value(140, help='Maximum number of epochs'),
'patience':
scfg.Value(
140, help='Maximum "bad" validation epochs before early stopping'),
'lr':
scfg.Value(1e-4, help='Base learning rate'),
'decay':
scfg.Value(1e-5, help='Base weight decay'),
'schedule':
scfg.Value(
'step90-120',
help=
('Special coercible netharn code. Eg: onecycle50, step50, gamma, ReduceLROnPlateau-p10-c10'
)),
'init':
scfg.Value(
'noop',
help=
'How to initialized weights: e.g. noop, kaiming_normal, path-to-a-pretrained-model)'
),
'pretrained':
scfg.Path(
help=('alternative way to specify a path to a pretrained model')),
}
def normalize(self):
if self['pretrained'] in ['null', 'None']:
self['pretrained'] = None
if self['pretrained'] is not None:
self['init'] = 'pretrained'
class GrabVOC_Config(scfg.Config):
default = {
'dpath': scfg.Path(
ub.expandpath('~/data/VOC'), help='download location')
}
class StyleTransferConfig(scfg.Config):
default = {
'name':
scfg.Value('style_example',
help='A human readable tag that is "name" for humans'),
'workdir':
scfg.Path('~/work/netharn', help='Dump all results in your workdir'),
'workers':
scfg.Value(2, help='number of parallel dataloading jobs'),
'xpu':
scfg.Value(
'auto',
help=
'See netharn.XPU for details. can be auto/cpu/xpu/cuda0/0,1,2,3)'),
'datasets':
scfg.Value('special:shapes256',
help='Either a special key or a coco file'),
'train_dataset':
scfg.Value(None),
'vali_dataset':
scfg.Value(None),
'test_dataset':
scfg.Value(None),
'sampler_backend':
scfg.Value(None, help='ndsampler backend'),
'channels':
scfg.Value('rgb', help='special channel code. See ChannelSpec'),
# 'arch': scfg.Value('resnet50', help='Network architecture code'),
'optim':
scfg.Value('adam',
help='Weight optimizer. Can be SGD, ADAM, ADAMW, etc..'),
'input_dims':
scfg.Value((256, 256), help='Window size to input to the network'),
# TODO
'normalize_inputs':
scfg.Value(
True,
help=(
'if True, precompute training mean and std for data whitening'
)),
'balance':
scfg.Value(None, help='balance strategy. Can be category or None'),
# 'augmenter': scfg.Value('simple', help='type of training dataset augmentation'),
'batch_size':
scfg.Value(3, help='number of items per batch'),
'num_batches':
scfg.Value(
'auto',
help=
'Number of batches per epoch (mainly for balanced batch sampling)'
),
'max_epoch':
scfg.Value(140, help='Maximum number of epochs'),
'patience':
scfg.Value(
140, help='Maximum "bad" validation epochs before early stopping'),
'lr':
scfg.Value(1e-4, help='Base learning rate'),
'decay':
scfg.Value(1e-5, help='Base weight decay'),
'schedule':
scfg.Value(
'step90-120',
help=
('Special coercible netharn code. Eg: onecycle50, step50, gamma, ReduceLROnPlateau-p10-c10'
)),
'init':
scfg.Value(
'noop',
help=
'How to initialized weights: e.g. noop, kaiming_normal, path-to-a-pretrained-model)'
),
# 'pretrained': scfg.Path(help=('alternative way to specify a path to a pretrained model')),
}