def setup_harn(cmdline=False, **kw):
"""
Ignore:
kw = {}
cmdline = False
harn = setup_harn()
"""
config = StyleTransferConfig(default=kw)
config.load(cmdline=cmdline)
print('config = {}'.format(ub.repr2(config.asdict())))
nh.configure_hacks(config)
dataset_info = nh.api.DatasetInfo.coerce(config)
# input_stats = dataset_info['input_stats']
model = (TransformerNetwork, {})
hyper = nh.HyperParams(name=config['name'],
workdir=config['workdir'],
xpu=nh.XPU.coerce(config['xpu']),
datasets=dataset_info['torch_datasets'],
loaders=dataset_info['torch_loaders'],
model=model,
criterion=None,
initializer=None,
optimizer=nh.Optimizer.coerce(config),
dynamics=nh.Dynamics.coerce(config),
scheduler=nh.Scheduler.coerce(config),
monitor=(nh.Monitor, {
'minimize': ['loss'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': 0.0,
}),
other={
'name': config['name'],
'batch_size': config['batch_size'],
'balance': config['balance'],
},
extra={
'argv': sys.argv,
'config': ub.repr2(config.asdict()),
})
harn = StyleTransferHarn(hyper=hyper)
harn.preferences.update({
'num_keep': 3,
'keep_freq': 10,
'tensorboard_groups': ['loss'],
'eager_dump_tensorboard': True,
})
harn.intervals.update({})
harn.script_config = config
return harn
def setup_harn(cmdline=True, **kw):
"""
Ignore:
>>> from object_detection import * # NOQA
>>> cmdline = False
>>> kw = {
>>> 'train_dataset': '~/data/VOC/voc-trainval.mscoco.json',
>>> 'vali_dataset': '~/data/VOC/voc-test-2007.mscoco.json',
>>> }
>>> harn = setup_harn(**kw)
"""
import ndsampler
from ndsampler import coerce_data
# Seed other global rngs just in case something uses them under the hood
kwarray.seed_global(1129989262, offset=1797315558)
config = DetectFitConfig(default=kw, cmdline=cmdline)
nh.configure_hacks(config) # fix opencv bugs
ub.ensuredir(config['workdir'])
# Load ndsampler.CocoDataset objects from info in the config
subsets = coerce_data.coerce_datasets(config)
samplers = {}
for tag, subset in subsets.items():
print('subset = {!r}'.format(subset))
sampler = ndsampler.CocoSampler(subset, workdir=config['workdir'])
samplers[tag] = sampler
torch_datasets = {
tag: DetectDataset(
sampler,
input_dims=config['input_dims'],
augment=config['augment'] if (tag == 'train') else False,
)
for tag, sampler in samplers.items()
}
print('make loaders')
loaders_ = {
tag:
torch.utils.data.DataLoader(dset,
batch_size=config['batch_size'],
num_workers=config['workers'],
shuffle=(tag == 'train'),
collate_fn=nh.data.collate.padded_collate,
pin_memory=True)
for tag, dset in torch_datasets.items()
}
# for x in ub.ProgIter(loaders_['train']):
# pass
if config['normalize_inputs']:
# Get stats on the dataset (todo: turn off augmentation for this)
_dset = torch_datasets['train']
stats_idxs = kwarray.shuffle(np.arange(len(_dset)),
rng=0)[0:min(1000, len(_dset))]
stats_subset = torch.utils.data.Subset(_dset, stats_idxs)
cacher = ub.Cacher('dset_mean', cfgstr=_dset.input_id + 'v2')
input_stats = cacher.tryload()
if input_stats is None:
# Use parallel workers to load data faster
loader = torch.utils.data.DataLoader(
stats_subset,
collate_fn=nh.data.collate.padded_collate,
num_workers=config['workers'],
shuffle=True,
batch_size=config['batch_size'])
# Track moving average
running = nh.util.RunningStats()
for batch in ub.ProgIter(loader, desc='estimate mean/std'):
try:
running.update(batch['im'].numpy())
except ValueError: # final batch broadcast error
pass
input_stats = {
'std': running.simple(axis=None)['mean'].round(3),
'mean': running.simple(axis=None)['std'].round(3),
}
cacher.save(input_stats)
else:
input_stats = None
print('input_stats = {!r}'.format(input_stats))
initializer_ = nh.Initializer.coerce(config, leftover='kaiming_normal')
print('initializer_ = {!r}'.format(initializer_))
arch = config['arch']
if arch == 'yolo2':
if False:
dset = samplers['train'].dset
print('dset = {!r}'.format(dset))
# anchors = yolo2.find_anchors(dset)
anchors = np.array([(1.3221, 1.73145), (3.19275, 4.00944),
(5.05587, 8.09892), (9.47112, 4.84053),
(11.2364, 10.0071)])
classes = samplers['train'].classes
model_ = (yolo2.Yolo2, {
'classes': classes,
'anchors': anchors,
'conf_thresh': 0.001,
'nms_thresh': 0.5 if not ub.argflag('--eav') else 0.4
})
model = model_[0](**model_[1])
model._initkw = model_[1]
criterion_ = (
yolo2.YoloLoss,
{
'coder': model.coder,
'seen': 0,
'coord_scale': 1.0,
'noobject_scale': 1.0,
'object_scale': 5.0,
'class_scale': 1.0,
'thresh': 0.6, # iou_thresh
# 'seen_thresh': 12800,
})
else:
raise KeyError(arch)
scheduler_ = nh.Scheduler.coerce(config)
print('scheduler_ = {!r}'.format(scheduler_))
optimizer_ = nh.Optimizer.coerce(config)
print('optimizer_ = {!r}'.format(optimizer_))
dynamics_ = nh.Dynamics.coerce(config)
print('dynamics_ = {!r}'.format(dynamics_))
xpu = nh.XPU.coerce(config['xpu'])
print('xpu = {!r}'.format(xpu))
import sys
hyper = nh.HyperParams(
**{
'nice':
config['nice'],
'workdir':
config['workdir'],
'datasets':
torch_datasets,
'loaders':
loaders_,
'xpu':
xpu,
'model':
model,
'criterion':
criterion_,
'initializer':
initializer_,
'optimizer':
optimizer_,
'dynamics':
dynamics_,
# 'optimizer': (torch.optim.SGD, {
# 'lr': lr_step_points[0],
# 'momentum': 0.9,
# 'dampening': 0,
# # multiplying by batch size was one of those unpublished details
# 'weight_decay': decay * simulated_bsize,
# }),
'scheduler':
scheduler_,
'monitor': (
nh.Monitor,
{
'minimize': ['loss'],
# 'maximize': ['mAP'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': .6,
}),
'other': {
# Other params are not used internally, so you are free to set any
# extra params specific to your algorithm, and still have them
# logged in the hyperparam structure. For YOLO this is `ovthresh`.
'batch_size': config['batch_size'],
'nice': config['nice'],
'ovthresh': config['ovthresh'], # used in mAP computation
},
'extra': {
'config': ub.repr2(config.asdict()),
'argv': sys.argv,
}
})
print('hyper = {!r}'.format(hyper))
print('make harn')
harn = DetectHarn(hyper=hyper)
harn.preferences.update({
'num_keep': 2,
'keep_freq': 30,
'export_modules': ['netharn'], # TODO
'prog_backend': 'progiter', # alternative: 'tqdm'
'keyboard_debug': True,
})
harn.intervals.update({
'log_iter_train': 50,
})
harn.fit_config = config
print('harn = {!r}'.format(harn))
print('samplers = {!r}'.format(samplers))
return harn
def setup_harn(cmdline=True, **kw):
"""
Example:
>>> # xdoctest: +REQUIRES(--download)
>>> import sys, ubelt
>>> sys.path.append(ubelt.expandpath('~/code/netharn/examples'))
>>> from sseg_camvid import * # NOQA
>>> kw = {'workers': 0, 'xpu': 'cpu', 'batch_size': 2}
>>> cmdline = False
>>> # Just sets up the harness, does not do any heavy lifting
>>> harn = setup_harn(cmdline=cmdline, **kw)
>>> #
>>> harn.initialize()
>>> #
>>> batch = harn._demo_batch(tag='train')
>>> epoch_metrics = harn._demo_epoch(tag='vali', max_iter=4)
"""
import sys
import ndsampler
config = SegmentationConfig(default=kw)
config.load(cmdline=cmdline)
nh.configure_hacks(config) # fix opencv bugs
assert config['datasets'] == 'special:camvid'
coco_datasets = setup_coco_datasets()
workdir = ub.ensuredir(ub.expandpath(config['workdir']))
samplers = {
# tag: ndsampler.CocoSampler(dset, workdir=workdir, backend='cog')
tag: ndsampler.CocoSampler(dset, workdir=workdir, backend='npy')
for tag, dset in coco_datasets.items()
}
torch_datasets = {
tag: SegmentationDataset(
sampler,
config['input_dims'],
input_overlap=((tag == 'train') and config['input_overlap']),
augment=((tag == 'train') and config['augment']),
)
for tag, sampler in samplers.items()
}
torch_loaders = {
tag: torch_data.DataLoader(dset,
batch_size=config['batch_size'],
num_workers=config['workers'],
shuffle=(tag == 'train'),
drop_last=True, pin_memory=True)
for tag, dset in torch_datasets.items()
}
if config['class_weights']:
mode = config['class_weights']
dset = torch_datasets['train']
class_weights = _precompute_class_weights(dset, mode=mode)
class_weights = torch.FloatTensor(class_weights)
class_weights[dset.classes.index('background')] = 0
else:
class_weights = None
initializer_ = nh.Initializer.coerce(config)
if config['arch'] == 'unet':
# Note: UNet can get through 256x256 images at a rate of ~17Hz with
# batch_size=8. This is pretty slow and can likely be improved by fixing
# some of the weird padding / mirror stuff I have to do in unet to get
# output_dims = input_dims.
from netharn.models.unet import UNet
model_ = (UNet, {
'classes': torch_datasets['train'].classes,
'in_channels': 3,
})
elif config['arch'] == 'segnet':
from netharn.models.segnet import Segnet
model_ = (Segnet, {
'classes': torch_datasets['train'].classes,
'in_channels': 3,
})
elif config['arch'] == 'psp':
from netharn.models.psp import PSPNet_Resnet50_8s
model_ = (PSPNet_Resnet50_8s, {
'classes': torch_datasets['train'].classes,
'in_channels': 3,
})
elif config['arch'] == 'deeplab':
from netharn.models.deeplab import DeepLab_ASPP
model_ = (DeepLab_ASPP, {
'classes': torch_datasets['train'].classes,
'in_channels': 3,
})
else:
raise KeyError(config['arch'])
if config['init'] == 'cls':
initializer_ = model_[0]._initializer_cls()
# Create hyperparameters
hyper = nh.HyperParams(
nice=config['nice'],
workdir=config['workdir'],
xpu=nh.XPU.coerce(config['xpu']),
datasets=torch_datasets,
loaders=torch_loaders,
model=model_,
initializer=initializer_,
scheduler=nh.Scheduler.coerce(config),
optimizer=nh.Optimizer.coerce(config),
dynamics=nh.Dynamics.coerce(config),
criterion=(nh.criterions.FocalLoss, {
'focus': config['focus'],
'weight': class_weights,
# 'reduction': 'none',
}),
monitor=(nh.Monitor, {
'minimize': ['loss'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': .6,
}),
other={
'batch_size': config['batch_size'],
},
extra={
'argv': sys.argv,
'config': ub.repr2(config.asdict()),
}
)
# Create harness
harn = SegmentationHarn(hyper=hyper)
harn.classes = torch_datasets['train'].classes
harn.preferences.update({
'num_keep': 5,
'keyboard_debug': True,
# 'export_modules': ['netharn'],
})
harn.intervals.update({
'vali': 1,
'test': 10,
})
harn.script_config = config
return harn
def setup_harn(cmdline=True, **kw):
"""
CommandLine:
xdoctest -m netharn.examples.segmentation setup_harn
Example:
>>> # xdoctest: +REQUIRES(--slow)
>>> kw = {'workers': 0, 'xpu': 'cpu', 'batch_size': 2}
>>> cmdline = False
>>> # Just sets up the harness, does not do any heavy lifting
>>> harn = setup_harn(cmdline=cmdline, **kw)
>>> #
>>> harn.initialize()
>>> #
>>> batch = harn._demo_batch(tag='train')
>>> epoch_metrics = harn._demo_epoch(tag='vali', max_iter=2)
"""
import sys
import ndsampler
import kwarray
# kwarray.seed_global(2108744082)
config = SegmentationConfig(default=kw)
config.load(cmdline=cmdline)
nh.configure_hacks(config) # fix opencv bugs
coco_datasets = nh.api.Datasets.coerce(config)
print('coco_datasets = {}'.format(ub.repr2(coco_datasets)))
for tag, dset in coco_datasets.items():
dset._build_hashid(hash_pixels=False)
workdir = ub.ensuredir(ub.expandpath(config['workdir']))
samplers = {
tag: ndsampler.CocoSampler(dset,
workdir=workdir,
backend=config['backend'])
for tag, dset in coco_datasets.items()
}
for tag, sampler in ub.ProgIter(list(samplers.items()),
desc='prepare frames'):
try:
sampler.frames.prepare(workers=config['workers'])
except AttributeError:
pass
torch_datasets = {
tag: SegmentationDataset(
sampler,
config['input_dims'],
input_overlap=((tag == 'train') and config['input_overlap']),
augmenter=((tag == 'train') and config['augmenter']),
)
for tag, sampler in samplers.items()
}
torch_loaders = {
tag: torch_data.DataLoader(dset,
batch_size=config['batch_size'],
num_workers=config['workers'],
shuffle=(tag == 'train'),
drop_last=True,
pin_memory=True)
for tag, dset in torch_datasets.items()
}
if config['class_weights']:
mode = config['class_weights']
dset = torch_datasets['train']
class_weights = _precompute_class_weights(dset,
mode=mode,
workers=config['workers'])
class_weights = torch.FloatTensor(class_weights)
class_weights[dset.classes.index('background')] = 0
else:
class_weights = None
if config['normalize_inputs']:
stats_dset = torch_datasets['train']
stats_idxs = kwarray.shuffle(np.arange(len(stats_dset)),
rng=0)[0:min(1000, len(stats_dset))]
stats_subset = torch.utils.data.Subset(stats_dset, stats_idxs)
cacher = ub.Cacher('dset_mean', cfgstr=stats_dset.input_id + 'v3')
input_stats = cacher.tryload()
if input_stats is None:
loader = torch.utils.data.DataLoader(
stats_subset,
num_workers=config['workers'],
shuffle=True,
batch_size=config['batch_size'])
running = nh.util.RunningStats()
for batch in ub.ProgIter(loader, desc='estimate mean/std'):
try:
running.update(batch['im'].numpy())
except ValueError: # final batch broadcast error
pass
input_stats = {
'std': running.simple(axis=None)['mean'].round(3),
'mean': running.simple(axis=None)['std'].round(3),
}
cacher.save(input_stats)
else:
input_stats = {}
print('input_stats = {!r}'.format(input_stats))
# TODO: infer numbr of channels
model_ = (SegmentationModel, {
'arch': config['arch'],
'input_stats': input_stats,
'classes': torch_datasets['train'].classes.__json__(),
'in_channels': 3,
})
initializer_ = nh.Initializer.coerce(config)
# if config['init'] == 'cls':
# initializer_ = model_[0]._initializer_cls()
# Create hyperparameters
hyper = nh.HyperParams(
nice=config['nice'],
workdir=config['workdir'],
xpu=nh.XPU.coerce(config['xpu']),
datasets=torch_datasets,
loaders=torch_loaders,
model=model_,
initializer=initializer_,
scheduler=nh.Scheduler.coerce(config),
optimizer=nh.Optimizer.coerce(config),
dynamics=nh.Dynamics.coerce(config),
criterion=(
nh.criterions.FocalLoss,
{
'focus': config['focus'],
'weight': class_weights,
# 'reduction': 'none',
}),
monitor=(nh.Monitor, {
'minimize': ['loss'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': .6,
}),
other={
'batch_size': config['batch_size'],
},
extra={
'argv': sys.argv,
'config': ub.repr2(config.asdict()),
})
# Create harness
harn = SegmentationHarn(hyper=hyper)
harn.classes = torch_datasets['train'].classes
harn.preferences.update({
'num_keep': 2,
'keyboard_debug': True,
# 'export_modules': ['netharn'],
})
harn.intervals.update({
'vali': 1,
'test': 10,
})
harn.script_config = config
return harn
def setup_harn(**kwargs):
"""
Args:
nice (str): Custom tag for this run
workdir (PathLike): path to dump all the intermedate results
batch_size (int):
Base batch size. Number of examples in GPU at any time.
p (int): num individuals per batch
k (int): num annots-per-individual per batch
bstep (int): Multiply by batch_size to simulate a larger batches.
lr (float|str): Base learning rate
decay (float): Weight decay (L2 regularization)
workers (int): Number of parallel data loader workers
xpu (str): Device to train on. Can be either `'cpu'`, `'gpu'`, a number
indicating a GPU (e.g. `0`), or a list of numbers (e.g. `[0,1,2]`)
indicating multiple GPUs
norm_desc (bool): if True normalizes the descriptors
pretrained (PathLike): path to a compatible pretrained model
margin (float): margin for loss criterion
soft (bool): use soft margin
"""
import ast
def trycast(x, type):
try:
return type(x)
except Exception:
return x
config = {}
config['init'] = kwargs.get('init', 'kaiming_normal')
config['pretrained'] = config.get('pretrained',
ub.argval('--pretrained', default=None))
config['margin'] = kwargs.get('margin', 3.0)
config['soft'] = kwargs.get('soft', False)
config['xpu'] = kwargs.get('xpu', 'argv')
config['nice'] = kwargs.get('nice', 'untitled')
config['workdir'] = kwargs.get('workdir', None)
config['workers'] = int(kwargs.get('workers', 1))
config['bstep'] = int(kwargs.get('bstep', 1))
config['optim'] = kwargs.get('optim', 'sgd')
config['scheduler'] = kwargs.get('scheduler', 'onecycle70')
config['lr'] = trycast(kwargs.get('lr', 0.0001), float)
config['decay'] = float(kwargs.get('decay', 1e-5))
config['max_epoch'] = int(kwargs.get('max_epoch', 100))
config['num_batches'] = trycast(kwargs.get('num_batches', 1000), int)
config['batch_size'] = int(kwargs.get('batch_size', 128))
config['p'] = float(kwargs.get('p', 10))
config['k'] = float(kwargs.get('k', 25))
config['arch'] = kwargs.get('arch', 'resnet')
config['hidden'] = trycast(kwargs.get('hidden', [128]), ast.literal_eval)
config['desc_size'] = kwargs.get('desc_size', 256)
config['norm_desc'] = kwargs.get('norm_desc', False)
config['dim'] = 28
xpu = nh.XPU.coerce(config['xpu'])
nh.configure_hacks(config)
datasets, workdir = setup_datasets()
loaders = {
tag: torch.utils.data.DataLoader(
dset,
batch_sampler=nh.data.batch_samplers.MatchingSamplerPK(
dset.pccs,
shuffle=(tag == 'train'),
batch_size=config['batch_size'],
num_batches=config['num_batches'],
k=config['k'],
p=config['p'],
),
num_workers=config['workers'],
)
for tag, dset in datasets.items()
}
if config['arch'] == 'simple':
model_ = (MNISTEmbeddingNet, {
'input_shape': (1, 1, config['dim'], config['dim']),
'desc_size': config['desc_size'],
})
elif config['arch'] == 'resnet':
model_ = (nh.models.DescriptorNetwork, {
'input_shape': (1, 1, config['dim'], config['dim']),
'norm_desc': config['norm_desc'],
'hidden_channels': config['hidden'],
'desc_size': config['desc_size'],
})
else:
raise KeyError(config['arch'])
if config['scheduler'] == 'steplr':
from torch.optim import lr_scheduler
scheduler_ = (lr_scheduler.StepLR,
dict(step_size=8, gamma=0.1, last_epoch=-1))
else:
scheduler_ = nh.Scheduler.coerce(config, scheduler='onecycle70')
# Here is the FitHarn magic.
# They nh.HyperParams object keeps track of and helps log all declarative
# info related to training a model.
hyper = nh.hyperparams.HyperParams(
nice=config['nice'],
xpu=xpu,
workdir=workdir,
datasets=datasets,
loaders=loaders,
model=model_,
initializer=nh.Initializer.coerce(config),
optimizer=nh.Optimizer.coerce(config),
scheduler=scheduler_,
criterion=(nh.criterions.TripletLoss, {
'margin': config['margin'],
'soft': config['soft'],
}),
monitor=nh.Monitor.coerce(
config,
minimize=['loss', 'pos_dist', 'brier'],
maximize=['accuracy', 'neg_dist', 'mcc'],
patience=100,
max_epoch=config['max_epoch'],
smoothing=0.4,
),
other={
'batch_size': config['batch_size'],
'num_batches': config['num_batches'],
})
harn = MNIST_MatchingHarness(hyper=hyper)
harn.preferences
return harn
def setup_harn(cmdline=True, **kw):
"""
This creates the "The Classification Harness" (i.e. core ClfHarn object).
This is where we programmatically connect our program arguments with the
netharn HyperParameter standards. We are using :module:`scriptconfig` to
capture these, but you could use click / argparse / etc.
This function has the responsibility of creating our torch datasets,
lazy computing input statistics, specifying our model architecture,
schedule, initialization, optimizer, dynamics, XPU etc. These can usually
be coerced using netharn API helpers and a "standardized" config dict. See
the function code for details.
Args:
cmdline (bool, default=True):
if True, behavior will be modified based on ``sys.argv``.
Note this will activate the scriptconfig ``--help``, ``--dump`` and
``--config`` interactions.
Kwargs:
**kw: the overrides the default config for :class:`ClfConfig`.
Note, command line flags have precedence if cmdline=True.
Returns:
ClfHarn: a fully-defined, but uninitialized custom :class:`FitHarn`
object.
Example:
>>> # xdoctest: +SKIP
>>> kw = {'datasets': 'special:shapes256'}
>>> cmdline = False
>>> harn = setup_harn(cmdline, **kw)
>>> harn.initialize()
"""
import ndsampler
config = ClfConfig(default=kw)
config.load(cmdline=cmdline)
print('config = {}'.format(ub.repr2(config.asdict())))
nh.configure_hacks(config)
coco_datasets = nh.api.Datasets.coerce(config)
print('coco_datasets = {}'.format(ub.repr2(coco_datasets, nl=1)))
for tag, dset in coco_datasets.items():
dset._build_hashid(hash_pixels=False)
workdir = ub.ensuredir(ub.expandpath(config['workdir']))
samplers = {
tag: ndsampler.CocoSampler(dset,
workdir=workdir,
backend=config['sampler_backend'])
for tag, dset in coco_datasets.items()
}
for tag, sampler in ub.ProgIter(list(samplers.items()),
desc='prepare frames'):
sampler.frames.prepare(workers=config['workers'])
torch_datasets = {
'train':
ClfDataset(
samplers['train'],
input_dims=config['input_dims'],
augmenter=config['augmenter'],
),
'vali':
ClfDataset(samplers['vali'],
input_dims=config['input_dims'],
augmenter=False),
}
if config['normalize_inputs']:
# Get stats on the dataset (todo: turn off augmentation for this)
_dset = torch_datasets['train']
stats_idxs = kwarray.shuffle(np.arange(len(_dset)),
rng=0)[0:min(1000, len(_dset))]
stats_subset = torch.utils.data.Subset(_dset, stats_idxs)
cacher = ub.Cacher('dset_mean', cfgstr=_dset.input_id + 'v3')
input_stats = cacher.tryload()
channels = ChannelSpec.coerce(config['channels'])
if input_stats is None:
# Use parallel workers to load data faster
from netharn.data.data_containers import container_collate
from functools import partial
collate_fn = partial(container_collate, num_devices=1)
loader = torch.utils.data.DataLoader(
stats_subset,
collate_fn=collate_fn,
num_workers=config['workers'],
shuffle=True,
batch_size=config['batch_size'])
# Track moving average of each fused channel stream
channel_stats = {
key: nh.util.RunningStats()
for key in channels.keys()
}
assert len(channel_stats) == 1, (
'only support one fused stream for now')
for batch in ub.ProgIter(loader, desc='estimate mean/std'):
for key, val in batch['inputs'].items():
try:
for part in val.numpy():
channel_stats[key].update(part)
except ValueError: # final batch broadcast error
pass
perchan_input_stats = {}
for key, running in channel_stats.items():
running = ub.peek(channel_stats.values())
perchan_stats = running.simple(axis=(1, 2))
perchan_input_stats[key] = {
'std': perchan_stats['mean'].round(3),
'mean': perchan_stats['std'].round(3),
}
input_stats = ub.peek(perchan_input_stats.values())
cacher.save(input_stats)
else:
input_stats = {}
torch_loaders = {
tag: dset.make_loader(
batch_size=config['batch_size'],
num_batches=config['num_batches'],
num_workers=config['workers'],
shuffle=(tag == 'train'),
balance=(config['balance'] if tag == 'train' else None),
pin_memory=True)
for tag, dset in torch_datasets.items()
}
initializer_ = None
classes = torch_datasets['train'].classes
modelkw = {
'arch': config['arch'],
'input_stats': input_stats,
'classes': classes.__json__(),
'channels': channels,
}
model = ClfModel(**modelkw)
model._initkw = modelkw
if initializer_ is None:
initializer_ = nh.Initializer.coerce(config)
hyper = nh.HyperParams(name=config['name'],
workdir=config['workdir'],
xpu=nh.XPU.coerce(config['xpu']),
datasets=torch_datasets,
loaders=torch_loaders,
model=model,
criterion=None,
optimizer=nh.Optimizer.coerce(config),
dynamics=nh.Dynamics.coerce(config),
scheduler=nh.Scheduler.coerce(config),
initializer=initializer_,
monitor=(nh.Monitor, {
'minimize': ['loss'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': 0.0,
}),
other={
'name': config['name'],
'batch_size': config['batch_size'],
'balance': config['balance'],
},
extra={
'argv': sys.argv,
'config': ub.repr2(config.asdict()),
})
harn = ClfHarn(hyper=hyper)
harn.preferences.update({
'num_keep': 3,
'keep_freq': 10,
'tensorboard_groups': ['loss'],
'eager_dump_tensorboard': True,
})
harn.intervals.update({})
harn.script_config = config
return harn
def setup_harn(cmdline=True, **kwargs):
"""
cmdline, kwargs = False, {}
"""
import sys
import ndsampler
config = ImageClfConfig(default=kwargs)
config.load(cmdline=cmdline)
nh.configure_hacks(config) # fix opencv bugs
cacher = ub.Cacher('tiny-imagenet', cfgstr='v4', verbose=3)
data = cacher.tryload()
if data is None:
data = grab_tiny_imagenet_as_coco()
cacher.save(data)
coco_datasets = data # setup_coco_datasets()
dset = coco_datasets['train']
print('train dset = {!r}'.format(dset))
workdir = ub.ensuredir(ub.expandpath(config['workdir']))
samplers = {
# tag: ndsampler.CocoSampler(dset, workdir=workdir, backend='cog')
tag: ndsampler.CocoSampler(dset, workdir=workdir, backend='npy')
for tag, dset in coco_datasets.items()
}
torch_datasets = {
tag: ImagClfDataset(
sampler, config['input_dims'],
augmenter=((tag == 'train') and config['augmenter']),
)
for tag, sampler in samplers.items()
}
torch_loaders = {
tag: torch_data.DataLoader(dset,
batch_size=config['batch_size'],
num_workers=config['workers'],
shuffle=(tag == 'train'),
pin_memory=True)
for tag, dset in torch_datasets.items()
}
import torchvision
# TODO: netharn should allow for this
model_ = torchvision.models.resnet50(pretrained=False)
# model_ = (, {
# 'classes': torch_datasets['train'].classes,
# 'in_channels': 3,
# })
initializer_ = nh.Initializer.coerce(config)
hyper = nh.HyperParams(
nice=config['nice'],
workdir=config['workdir'],
xpu=nh.XPU.coerce(config['xpu']),
datasets=torch_datasets,
loaders=torch_loaders,
model=model_,
initializer=initializer_,
scheduler=nh.Scheduler.coerce(config),
optimizer=nh.Optimizer.coerce(config),
dynamics=nh.Dynamics.coerce(config),
criterion=(nh.criterions.FocalLoss, {
'focus': 0.0,
}),
monitor=(nh.Monitor, {
'minimize': ['loss'],
'patience': config['patience'],
'max_epoch': config['max_epoch'],
'smoothing': .6,
}),
other={
'batch_size': config['batch_size'],
},
extra={
'argv': sys.argv,
'config': ub.repr2(config.asdict()),
}
)
# Create harness
harn = ImageClfHarn(hyper=hyper)
harn.classes = torch_datasets['train'].classes
harn.preferences.update({
'num_keep': 5,
'keyboard_debug': True,
# 'export_modules': ['netharn'],
})
harn.intervals.update({
'vali': 1,
'test': 10,
})
harn.script_config = config
return harn
def setup_yolo_harness(bsize=16, workers=0):
"""
CommandLine:
python -m netharn.examples.yolo_voc setup_yolo_harness
Example:
>>> # DISABLE_DOCTSET
>>> harn = setup_yolo_harness()
>>> harn.initialize()
"""
xpu = nh.XPU.coerce('argv')
nice = ub.argval('--nice', default='Yolo2Baseline')
batch_size = int(ub.argval('--batch_size', default=bsize))
bstep = int(ub.argval('--bstep', 4))
workers = int(ub.argval('--workers', default=workers))
decay = float(ub.argval('--decay', default=0.0005))
lr = float(ub.argval('--lr', default=0.001))
ovthresh = 0.5
simulated_bsize = bstep * batch_size
nh.configure_hacks(workers=workers)
# We will divide the learning rate by the simulated batch size
datasets = {
'train': YoloVOCDataset(years=[2007, 2012], split='trainval'),
# 'test': YoloVOCDataset(years=[2007], split='test'),
}
loaders = {
key: dset.make_loader(batch_size=batch_size,
num_workers=workers,
shuffle=(key == 'train'),
pin_memory=True,
resize_rate=10 * bstep,
drop_last=True)
for key, dset in datasets.items()
}
anchors = np.array([(1.3221, 1.73145), (3.19275, 4.00944),
(5.05587, 8.09892), (9.47112, 4.84053),
(11.2364, 10.0071)])
if not ub.argflag('--eav'):
lr_step_points = {
# 0: lr * 0.1 / simulated_bsize, # burnin
# 4: lr * 1.0 / simulated_bsize,
0: lr * 1.0 / simulated_bsize,
154: lr * 1.0 / simulated_bsize,
155: lr * 0.1 / simulated_bsize,
232: lr * 0.1 / simulated_bsize,
233: lr * 0.01 / simulated_bsize,
}
max_epoch = 311
scheduler_ = (
nh.schedulers.core.YOLOScheduler,
{
'points': lr_step_points,
# 'interpolate': False,
'interpolate': True,
'burn_in': 0.96899225 if ub.argflag('--eav') else
3.86683584, # number of epochs to burn_in for. approx 1000 batches?
'dset_size': len(datasets['train']), # when drop_last=False
# 'dset_size': (len(datasets['train']) // simulated_bsize) * simulated_bsize, # make a multiple of batch_size because drop_last=True
'batch_size': batch_size,
})
from netharn.models.yolo2 import light_region_loss
criterion_ = (
light_region_loss.RegionLoss,
{
'num_classes': datasets['train'].num_classes,
'anchors': anchors,
'object_scale': 5.0,
'noobject_scale': 1.0,
# eav version originally had a random *2 in cls loss,
# we removed, that but we can replicate it here.
'class_scale': 1.0 if not ub.argflag('--eav') else 2.0,
'coord_scale': 1.0,
'thresh': 0.6, # iou_thresh
'seen_thresh': 12800,
# 'small_boxes': not ub.argflag('--eav'),
'small_boxes': True,
'mse_factor': 0.5 if not ub.argflag('--eav') else 1.0,
})
else:
lr_step_points = {
# dividing by batch size was one of those unpublished details
0: lr * 0.1 / simulated_bsize,
1: lr * 1.0 / simulated_bsize,
96: lr * 1.0 / simulated_bsize,
97: lr * 0.1 / simulated_bsize,
135: lr * 0.1 / simulated_bsize,
136: lr * 0.01 / simulated_bsize,
}
max_epoch = 176
scheduler_ = (nh.schedulers.ListedLR, {
'points': lr_step_points,
'interpolate': False,
})
from netharn.models.yolo2 import region_loss2
criterion_ = (
region_loss2.RegionLoss,
{
'num_classes': datasets['train'].num_classes,
'anchors': anchors,
'reduction': 32,
'seen': 0,
'coord_scale': 1.0,
'noobject_scale': 1.0,
'object_scale': 5.0,
'class_scale': 1.0,
'thresh': 0.6, # iou_thresh
# 'seen_thresh': 12800,
})
weights = ub.argval('--weights', default=None)
if weights is None or weights == 'imagenet':
weights = light_yolo.initial_imagenet_weights()
elif weights == 'lightnet':
weights = light_yolo.demo_voc_weights()
else:
print('weights = {!r}'.format(weights))
hyper = nh.HyperParams(
**{
'nice':
nice,
'workdir':
ub.expandpath('~/work/voc_yolo2'),
'datasets':
datasets,
'loaders':
loaders,
'xpu':
xpu,
'model': (
light_yolo.Yolo,
{
'num_classes': datasets['train'].num_classes,
'anchors': anchors,
'conf_thresh': 0.001,
# 'conf_thresh': 0.1, # make training a bit faster
'nms_thresh': 0.5 if not ub.argflag('--eav') else 0.4
}),
'criterion':
criterion_,
'initializer': (nh.initializers.Pretrained, {
'fpath': weights,
}),
'optimizer': (
torch.optim.SGD,
{
'lr': lr_step_points[0],
'momentum': 0.9,
'dampening': 0,
# multiplying by batch size was one of those unpublished details
'weight_decay': decay * simulated_bsize,
}),
'scheduler':
scheduler_,
'monitor': (nh.Monitor, {
'minimize': ['loss'],
'maximize': ['mAP'],
'patience': max_epoch,
'max_epoch': max_epoch,
}),
# 'augment': datasets['train'].augmenter,
'dynamics': {
# Controls how many batches to process before taking a step in the
# gradient direction. Effectively simulates a batch_size that is
# `bstep` times bigger.
'batch_step': bstep,
},
'other': {
# Other params are not used internally, so you are free to set any
# extra params specific to your algorithm, and still have them
# logged in the hyperparam structure. For YOLO this is `ovthresh`.
'batch_size': batch_size,
'nice': nice,
'ovthresh': ovthresh, # used in mAP computation
'input_range': 'norm01',
},
})
print('max_epoch = {!r}'.format(max_epoch))
harn = YoloHarn(hyper=hyper)
harn.preferences['prog_backend'] = 'progiter'
harn.intervals['log_iter_train'] = None
harn.intervals['log_iter_test'] = None
harn.intervals['log_iter_vali'] = None
harn.preferences[
'large_loss'] = 1000 # tell netharn when to check for divergence
return harn
def setup_harn(**kwargs):
"""
CommandLine:
python ~/code/netharn/netharn/examples/ggr_matching.py setup_harn
Args:
dbname (str): Name of IBEIS database to use
nice (str): Custom tag for this run
workdir (PathLike): path to dump all the intermedate results
dim (int): Width and height of the network input
batch_size (int): Base batch size. Number of examples in GPU at any time.
bstep (int): Multiply by batch_size to simulate a larger batches.
lr (float): Base learning rate
decay (float): Weight decay (L2 regularization)
workers (int): Number of parallel data loader workers
xpu (str): Device to train on. Can be either `'cpu'`, `'gpu'`, a number
indicating a GPU (e.g. `0`), or a list of numbers (e.g. `[0,1,2]`)
indicating multiple GPUs
triple (bool): if True uses triplet loss, otherwise contrastive loss
norm_desc (bool): if True normalizes the descriptors
pretrained (PathLike): path to a compatible pretrained model
margin (float): margin for loss criterion
soft (bool): use soft margin
Example:
>>> harn = setup_harn(dbname='PZ_MTEST')
>>> harn.initialize()
"""
config = parse_config(**kwargs)
nh.configure_hacks(config)
datasets, workdir = setup_datasets(config)
loaders = {
tag: dset.make_loader(
shuffle=(tag == 'train'),
batch_size=config['batch_size'],
num_batches=(config['num_batches']
if tag == 'train' else config['num_batches'] // 10),
k=config['k'],
p=config['p'],
num_workers=config['workers'],
)
for tag, dset in datasets.items()
}
if config['scheduler'] == 'steplr':
from torch.optim import lr_scheduler
scheduler_ = (lr_scheduler.StepLR,
dict(step_size=8, gamma=0.1, last_epoch=-1))
else:
scheduler_ = nh.Scheduler.coerce(config, scheduler='onecycle70')
hyper = nh.HyperParams(
**{
'nice':
config['nice'],
'workdir':
config['workdir'],
'datasets':
datasets,
'loaders':
loaders,
'xpu':
nh.XPU.coerce(config['xpu']),
'model': (
nh.models.DescriptorNetwork,
{
'input_shape': (1, 3, config['dim'], config['dim']),
'norm_desc': config['norm_desc'],
# 'hidden_channels': [512, 256]
'hidden_channels': [256],
'desc_size': 128,
}),
'initializer':
nh.Initializer.coerce(config),
'optimizer':
nh.Optimizer.coerce(config),
'scheduler':
scheduler_,
'criterion': (nh.criterions.TripletLoss, {
'margin': config['margin'],
'soft': config['soft'],
}),
'monitor':
nh.Monitor.coerce(
config,
minimize=['loss', 'pos_dist', 'brier'],
maximize=['accuracy', 'neg_dist', 'mcc'],
patience=100,
max_epoch=100,
),
'dynamics':
nh.Dynamics.coerce(config),
'other': {
'n_classes': 2,
},
})
harn = MatchingHarness(hyper=hyper)
harn.preferences['prog_backend'] = 'progiter'
harn.intervals['log_iter_train'] = 1
harn.intervals['log_iter_test'] = None
harn.intervals['log_iter_vali'] = None
return harn