def load_dataset(dataset, hyperparams):
datasets = {
"canevet-icml2016-jittered": partial(CanevetICML2016, smooth=10),
"canevet-icml2016": CanevetICML2016,
"canevet-icml2016-smooth": partial(CanevetICML2016, smooth=15000),
"cifar-sanity-check": CIFARSanityCheck,
"mnist": MNIST,
"cifar10": CIFAR10,
"cifar100": CIFAR100,
"cifar10-augmented": compose(
partial(OntheflyAugmentedImages, ___, dict(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False
)),
CIFAR10
),
"cifar10-whitened-augmented": compose(
partial(OntheflyAugmentedImages, ___, dict(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=True,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False
)),
CIFAR10
),
"cifar100-augmented": compose(
partial(OntheflyAugmentedImages, ___, dict(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False
)),
CIFAR100
),
"ptb": partial(PennTreeBank, 20)
}
return datasets[dataset]()
def get_samplers_dictionary(model, hyperparams={}, reweighting=None):
# Add some base samplers
samplers = {
"uniform":
partial(UniformSampler, ___, reweighting),
"model":
partial(ModelSampler,
___,
reweighting,
model,
large_batch=hyperparams.get("presample", 1024)),
"lstm":
partial(LSTMSampler,
___,
reweighting,
log=hyperparams.get("lstm_log", 0) != 0,
warmup=hyperparams.get("lstm_warmup", 100),
presample=hyperparams.get("presample", 2048)),
"pcg":
partial(PerClassGaussian,
___,
reweighting,
alpha=hyperparams.get("alpha", 0.9),
presample=hyperparams.get("presample", 2048)),
"obs":
partial(OnlineBatchSelectionSampler,
___,
reweighting,
model,
steps_per_epoch=hyperparams.get("steps_per_epoch", 300),
recompute=hyperparams.get("obs_recompute", 2),
s_e=(hyperparams.get("obs_se0",
1), hyperparams.get("obs_seend", 1)),
n_epochs=hyperparams.get("obs_epochs", 1))
}
samplers["lstm-comparison"] = lambda x: LSTMComparisonSampler(
x,
samplers["lstm"](x),
samplers["model"](x),
subset=hyperparams.get("lstm_comparison", 1024))
# Add some decorated samplers
samplers_for_decoration = ["model", "lstm", "pcg"]
for sampler in samplers_for_decoration:
samplers["smooth-" + sampler] = compose(
partial(AdditiveSmoothingSampler,
c=hyperparams.get("smooth", 1.0)), samplers[sampler])
samplers["adaptive-smooth-" + sampler] = compose(
partial(AdaptiveAdditiveSmoothingSampler,
percentage=hyperparams.get("smooth", 0.5)),
samplers[sampler])
samplers["power-smooth-" + sampler] = compose(
partial(PowerSmoothingSampler,
power=hyperparams.get("smooth", 0.5)), samplers[sampler])
return samplers
def test_datasets(self):
datasets = [(CIFAR10, 50000, 10000, (32, 32, 3), 10),
(CIFARSanityCheck, 40000, 2000, (32, 32, 3), 2),
(MNIST, 60000, 10000, (28, 28, 1), 10),
(partial(CanevetICML2016,
N=256), int(256**2 - 256**2 * 0.33) + 1,
int(256**2 * 0.33), (2, ), 2),
(compose(
partial(OntheflyAugmentedImages,
augmentation_params=dict(
featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)),
CIFAR10), 500000, 10000, (32, 32, 3), 10),
(partial(PennTreeBank, 20), 887521, 70390, (20, ), 10000)]
for args in datasets:
self._test_dset(*args)
def load_dataset(dataset):
datasets = {
"mnist":
MNIST,
"cifar10":
CIFAR10,
"cifar100":
CIFAR100,
"cifar10-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR10),
"cifar100-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR100),
"ptb":
partial(PennTreeBank, 20)
}
return datasets[dataset]()
def get_samplers_dictionary(model, hyperparams={}, reweighting=None):
# Add some base samplers
samplers = {
"uniform":
partial(UniformSampler, ___, reweighting),
"model":
partial(ModelSampler,
___,
reweighting,
model,
large_batch=hyperparams.get("presample", 1024)),
"lstm":
partial(LSTMSampler,
___,
reweighting,
log=hyperparams.get("lstm_log", 0) != 0,
presample=hyperparams.get("presample", 2048)),
"pcg":
partial(PerClassGaussian,
___,
reweighting,
alpha=hyperparams.get("alpha", 0.9),
presample=hyperparams.get("presample", 2048)),
"obs":
partial(OnlineBatchSelectionSampler,
___,
reweighting,
model,
steps_per_epoch=hyperparams.get("steps_per_epoch", 300),
recompute=hyperparams.get("obs_recompute", 2),
s_e=(hyperparams.get("obs_se0",
1), hyperparams.get("obs_seend", 1)),
n_epochs=hyperparams.get("obs_epochs", 1)),
"history":
partial(HistorySampler,
___,
reweighting,
model,
recompute=hyperparams.get("history_recompute", 600)),
"cache":
partial(CacheSampler,
___,
reweighting,
staleness=hyperparams.get("staleness", 3),
cache_prob=hyperparams.get("cache_prob", 0.5),
smooth=hyperparams.get("smooth", 0.2))
}
samplers["lstm-comparison"] = lambda x: LSTMComparisonSampler(
x,
samplers["lstm"](x),
samplers["model"](x),
subset=hyperparams.get("lstm_comparison", 1024))
# Add some decorated samplers
samplers_for_decoration = ["model", "lstm", "pcg", "history"]
for sampler in samplers_for_decoration:
samplers["smooth-" + sampler] = compose(
partial(AdditiveSmoothingSampler,
c=hyperparams.get("smooth", 1.0)), samplers[sampler])
samplers["adaptive-smooth-" + sampler] = compose(
partial(AdaptiveAdditiveSmoothingSampler,
percentage=hyperparams.get("smooth", 0.5)),
samplers[sampler])
samplers["power-smooth-" + sampler] = compose(
partial(PowerSmoothingSampler,
power=hyperparams.get("smooth", 0.5)), samplers[sampler])
for sampler in samplers.keys():
if "model" in sampler or "lstm" in sampler:
samplers["warmup-" + sampler] = compose(
partial(ConditionalStartSampler, ___,
WarmupCondition(hyperparams.get("warmup", 100))),
samplers[sampler])
samplers["exp-" + sampler] = compose(
partial(ConditionalStartSampler, ___,
ExpCondition(hyperparams.get("exp_th", 2.0))),
samplers[sampler])
samplers["tv-" + sampler] = compose(
partial(ConditionalStartSampler, ___,
TotalVariationCondition(hyperparams.get("tv_th",
0.5))),
samplers[sampler])
samplers["vr-" + sampler] = compose(
partial(
ConditionalStartSampler, ___,
VarianceReductionCondition(hyperparams.get("vr_th", 0.5))),
samplers[sampler])
return samplers
def load_dataset(dataset, hyperparams):
datasets = {
"canevet-icml2016-jittered":
partial(CanevetICML2016, smooth=10),
"canevet-icml2016":
CanevetICML2016,
"canevet-icml2016-smooth":
partial(CanevetICML2016, smooth=15000),
"cifar-sanity-check":
CIFARSanityCheck,
"mnist":
MNIST,
"cifar10":
CIFAR10,
"cifar100":
CIFAR100,
"cifar10-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR10),
"cifar10-whitened-augmented":
compose(
partial(OntheflyAugmentedImages,
___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False),
N=15 * 10**5), ZCAWhitening, CIFAR10),
"cifar100-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR100),
"cifar100-whitened-augmented":
compose(
partial(OntheflyAugmentedImages,
___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False),
N=15 * 10**5), ZCAWhitening, CIFAR100),
"ptb":
partial(PennTreeBank, 20),
"imagenet-32x32":
partial(ImageNetDownsampled,
hyperparams.get("imagenet", os.getenv("IMAGENET")),
size=32),
"imagenet-64x64":
partial(ImageNetDownsampled,
hyperparams.get("imagenet", os.getenv("IMAGENET")),
size=64),
"timit":
partial(TIMIT, 20, hyperparams.get("timit", os.getenv("TIMIT"))),
"mit-67":
partial(MIT67, hyperparams.get("mit67", os.getenv("MIT67"))),
"casia":
partial(CASIAWebFace,
hyperparams.get("casia", os.getenv("CASIA")),
alpha=hyperparams.get("triplet_alpha", 0.2),
embedding=hyperparams.get("triplet_embedding", 128)),
"casia-cls":
partial(
CASIAWebFace2,
hyperparams.get("casia", os.getenv("CASIA")),
),
"pixel-mnist":
PixelByPixelMNIST
}
return datasets[dataset]()
def load_dataset(dataset):
datasets = {
"mnist":
MNIST,
"cifar10":
CIFAR10,
"cifar100":
CIFAR100,
"cifar10-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR10),
"cifar10-whitened-augmented":
compose(
partial(OntheflyAugmentedImages,
___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False),
N=15 * 10**5), ZCAWhitening, CIFAR10),
"cifar100-augmented":
compose(
partial(
OntheflyAugmentedImages, ___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False)), CIFAR100),
"cifar100-whitened-augmented":
compose(
partial(OntheflyAugmentedImages,
___,
dict(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=0,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False),
N=15 * 10**5), ZCAWhitening, CIFAR100),
"imagenet-32x32":
partial(ImageNetDownsampled, os.getenv("IMAGENET"), size=32),
"ptb":
partial(PennTreeBank, 20),
}
return datasets[dataset]()
