head class to instantiate. For instance, a config `{"name": "my_head",
"foo": "bar"}` will find a class that was registered as "my_head"
(see :func:`register_head`) and call .from_config on it."""
assert "name" in config, "Expect name in config"
assert "unique_id" in config, "Expect a global unique id in config"
assert config["name"] in HEAD_REGISTRY, "unknown head {}".format(
config["name"])
name = config["name"]
head_config = copy.deepcopy(config)
del head_config["name"]
return HEAD_REGISTRY[name].from_config(head_config)
# automatically import any Python files in the heads/ directory
import_all_modules(FILE_ROOT, "classy_vision.heads")
from .fully_connected_head import FullyConnectedHead # isort:skip
from .fully_convolutional_linear_head import FullyConvolutionalLinearHead # isort:skip
from .identity_head import IdentityHead # isort:skip
from .vision_transformer_head import VisionTransformerHead # isort:skip
__all__ = [
"ClassyHead",
"FullyConnectedHead",
"FullyConvolutionalLinearHead",
"IdentityHead",
"VisionTransformerHead",
"build_head",
"register_head",
]
return cls
return register_model_head_cls
def get_model_head(name: str):
"""
Given the model head name, construct the head if it's registered
with VISSL.
"""
assert name in MODEL_HEADS_REGISTRY, "Unknown model head"
return MODEL_HEADS_REGISTRY[name]
# automatically import any Python files in the heads/ directory
import_all_modules(FILE_ROOT, "vissl.models.heads")
from vissl.models.heads.linear_eval_mlp import LinearEvalMLP # isort:skip # noqa
from vissl.models.heads.mlp import MLP # isort:skip # noqa
from vissl.models.heads.siamese_concat_view import ( # isort:skip # noqa
SiameseConcatView, )
from vissl.models.heads.swav_prototypes_head import ( # isort:skip # noqa
SwAVPrototypesHead, )
__all__ = [
"get_model_head",
"LinearEvalMLP",
"MLP",
"SiameseConcatView",
"SwAVPrototypesHead",
]
def register_loss_cls(cls):
if name in LOSS_REGISTRY:
raise ValueError(
"Cannot register duplicate optimizer ({})".format(name))
if not issubclass(cls, ClassyLoss):
raise ValueError("Loss ({}: {}) must extend ClassyLoss".format(
name, cls.__name__))
LOSS_REGISTRY[name] = cls
LOSS_CLASS_NAMES.add(cls.__name__)
return cls
return register_loss_cls
# automatically import any Python files in the losses/ directory
import_all_modules(FILE_ROOT, "classy_vision.losses")
from .barron_loss import BarronLoss # isort:skip
from .label_smoothing_loss import LabelSmoothingCrossEntropyLoss # isort:skip
from .multi_output_sum_loss import MultiOutputSumLoss # isort:skip
from .soft_target_cross_entropy_loss import SoftTargetCrossEntropyLoss # isort:skip
from .sum_arbitrary_loss import SumArbitraryLoss # isort:skip
__all__ = [
"BarronLoss",
"ClassyLoss",
"LabelSmoothingCrossEntropyLoss",
"MultiOutputSumLoss",
"SoftTargetCrossEntropyLoss",
"SumArbitraryLoss",
"build_loss",
)
if cls.__name__ in OPTIMIZER_CLASS_NAMES:
raise ValueError(
"Cannot register optimizer with duplicate class name({})".format(
cls.__name__
)
)
OPTIMIZER_REGISTRY[name] = cls
OPTIMIZER_CLASS_NAMES.add(cls.__name__)
return cls
return register_optimizer_cls
# automatically import any Python files in the optim/ directory
import_all_modules(FILE_ROOT, "classy_vision.optim")
from .adam import Adam # isort:skip
from .rmsprop import RMSProp # isort:skip
from .rmsprop_tf import RMSPropTF # isort:skip
from .sgd import SGD # isort:skip
__all__ = [
"Adam",
"ClassyOptimizer",
"RMSProp",
"RMSPropTF",
"SGD",
"build_optimizer",
"register_optimizer",
]
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # Automatically import any Python files in the losses/ directory import_all_modules(FILE_ROOT, "losses")
...
To get a model trunk from a configuration file, see :func:`get_model_trunk`."""
def register_model_trunk_cls(cls: Callable[..., Callable]):
if name in MODEL_TRUNKS_REGISTRY:
raise ValueError(
"Cannot register duplicate model trunk ({})".format(name))
if cls.__name__ in MODEL_TRUNKS_NAMES:
raise ValueError(
"Cannot register task with duplicate model trunk name ({})".
format(cls.__name__))
MODEL_TRUNKS_REGISTRY[name] = cls
MODEL_TRUNKS_NAMES.add(cls.__name__)
return cls
return register_model_trunk_cls
def get_model_trunk(name: str):
"""
Given the model trunk name, construct the trunk if it's registered
with VISSL.
"""
assert name in MODEL_TRUNKS_REGISTRY, "Unknown model trunk"
return MODEL_TRUNKS_REGISTRY[name]
# automatically import any Python files in the trunks/ directory
import_all_modules(FILE_ROOT, "vissl.models.trunks")
raise ValueError("Task ({}: {}) must extend ClassyTask".format(
name, cls.__name__))
if cls.__name__ in TASK_CLASS_NAMES:
msg = ("Cannot register task with duplicate class name({})." +
"Previously registered at \n{}\n")
raise ValueError(
msg.format(cls.__name__, TASK_CLASS_NAMES_TB[cls.__name__]))
tb = "".join(traceback.format_stack())
TASK_REGISTRY[name] = cls
TASK_CLASS_NAMES.add(cls.__name__)
TASK_REGISTRY_TB[name] = tb
TASK_CLASS_NAMES_TB[cls.__name__] = tb
return cls
return register_task_cls
from .classification_task import ClassificationTask # isort:skip
from .fine_tuning_task import FineTuningTask # isort:skip
__all__ = [
"ClassyTask",
"FineTuningTask",
"build_task",
"register_task",
"ClassificationTask",
]
# automatically import any Python files in the tasks/ directory
import_all_modules(FILE_ROOT, "classy_vision.tasks")
if hasattr(transforms, name) or hasattr(transforms_video, name):
raise ValueError(
"{} has existed in torchvision.transforms, Please change the name!".format(
name
)
)
TRANSFORM_REGISTRY[name] = cls
tb = "".join(traceback.format_stack())
TRANSFORM_REGISTRY_TB[name] = tb
return cls
return register_transform_cls
# automatically import any Python files in the transforms/ directory
import_all_modules(FILE_ROOT, "classy_vision.dataset.transforms")
from .lighting_transform import LightingTransform # isort:skip
from .util import ApplyTransformToKey # isort:skip
from .util import ImagenetAugmentTransform # isort:skip
from .util import ImagenetAugmentTransform # isort:skip
from .util import ImagenetNoAugmentTransform # isort:skip
from .util import GenericImageTransform # isort:skip
from .util import TupleToMapTransform # isort:skip
__all__ = [
"ClassyTransform",
"ImagenetAugmentTransform",
"ImagenetNoAugmentTransform",
"GenericImageTransform",
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # automatically import any Python files in the meters/ directory import_all_modules(FILE_ROOT, "vissl.meters")
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # automatically import any Python files in the param_scheduler/ directory import_all_modules(FILE_ROOT, "vissl.optimizers.param_scheduler")
"Previously registered at \n{}\n")
raise ValueError(
msg.format(cls.__name__,
DATASET_CLASS_NAMES_TB[cls.__name__]))
tb = "".join(traceback.format_stack())
DATASET_REGISTRY[name] = cls
DATASET_CLASS_NAMES.add(cls.__name__)
DATASET_REGISTRY_TB[name] = tb
DATASET_CLASS_NAMES_TB[cls.__name__] = tb
return cls
return register_dataset_cls
# automatically import any Python files in the dataset/ directory
import_all_modules(FILE_ROOT, "classy_vision.dataset")
from .classy_cifar import CIFARDataset # isort:skip
from .classy_hmdb51 import HMDB51Dataset # isort:skip
from .classy_kinetics400 import Kinetics400Dataset # isort:skip
from .classy_synthetic_image import SyntheticImageDataset # isort:skip
from .classy_synthetic_image_streaming import ( # isort:skip
SyntheticImageStreamingDataset, # isort:skip
) # isort:skip
from .classy_synthetic_video import SyntheticVideoDataset # isort:skip
from .classy_ucf101 import UCF101Dataset # isort:skip
from .classy_video_dataset import ClassyVideoDataset # isort:skip
from .dataloader_async_gpu_wrapper import DataloaderAsyncGPUWrapper # isort:skip
from .dataloader_limit_wrapper import DataloaderLimitWrapper # isort:skip
from .dataloader_skip_none_wrapper import DataloaderSkipNoneWrapper # isort:skip
from .dataloader_wrapper import DataloaderWrapper # isort:skip
def register_param_scheduler_cls(cls):
if name in PARAM_SCHEDULER_REGISTRY:
raise ValueError(
"Cannot register duplicate param scheduler ({})".format(name))
if not issubclass(cls, ParamScheduler):
raise ValueError(
"Param Scheduler ({}: {}) must extend ParamScheduler".format(
name, cls.__name__))
PARAM_SCHEDULER_REGISTRY[name] = cls
return cls
return register_param_scheduler_cls
# automatically import any Python files in the optim/param_scheduler/ directory
import_all_modules(FILE_ROOT, "classy_vision.optim.param_scheduler")
from .composite_scheduler import CompositeParamScheduler, IntervalScaling # isort:skip
from .fvcore_schedulers import (
ConstantParamScheduler,
CosineParamScheduler,
LinearParamScheduler,
MultiStepParamScheduler,
PolynomialDecayParamScheduler,
StepParamScheduler,
StepWithFixedGammaParamScheduler,
) # isort:skip
__all__ = [
"ParamScheduler",
"ClassyParamScheduler",
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # automatically import any Python files in the losses/ directory import_all_modules(FILE_ROOT, "vissl.losses")
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from pathlib import Path
import torchvision.transforms as pth_transforms
from classy_vision.generic.registry_utils import import_all_modules
from vissl.data.ssl_transforms.ssl_transforms_wrapper import SSLTransformsWrapper
def get_transform(input_transforms_list):
"""
Given the list of user specified transforms, return the
torchvision.transforms.Compose() version of the transforms. Each transform
in the composition is SSLTransformsWrapper which wraps the original
transforms to handle multi-modal nature of input.
"""
output_transforms = []
for transform_config in input_transforms_list:
transform = SSLTransformsWrapper.from_config(transform_config)
output_transforms.append(transform)
return pth_transforms.Compose(output_transforms)
FILE_ROOT = Path(__file__).parent
import_all_modules(FILE_ROOT, "vissl.data.ssl_transforms")
__all__ = ["get_transform"]
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # Automatically import any Python files in the datasets/ directory import_all_modules(FILE_ROOT, "datasets")
def register_collator_fn(func):
if name in COLLATOR_REGISTRY:
raise ValueError(
"Cannot register duplicate collator ({})".format(name))
if func.__name__ in COLLATOR_NAMES:
raise ValueError(
"Cannot register task with duplicate collator name ({})".
format(func.__name__))
COLLATOR_REGISTRY[name] = func
COLLATOR_NAMES.add(func.__name__)
return func
return register_collator_fn
def get_collator(collator_name, collate_params):
"""
Given the collator name and the collator params, return the collator
if registered with VISSL. Also supports pytorch default collators.
"""
if collator_name == "default_collate":
return default_collate
else:
assert collator_name in COLLATOR_REGISTRY, "Unknown collator"
return partial(COLLATOR_REGISTRY[collator_name], **collate_params)
# automatically import any Python files in the collators/ directory
import_all_modules(FILE_ROOT, "vissl.data.collators")
#!/usr/bin/env python3 # Copyright (c) Facebook, Inc. and its affiliates. # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. from pathlib import Path from classy_vision.generic.registry_utils import import_all_modules FILE_ROOT = Path(__file__).parent # Automatically import any Python files in the models/ directory import_all_modules(FILE_ROOT, "models")
if "heads" in config:
heads = defaultdict(list)
for head_config in config["heads"]:
assert "fork_block" in head_config, "Expect fork_block in config"
fork_block = head_config["fork_block"]
updated_config = copy.deepcopy(head_config)
del updated_config["fork_block"]
head = build_head(updated_config)
heads[fork_block].append(head)
model.set_heads(heads)
return model
# automatically import any Python files in the models/ directory
import_all_modules(FILE_ROOT, "classy_vision.models")
from .classy_block import ClassyBlock # isort:skip
from .classy_model import ( # isort:skip
ClassyModelWrapper, # isort:skip
ClassyModelHeadExecutorWrapper, # isort:skip
) # isort:skip
from .densenet import DenseNet # isort:skip
from .efficientnet import EfficientNet # isort:skip
from .mlp import MLP # isort:skip
from .regnet import RegNet # isort:skip
from .resnet import ResNet # isort:skip
from .resnext import ResNeXt # isort:skip
from .resnext3d import ResNeXt3D # isort:skip
from .squeeze_and_excitation_layer import SqueezeAndExcitationLayer # isort:skip
To get a train step from a configuration file, see :func:`get_train_step`.
"""
def register_train_step_fn(func):
if name in TRAIN_STEP_REGISTRY:
raise ValueError(
"Cannot register duplicate train step ({})".format(name))
if func.__name__ in TRAIN_STEP_NAMES:
raise ValueError(
"Cannot register task with duplicate train step name ({})".
format(func.__name__))
TRAIN_STEP_REGISTRY[name] = func
TRAIN_STEP_NAMES.add(func.__name__)
return func
return register_train_step_fn
def get_train_step(train_step_name: str):
"""
Lookup the train_step_name in the train step registry and return.
If the train step is not implemented, asserts will be thrown and workflow will exit.
"""
assert train_step_name in TRAIN_STEP_REGISTRY, "Unknown train step"
return TRAIN_STEP_REGISTRY[train_step_name]
# automatically import any Python files in the train_steps/ directory
import_all_modules(FILE_ROOT, "vissl.trainer.train_steps")