def build(task: str, cfg: CfgNode):
r"""
Builder function.
Arguments
---------
task: str
builder task name (track|vos)
cfg: CfgNode
buidler configuration
Returns
-------
torch.nn.Module
module built by builder
"""
if task in TASK_BACKBONES:
modules = TASK_BACKBONES[task]
else:
logger.error("no backbone for task {}".format(task))
exit(-1)
name = cfg.name
assert name in modules, "backbone {} not registered for {}!".format(
name, task)
module = modules[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode, seed: int = 0) -> TransformerBase:
r"""
Arguments
---------
task: str
task
cfg: CfgNode
node name: transformer
"""
assert task in TASK_TRANSFORMERS, "invalid task name"
MODULES = TASK_TRANSFORMERS[task]
names = cfg.names
modules = []
for name in names:
module = MODULES[name](seed=seed)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
modules.append(module)
return modules
def build(task: str, cfg: CfgNode):
r"""
Builder function.
Arguments
---------
task: str
builder task name (track|vos)
cfg: CfgNode
buidler configuration
Returns
-------
torch.nn.Module
module built by builder
"""
if task in TASK_HEADS:
head_modules = TASK_HEADS[task]
else:
logger.error("no task model for task {}".format(task))
exit(-1)
name = cfg.name
head_module = head_modules[name]()
hps = head_module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
head_module.set_hps(hps)
head_module.update_params()
return head_module
def build(task: str, cfg: CfgNode, seed: int = 0) -> DatasetBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: sampler
seed: int
seed for rng initialization
"""
assert task in TASK_SAMPLERS, "invalid task name"
MODULES = TASK_SAMPLERS[task]
submodules_cfg = cfg.submodules
dataset_cfg = submodules_cfg.dataset
datasets = dataset_builder.build(task, dataset_cfg)
filter_cfg = getattr(submodules_cfg, "filter", None)
filt = filter_builder.build(task,
filter_cfg) if filter_cfg is not None else None
name = cfg.name
module = MODULES[name](datasets, seed=seed, filt=filt)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode, seed: int = 0) -> DatapipelineBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: data
seed: int
seed for rng initialization
"""
assert task in TASK_DATAPIPELINES, "invalid task name"
MODULES = TASK_DATAPIPELINES[task]
sampler = build_sampler(task, cfg.sampler, seed=seed)
transformers = build_transformer(task, cfg.transformer, seed=seed)
target = build_target(task, cfg.target)
pipeline = []
pipeline.extend(transformers)
pipeline.append(target)
cfg = cfg.datapipeline
name = cfg.name
module = MODULES[name](sampler, pipeline)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode, model: ModuleBase):
r"""
Builder function.
Arguments
---------
task: str
task name
cfg: CfgNode
buidler configuration
model: ModuleBase
model instance
Returns
-------
torch.nn.Module
module built by builder
"""
assert task in PIPELINES, "no pipeline for task {}".format(task)
pipelines = PIPELINES[task]
pipeline_name = cfg.name
pipeline = pipelines[pipeline_name](model)
hps = pipeline.get_hps()
hps = merge_cfg_into_hps(cfg[pipeline_name], hps)
pipeline.set_hps(hps)
pipeline.update_params()
return pipeline
def build(task: str, cfg: CfgNode, pipeline: PipelineBase):
r"""
Builder function.
Arguments
---------
task: str
builder task name (track|vos)
cfg: CfgNode
buidler configuration
Returns
-------
TesterBase
tester built by builder
"""
assert task in TASK_TESTERS, "no tester for task {}".format(task)
MODULES = TASK_TESTERS[task]
names = cfg.tester.names
testers = []
# tester for multiple experiments
for name in names:
tester = MODULES[name](pipeline)
hps = tester.get_hps()
hps = merge_cfg_into_hps(cfg.tester[name], hps)
tester.set_hps(hps)
tester.update_params()
testers.append(tester)
return testers
def build(task: str, cfg: CfgNode) -> DatasetBase:
assert task in TASK_FILTERS, "invalid task name"
MODULES = TASK_FILTERS[task]
name = cfg.name
module = MODULES[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str,
cfg: CfgNode,
backbone: ModuleBase,
head: ModuleBase,
loss: ModuleBase = None):
r"""
Builder function.
Arguments
---------
task: str
builder task name
cfg: CfgNode
buidler configuration
backbone: torch.nn.Module
backbone used by task module.
head: torch.nn.Module
head network used by task module.
loss: torch.nn.Module
criterion module used by task module (for training). None in case other than training.
Returns
-------
torch.nn.Module
task module built by builder
"""
if task == "track":
task_modules = TRACK_TASKMODELS
elif task == "vos":
task_modules = VOS_TASKMODELS
else:
logger.error("no task model for task {}".format(task))
exit(-1)
if task == "track":
name = cfg.name
task_module = task_modules[name](backbone, head, loss)
hps = task_module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
task_module.set_hps(hps)
task_module.update_params()
return task_module
else:
logger.error("task model {} is not completed".format(task))
exit(-1)
def build(
task: str,
cfg: CfgNode,
model: ModuleBase = None,
segmenter: ModuleBase = None,
tracker: ModuleBase = None,
):
r"""
Builder function.
Arguments
---------
task: str
task name
cfg: CfgNode
buidler configuration
model: ModuleBase
model instance for siamfcpp
segmenter: ModuleBase
segmenter instance for tracker
tracker: ModuleBase
model instance for tracker
Returns
-------
torch.nn.Module
module built by builder
"""
assert task in PIPELINES, "no pipeline for task {}".format(task)
pipelines = PIPELINES[task]
pipeline_name = cfg.name
if task == 'track':
pipeline = pipelines[pipeline_name](model)
elif task == 'vos':
pipeline = pipelines[pipeline_name](segmenter, tracker)
else:
logger.error("unknown task {} for pipline".format(task))
exit(-1)
hps = pipeline.get_hps()
hps = merge_cfg_into_hps(cfg[pipeline_name], hps)
pipeline.set_hps(hps)
pipeline.update_params()
return pipeline
def build(task: str, cfg: CfgNode, basemodel=None):
r"""
backbone build 函数
根据传入的 task(track|vos) 及配置信息构建并返回 backbone 模块
Builder function.
Arguments
---------
task: str
builder task name (track|vos)
cfg: CfgNode
buidler configuration
basemodel:
warp backbone into encoder if not None
Returns
-------
torch.nn.Module
module built by builder
"""
if task in TASK_BACKBONES:
modules = TASK_BACKBONES[
task] # 检索对应的 backbones(TRACK_BACKBONES|VOS_BACKBONES)
else:
logger.error("no backbone for task {}".format(task))
exit(-1) # 检索失败
name = cfg.name # 配置信息中的目标 backbone 名称
assert name in modules, "backbone {} not registered for {}!".format(
name, task)
if basemodel:
module = modules[name](basemodel)
else:
module = modules[name](
) # 根据 backbone 名称检索对应的 backbone module （如 AlexNet）并实例化
hps = module.get_hps() # 获取 backbone 实例的超参数字典
hps = merge_cfg_into_hps(cfg[name], hps) # 将配置信息中的该 backbone 类的超参数数据合并入字典中
module.set_hps(hps) # 重新设置 backbone 实例的超参数
module.update_params()
return module
def build(task: str, cfg: CfgNode, model: nn.Module) -> OptimizerBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: optim
"""
name = cfg.name
module = OPTIMIZERS[name](cfg, model)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build_sat_model(task: str,
cfg: CfgNode,
gml_extractor=None,
joint_encoder=None,
decoder=None,
loss: ModuleBase = None):
r"""
Builder function for SAT.
Arguments
---------
task: str
builder task name
cfg: CfgNode
buidler configuration
gml_extractor: torch.nn.Module
feature extractor for global modeling loop
joint_encoder: torch.nn.Module
joint encoder
decoder: torch.nn.Module
decoder for SAT
loss: torch.nn.Module
criterion module used by task module (for training). None in case other than training.
Returns
-------
torch.nn.Module
task module built by builder
"""
if task == "vos":
task_modules = TASK_TASKMODELS[task]
else:
logger.error("sat model builder could not build task {}".format(task))
exit(-1)
name = cfg.name
#SatVOS
task_module = task_modules[name](gml_extractor, joint_encoder, decoder,
loss)
hps = task_module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
task_module.set_hps(hps)
task_module.update_params()
return task_module
def build(task: str, cfg: CfgNode) -> TargetBase:
r"""
Arguments
---------
task: str
task
cfg: CfgNode
node name: target
"""
assert task in TASK_TARGETS, "invalid task name"
MODULES = TASK_TARGETS[task]
name = cfg.name
module = MODULES[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode, model: nn.Module) -> OptimizerBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: optim
"""
assert task in TASK_OPTIMIZERS, "invalid task name"
MODULES = TASK_OPTIMIZERS[task]
name = cfg.name
module = MODULES[name](cfg, model)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode) -> ContribModule:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: contrib_module
"""
assert task in TASK_CONTRIB_MODULES, "invalid task name"
MODULES = TASK_CONTRIB_MODULES[task]
name = cfg.name
module = MODULES[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode) -> GradModifierBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: scheduler
seed: int
seed for rng initialization
"""
name = cfg.name
module = GRAD_MODIFIERS[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode) -> TemplateModuleBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: template_module
"""
assert task in TASK_TEMPLATE_MODULES, "invalid task name"
MODULES = TASK_TEMPLATE_MODULES[task] # # 检索 task 对应的 module
name = cfg.name
module = MODULES[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build(task: str, cfg: CfgNode, pyramid_model=None):
r"""
Builder function.
Arguments
---------
task: str
builder task name (track|vos)
cfg: CfgNode
builder configuration
pyramid_model:
warp pyramid into encoder if not None
Returns
-------
torch.nn.Module
module built by builder
"""
if task in TASK_PYRAMIDS:
modules = TASK_PYRAMIDS[task]
else:
logger.error("no pyramid for task {}".format(task))
exit(-1)
name = cfg.name
assert name in modules, "pyramid {} not registered for {}!".format(
name, task)
if pyramid_model:
module = modules[name](pyramid_model)
else:
module = modules[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
def build_track_dual_backbone(task: str,
cfg: CfgNode,
basemodel_target=None,
basemodel_search=None,
head=None,
loss: ModuleBase = None):
r"""
Builder function for SiamFCpp
In case of the siamese branches do not share weights
Arguments
---------
task: str
builder task name
cfg: CfgNode
buidler configuration
basemodel_target: torch.nn.Module
backbone used by target image backbone.
basemodel_search: torch.nn.Module
backbone used by search image backbone.
head: torch.nn.Module
head network used by task module.
loss: torch.nn.Module
criterion module used by task module (for training). None in case other than training.
Returns
-------
torch.nn.Module
task module built by builder
"""
task_modules = TASK_TASKMODELS[task]
name = cfg.name
task_module = task_modules[name](basemodel_target, basemodel_search, head,
loss)
hps = task_module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
task_module.set_hps(hps)
task_module.update_params()
return task_module
def build(task: str, cfg: CfgNode) -> DatasetBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: dataset
"""
assert task in TASK_DATASETS, "invalid task name"
dataset_modules = TASK_DATASETS[task]
names = cfg.names
modules = []
for name in names:
module = dataset_modules[name]()
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
modules.append(module)
return modules
def build(task: str, cfg: CfgNode, seed: int = 0) -> DatapipelineBase:
r"""
Arguments
---------
task: str
task name (track|vos)
cfg: CfgNode
node name: data
seed: int
seed for rng initialization
"""
assert task in TASK_DATAPIPELINES, "invalid task name"
MODULES = TASK_DATAPIPELINES[task]
sampler = build_sampler(
task, cfg.sampler,
seed=seed) # 从同一个序列中产生一个positive pair或者是从不同序列中产生一个negative pair
transformers = build_transformer(
task, cfg.transformer, seed=seed
) # 数据增广，主要有scale和shift，使search patch和templete patch并不是正好目标中心对着的，会有一点偏移
target = build_target(task, cfg.target) # 构造label
pipeline = []
pipeline.extend(transformers)
pipeline.append(target)
cfg = cfg.datapipeline
name = cfg.name
module = MODULES[name](sampler, pipeline)
hps = module.get_hps()
hps = merge_cfg_into_hps(cfg[name], hps)
module.set_hps(hps)
module.update_params()
return module
