bridge.data.vlm_datasets.step37_flickr8k.samplers

Looped sequential / shuffle / weighted-random samplers.

These drive :class:MixedPackedDataloader._schedule_all. The seeds and the exact torch.randperm / heapq-based selection order define the packed batch sequence, so they must not be changed if reproducible packs are required.

Module Contents

Classes

LoopedSequentialSampler	size=3: [0, 1, 2, 0, 1, 2, …]
LoopedShuffleSampler	Looped shuffle sampler.
WeightedRandomSampler	Heap-based balanced weighted sampler.

API

class bridge.data.vlm_datasets.step37_flickr8k.samplers.LoopedSequentialSampler(size: int)

size=3: [0, 1, 2, 0, 1, 2, …]

Initialization

__iter__()

__next__() → int

get() → int

update() → None

state_dict() → dict[str, Any]

load_state_dict(state_dict: dict[str, Any]) → None

class bridge.data.vlm_datasets.step37_flickr8k.samplers.LoopedShuffleSampler(

size: int = 0,

base_seed: int = 1234,

same_order_for_each_epoch: bool = False,

)

Looped shuffle sampler.

Yields a fresh torch.randperm permutation per epoch using seed = base_seed + epoch (or just base_seed if same_order_for_each_epoch is set).

Initialization

__iter__()

__next__() → int

get() → int

update() → None

state_dict() → dict[str, Any]

load_state_dict(state_dict: dict[str, Any]) → None

_reset_idx_cur_epoch() → None

class bridge.data.vlm_datasets.step37_flickr8k.samplers.WeightedRandomSampler(

size: int = 0,

base_seed: int = 1234,

weights: Optional[Union[Sequence[float], torch.Tensor]] = None,

)

Heap-based balanced weighted sampler.

Maintains a min-heap of cumulative scores count[i] / weight[i] and always picks the lowest-score index (ties broken by lower index, per Python’s heapq invariant). This produces an exactly reproducible weighted order without floating-point randomness in the selection.

Initialization

__iter__()

__next__() → int

_select_idx() → int

get() → int

update() → None

state_dict() → dict[str, Any]

load_state_dict(state_dict: dict[str, Any]) → None