from functools import reduce from operator import add from typing import Any, Callable, Dict, List, Optional, Tuple, Union import numpy as np import torch from lhotse import CutSet from lhotse.cut import Cut from lhotse.dataset.sampling.base import CutSampler, SamplingDiagnostics class RoundRobinSampler(CutSampler): """ :class:`.RoundRobinSampler` takes several samplers as input, and yields a mini-batch of cuts from each of those samplers in turn. E.g., with two samplers, the first mini-batch is from ``sampler0``, the seconds from ``sampler1``, the third from ``sampler0``, and so on. It is helpful for alternating mini-batches from multiple datasets or manually creating batches of different sizes. The input samplers do not have to provide the same number of batches -- when any of the samplers becomes depleted, we continue to iterate the non-depleted samplers, until all of them are exhausted. Example:: >>> sampler = RoundRobinSampler( ... SimpleCutSampler(cuts_corpusA, max_cuts=32, shuffle=True), ... SimpleCutSampler(cuts_corpusB, max_cuts=64, shuffle=True), ... ) >>> for cut in sampler: ... pass # profit """ def __init__( self, *samplers: CutSampler, stop_early: bool = False, randomize: Union[bool, List[float]] = False, seed: int = 0, ) -> None: """ RoundRobinSampler's constructor. :param samplers: The list of samplers from which we sample batches in turns. :param stop_early: Should we finish the epoch once any of the samplers becomes depleted. By default, we will keep iterating until all the samplers are exhausted. This setting can be used to balance datasets of different sizes. :param randomize: Select the next sampler according to a distribution, instead of in order. If a list of floats is provided, it must contain the same number of elements as the number of samplers, and the values will be used as probabilities. If ``True`` is provided, the probabilities will be uniform. If ``False`` is provided, the samplers will be selected in order. :param seed: Random seed used to select the next sampler (only used if ``randomize`` is True) """ super().__init__(rank=0, world_size=1, seed=seed) self.samplers = samplers self.stop_early = stop_early self.rng = None self._nondepleted_samplers_indices = list(range(len(self.samplers))) self._cur_sampler_idx = 0 self._num_dl_workers = 1 if isinstance(randomize, list): assert len(randomize) == len(self.samplers) elif randomize == True: randomize = [1.0 / len(self.samplers)] * len(self.samplers) self.randomize = randomize @property def remaining_duration(self) -> Optional[float]: """ Remaining duration of data left in the sampler (may be inexact due to float arithmetic). """ try: return sum(s.remaining_duration for s in self.samplers) except TypeError: return None @property def remaining_cuts(self) -> Optional[int]: """ Remaining number of cuts in the sampler. Not available when the CutSet is read in lazy mode (returns None). """ try: return sum(s.remaining_cuts for s in self.samplers) except TypeError: return None @property def num_cuts(self) -> Optional[int]: """ Total number of cuts in the sampler. Not available when the CutSet is read in lazy mode (returns None). """ try: return sum(s.num_cuts for s in self.samplers) except TypeError: return None def allow_iter_to_reset_state(self): """ Enables re-setting to the start of an epoch when iter() is called. This is only needed in one specific scenario: when we restored previous sampler state via ``sampler.load_state_dict()`` but want to discard the progress in the current epoch and start from the beginning. """ super().allow_iter_to_reset_state() for s in self.samplers: s.allow_iter_to_reset_state() def state_dict(self) -> Dict[str, Any]: """ Return the current state of the sampler in a state_dict. Together with ``load_state_dict()``, this can be used to restore the training loop's state to the one stored in the state_dict. """ state_dict = super().state_dict() state_dict.update( { "samplers": [s.state_dict() for s in self.samplers], "stop_early": self.stop_early, "randomize": self.randomize, "_cur_sampler_idx": self._cur_sampler_idx, "_num_dl_workers": self._num_dl_workers, # Explicit list copy below allows to restore within the same process. "_nondepleted_samplers_indices": list( self._nondepleted_samplers_indices ), } ) return state_dict def load_state_dict(self, state_dict: Dict[str, Any]) -> None: """ Restore the state of the sampler that is described in a state_dict. This will result in the sampler yielding batches from where the previous training left it off. .. caution:: The samplers are expected to be initialized with the same CutSets, but this is not explicitly checked anywhere. .. caution:: The input ``state_dict`` is being mutated: we remove each consumed key, and expect it to be empty at the end of loading. If you don't want this behavior, pass a copy inside of this function (e.g., using ``import deepcopy``). .. note:: For implementers of sub-classes of CutSampler: the flag ``self._just_restored_state`` has to be handled in ``__iter__`` to make it avoid resetting the just-restored state (only once). """ self.stop_early = state_dict.pop("stop_early") self.randomize = state_dict.pop("randomize") self._cur_sampler_idx = state_dict.pop("_cur_sampler_idx") self._num_dl_workers = state_dict.pop("_num_dl_workers") self._nondepleted_samplers_indices = state_dict.pop( "_nondepleted_samplers_indices" ) assert len(self.samplers) == len(state_dict["samplers"]), ( "Error in RoundRobinSampler.load_state_dict(): Inconsistent number of samplers: " f"current RoundRobinSampler has {len(self.samplers)}, the state_dict has {len(state_dict['samplers'])}." ) for sampler, sampler_sd in zip(self.samplers, state_dict.pop("samplers")): sampler.load_state_dict(sampler_sd) super().load_state_dict(state_dict) def __iter__(self): self.rng = np.random.default_rng(seed=self.seed + self.epoch) for sampler in self.samplers: iter(sampler) if self._just_restored_state: return self self._nondepleted_samplers_indices = list(range(len(self.samplers))) # In case this sampler lives in the dataloading worker subprocess, # set the starting index to a different value on each dataloading worker. # This helps avoid situations where the round robin sampler chooses # the same underlying sampler for N consecutive mini-batches, where N = num_workers (>1). self._cur_sampler_idx = 0 self._num_dl_workers = 1 worker_info = torch.utils.data.get_worker_info() if worker_info is not None: self._cur_sampler_idx = worker_info.id % len( self._nondepleted_samplers_indices ) self._num_dl_workers = worker_info.num_workers return self def _next_batch(self) -> Union[CutSet, Tuple[CutSet]]: if len(self._nondepleted_samplers_indices) == 0: raise StopIteration() sampler_idx = self._nondepleted_samplers_indices[self._cur_sampler_idx] sampler = self.samplers[sampler_idx] try: batch = next(sampler) except StopIteration: # Try again recursively as long as there is at least one non depleted sampler left. self._nondepleted_samplers_indices.pop(self._cur_sampler_idx) if self.stop_early or len(self._nondepleted_samplers_indices) == 0: raise self._set_next_idx() return self._next_batch() self._set_next_idx() return batch def _set_next_idx(self) -> None: if self.randomize is not False and len(self._nondepleted_samplers_indices) > 1: N = range(len(self._nondepleted_samplers_indices)) p = [self.randomize[i] for i in self._nondepleted_samplers_indices] # Normalize the probabilities p = [x / sum(p) for x in p] self._cur_sampler_idx = self.rng.choice(N, size=1, replace=False, p=p)[0] else: self._cur_sampler_idx = ( self._cur_sampler_idx + self._num_dl_workers ) % len(self._nondepleted_samplers_indices) def set_epoch(self, epoch: int) -> None: """ Sets the epoch for this sampler. When :attr:`shuffle=True`, this ensures all replicas use a different random ordering for each epoch. Otherwise, the next iteration of this sampler will yield the same ordering. :param epoch: Epoch number. """ for s in self.samplers: s.set_epoch(epoch) super().set_epoch(epoch) def filter(self, predicate: Callable[[Cut], bool]) -> None: """ Add a constraint on individual cuts that has to be satisfied to consider them. Can be useful when handling large, lazy manifests where it is not feasible to pre-filter them before instantiating the sampler. Example: >>> cuts = CutSet(...) ... sampler = SimpleCutSampler(cuts, max_duration=100.0) ... # Retain only the cuts that have at least 1s and at most 20s duration. ... sampler.filter(lambda cut: 1.0 <= cut.duration <= 20.0) """ for sampler in self.samplers: sampler.filter(predicate) def _log_diagnostics(self, batch: Union[CutSet, Tuple[CutSet, ...]]) -> None: return # do nothing @property def diagnostics(self) -> SamplingDiagnostics: return reduce(add, (s.diagnostics for s in self.samplers)) def get_report(self) -> str: """Returns a string describing the statistics of the sampling process so far.""" return self.diagnostics.get_report()