update

speech_asr

2023-04-18 da9ac240cb3298248acb3262ed96b87fa3c1fa56

update

 funasr/datasets/small_datasets/build_loader.py

@@ -6,7 +6,7 @@

from funasr.datasets.small_datasets.dataset import ESPnetDataset
from funasr.datasets.small_datasets.preprocessor import build_preprocess
from funasr.samplers.length_batch_sampler import LengthBatchSampler
from funasr.datasets.small_datasets.length_batch_sampler import LengthBatchSampler


def build_dataloader(args, mode="train"):
@@ -26,10 +26,6 @@
        preprocess=preprocess_fn,
        dest_sample_rate=dest_sample_rate,
    )
    if os.path.exists(os.path.join(data_path_and_name_and_type[0][0].parent, "utt2category")):
        utt2category_file = os.path.join(data_path_and_name_and_type[0][0].parent, "utt2category")
    else:
        utt2category_file = None

    dataset_conf = args.dataset_conf
    batch_sampler = LengthBatchSampler(

 funasr/datasets/small_datasets/length_batch_sampler.py

New file
@@ -0,0 +1,147 @@
from typing import Iterator
from typing import List
from typing import Dict
from typing import Tuple
from typing import Union

from typeguard import check_argument_types

from funasr.fileio.read_text import load_num_sequence_text
from funasr.samplers.abs_sampler import AbsSampler


class LengthBatchSampler(AbsSampler):
    def __init__(
        self,
        batch_bins: int,
        shape_files: Union[Tuple[str, ...], List[str], Dict],
        min_batch_size: int = 1,
        sort_in_batch: str = "descending",
        sort_batch: str = "ascending",
        drop_last: bool = False,
        padding: bool = True,
    ):
        assert check_argument_types()
        assert batch_bins > 0
        if sort_batch != "ascending" and sort_batch != "descending":
            raise ValueError(
                f"sort_batch must be ascending or descending: {sort_batch}"
            )
        if sort_in_batch != "descending" and sort_in_batch != "ascending":
            raise ValueError(
                f"sort_in_batch must be ascending or descending: {sort_in_batch}"
            )

        self.batch_bins = batch_bins
        self.shape_files = shape_files
        self.sort_in_batch = sort_in_batch
        self.sort_batch = sort_batch
        self.drop_last = drop_last

        # utt2shape: (Length, ...)
        #    uttA 100,...
        #    uttB 201,...
        if isinstance(shape_files, dict):
            utt2shapes = [shape_files]
        else:
            utt2shapes = [
                load_num_sequence_text(s, loader_type="csv_int") for s in shape_files
            ]

        first_utt2shape = utt2shapes[0]
        for s, d in zip(shape_files, utt2shapes):
            if set(d) != set(first_utt2shape):
                raise RuntimeError(
                    f"keys are mismatched between {s} != {shape_files[0]}"
                )

        # Sort samples in ascending order
        # (shape order should be like (Length, Dim))
        keys = sorted(first_utt2shape, key=lambda k: first_utt2shape[k][0])
        if len(keys) == 0:
            raise RuntimeError(f"0 lines found: {shape_files[0]}")

        # Decide batch-sizes
        batch_sizes = []
        current_batch_keys = []
        for key in keys:
            current_batch_keys.append(key)
            # shape: (Length, dim1, dim2, ...)
            if padding:
                # bins = bs x max_length
                bins = sum(len(current_batch_keys) * sh[key][0] for sh in utt2shapes)
            else:
                # bins = sum of lengths
                bins = sum(d[k][0] for k in current_batch_keys for d in utt2shapes)

            if bins > batch_bins and len(current_batch_keys) >= min_batch_size:
                batch_sizes.append(len(current_batch_keys))
                current_batch_keys = []
        else:
            if len(current_batch_keys) != 0 and (
                not self.drop_last or len(batch_sizes) == 0
            ):
                batch_sizes.append(len(current_batch_keys))

        if len(batch_sizes) == 0:
            # Maybe we can't reach here
            raise RuntimeError("0 batches")

        # If the last batch-size is smaller than minimum batch_size,
        # the samples are redistributed to the other mini-batches
        if len(batch_sizes) > 1 and batch_sizes[-1] < min_batch_size:
            for i in range(batch_sizes.pop(-1)):
                batch_sizes[-(i % len(batch_sizes)) - 1] += 1

        if not self.drop_last:
            # Bug check
            assert sum(batch_sizes) == len(keys), f"{sum(batch_sizes)} != {len(keys)}"

        # Set mini-batch
        self.batch_list = []
        iter_bs = iter(batch_sizes)
        bs = next(iter_bs)
        minibatch_keys = []
        for key in keys:
            minibatch_keys.append(key)
            if len(minibatch_keys) == bs:
                if sort_in_batch == "descending":
                    minibatch_keys.reverse()
                elif sort_in_batch == "ascending":
                    # Key are already sorted in ascending
                    pass
                else:
                    raise ValueError(
                        "sort_in_batch must be ascending"
                        f" or descending: {sort_in_batch}"
                    )
                self.batch_list.append(tuple(minibatch_keys))
                minibatch_keys = []
                try:
                    bs = next(iter_bs)
                except StopIteration:
                    break

        if sort_batch == "ascending":
            pass
        elif sort_batch == "descending":
            self.batch_list.reverse()
        else:
            raise ValueError(
                f"sort_batch must be ascending or descending: {sort_batch}"
            )

    def __repr__(self):
        return (
            f"{self.__class__.__name__}("
            f"N-batch={len(self)}, "
            f"batch_bins={self.batch_bins}, "
            f"sort_in_batch={self.sort_in_batch}, "
            f"sort_batch={self.sort_batch})"
        )

    def __len__(self):
        return len(self.batch_list)

    def __iter__(self) -> Iterator[Tuple[str, ...]]:
        return iter(self.batch_list)