Support downloading specific splits in load_dataset
#6832
Conversation
…and_prepare-if-missing-splits
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
| _dataset_name = self.name if self._check_legacy_cache() else self.dataset_name | ||
| splits: Optional[List[str]] = None | ||
| cached_split_filepatterns = [] | ||
| supports_partial_generation = self._supports_partial_generation() |
There was a problem hiding this comment.
(nit) maybe rename _supports_partial_generation -> _supports_split_by_split_generation
(or alternatively isinstance(DatasetSplitBuilder) and DatasetSplitBuilder would implement the method to list the splits and have the extended _split_generators signature, but maybe it's too much)
| split_names = [rel_instr.splitname for rel_instr in split._relative_instructions] | ||
| splits.extend(split_names) | ||
| splits = list(unique_values(splits)) # remove duplicates | ||
| available_splits = self._available_splits() |
There was a problem hiding this comment.
it can also be a property .splits
| # We cannot use info as the source of truth if the builder supports partial generation | ||
| # as the info can be incomplete in that case | ||
| requested_splits_exist = not splits if supports_partial_generation else info_exists |
There was a problem hiding this comment.
is this just an optimization to avoid checking for files for all the splits ?
| shutil.rmtree(dirname) | ||
| # LocalFileSystem.mv does copy + rm, it is more efficient to simply rename a local directory | ||
| shutil.move(tmp_dir, dirname) | ||
| for root, dirnames, filenames in os.walk(dirname, topdown=False): |
There was a problem hiding this comment.
iirc this directory is a flat directory no ? (no subdirectories)
maybe this code can be simplified a bit
| # We need to update the info in case some splits were added in the meantime | ||
| # for example when calling load_dataset from multiple workers. | ||
| self.info = self._load_info() | ||
| _dataset_name = self.name if self._check_legacy_cache() else self.dataset_name | ||
| splits: Optional[List[str]] = None | ||
| cached_split_filepatterns = [] |
There was a problem hiding this comment.
maybe rename patterns_of_split_files_to_overwrite
| @@ -1031,8 +1110,14 @@ def incomplete_dir(dirname): | |||
| **download_and_prepare_kwargs, | |||
| ) | |||
| # Sync info | |||
| if supports_partial_generation and self.info.download_checksums is not None: | |||
There was a problem hiding this comment.
(nit) not sure supports_partial_generation is needed here
| if supports_partial_generation and self.info.download_checksums is not None: | |
| if self.info.download_checksums is not None: |
|
Friendly ping on this! This feature would be really helpful and useful to me (and likely others with limited download speed and storage space!). Thanks so much! |
|
No one is working on this atm afaik :/ |
|
No worries! I've patched the ImageNet dataset in: https://huggingface.co/datasets/ILSVRC/imagenet-1k/blob/refs%2Fpr%2F20/imagenet-1k.py Together with: dataset = load_dataset(
"ILSVRC/imagenet-1k",
split="validation",
data_files={"val": "data/val_images.tar.gz"},
revision="refs/pr/20",
trust_remote_code=True,
download_config=DownloadConfig(resume_download=True),
verification_mode=VerificationMode.NO_CHECKS,
)It only downloads the validation set this way (NO_CHECKS is a bit annoying because I'd rather have md5 checks, but I guess I can't have everything) ^^' The patch is not perfect, but it does the job. |
This PR builds on #6639 to support downloading only the specified splits in
load_dataset. For this to work, a builder's_split_generatorsneed to be able to accept the requested splits (as a list) via asplitsargument to avoid processing the non-requested ones. Also, the builder has to define a_available_splitsmethod that lists all the possiblesplitsvalues.Close #4101, close #2538 (I'm probably missing some)
Should also make it possible to address #6793