def upload(self, filepaths: List[str], split: str, col_mapping: Dict[str, str]):
"""Uploads files to the project"""
local_dataset_dir = os.path.expanduser(f"~/.huggingface/autonlp/projects/{self.dataset_id}")
if os.path.exists(local_dataset_dir):
if os.path.isdir(os.path.join(local_dataset_dir, "git")):
clone_from = None
else:
shutil.rmtree(local_dataset_dir)
clone_from = "https://huggingface.co/datasets/" + self.dataset_id
else:
clone_from = "https://huggingface.co/datasets/" + self.dataset_id
dataset_repo = Repository(
local_dir=local_dataset_dir,
clone_from=clone_from,
use_auth_token=self._token,
)
dataset_repo.git_pull()
for idx, file_path in enumerate(filepaths):
if not os.path.isfile(file_path):
logger.error(f"[{idx + 1}/{len(filepaths)}] ❌ '{file_path}' does not exist or is not a file!")
continue
file_name = os.path.basename(file_path)
file_extension = file_name.split(".")[-1]
src = os.path.expanduser(file_path)
dst = os.path.join(local_dataset_dir, "raw", file_name)
logger.info(f"[{idx + 1}/{len(filepaths)}] 📦 Copying {src} to {dst}...")
os.makedirs(os.path.dirname(dst), exist_ok=True)
shutil.copyfile(src, dst)
logger.info(f"[{idx + 1}/{len(filepaths)}] 🔎 Validating {dst} and column mapping...")
validate_file(path=dst, task=self.task, file_ext=file_extension, col_mapping=col_mapping)
dataset_repo.lfs_track(patterns=[f"raw/*.{file_extension}"])
dataset_repo.git_pull()
try:
logger.info("☁ Uploading files to the dataset hub...")
dataset_repo.push_to_hub(commit_message="Upload from AutoNLP CLI")
logger.info("✅ Successfully uploaded the files!")
except OSError as err:
if "nothing to commit, working tree clean" in err.args[0]:
logger.info("❔ Files did not change since last upload!")
dataset_repo.git_push()
return
logger.error("❌ Something went wrong when uploading the files!")
raise
for idx, file_path in enumerate(filepaths):
file_name = os.path.basename(file_path)
logger.info(f"[{idx + 1}/{len(filepaths)}] 📁 Registering file {file_name} into project '{file_name}'...")
payload = {
"split": split,
"col_mapping": col_mapping,
"data_files": [{"fname": file_name, "username": self.user}],
}
http_post(path=f"/projects/{self.proj_id}/data/add", payload=payload, token=self._token)
logger.info(f"[{idx + 1}/{len(filepaths)}] ✅ Success!")
def save_to_hub(self,
repo_name: str,
organization: Optional[str] = None,
private: Optional[bool] = None,
commit_message: str = "Add new SentenceTransformer model.",
local_model_path: Optional[str] = None,
exist_ok: bool = False,
replace_model_card: bool = False):
"""
Uploads all elements of this Sentence Transformer to a new HuggingFace Hub repository.
:param repo_name: Repository name for your model in the Hub.
:param organization: Organization in which you want to push your model or tokenizer (you must be a member of this organization).
:param private: Set to true, for hosting a prive model
:param commit_message: Message to commit while pushing.
:param local_model_path: Path of the model locally. If set, this file path will be uploaded. Otherwise, the current model will be uploaded
:param exist_ok: If true, saving to an existing repository is OK. If false, saving only to a new repository is possible
:param replace_model_card: If true, replace an existing model card in the hub with the automatically created model card
:return: The url of the commit of your model in the given repository.
"""
token = HfFolder.get_token()
if token is None:
raise ValueError(
"You must login to the Hugging Face hub on this computer by typing `transformers-cli login`."
)
if '/' in repo_name:
splits = repo_name.split('/', maxsplit=1)
if organization is None or organization == splits[0]:
organization = splits[0]
repo_name = splits[1]
else:
raise ValueError(
"You passed and invalid repository name: {}.".format(
repo_name))
endpoint = "https://huggingface.co"
repo_url = HfApi(endpoint=endpoint).create_repo(
token,
repo_name,
organization=organization,
private=private,
repo_type=None,
exist_ok=exist_ok,
)
full_model_name = repo_url[len(endpoint) + 1:].strip("/")
with tempfile.TemporaryDirectory() as tmp_dir:
# First create the repo (and clone its content if it's nonempty).
logging.info("Create repository and clone it if it exists")
repo = Repository(tmp_dir, clone_from=repo_url)
# If user provides local files, copy them.
if local_model_path:
copy_tree(local_model_path, tmp_dir)
else: # Else, save model directly into local repo.
create_model_card = replace_model_card or not os.path.exists(
os.path.join(tmp_dir, 'README.md'))
self.save(tmp_dir,
model_name=full_model_name,
create_model_card=create_model_card)
#Find files larger 5M and track with git-lfs
large_files = []
for root, dirs, files in os.walk(tmp_dir):
for filename in files:
file_path = os.path.join(root, filename)
rel_path = os.path.relpath(file_path, tmp_dir)
if os.path.getsize(file_path) > (5 * 1024 * 1024):
large_files.append(rel_path)
if len(large_files) > 0:
logging.info("Track files with git lfs: {}".format(
", ".join(large_files)))
repo.lfs_track(large_files)
logging.info("Push model to the hub. This might take a while")
push_return = repo.push_to_hub(commit_message=commit_message)
def on_rm_error(func, path, exc_info):
# path contains the path of the file that couldn't be removed
# let's just assume that it's read-only and unlink it.
try:
os.chmod(path, stat.S_IWRITE)
os.unlink(path)
except:
pass
# Remove .git folder. On Windows, the .git folder might be read-only and cannot be deleted
# Hence, try to set write permissions on error
try:
for f in os.listdir(tmp_dir):
shutil.rmtree(os.path.join(tmp_dir, f),
onerror=on_rm_error)
except Exception as e:
logging.warning("Error when deleting temp folder: {}".format(
str(e)))
pass
return push_return
def push_to_hf(
repo_name: str,
serialization_dir: Optional[Union[str, PathLike]] = None,
archive_path: Optional[Union[str, PathLike]] = None,
organization: Optional[str] = None,
commit_message: str = "Update repository",
local_repo_path: Union[str, PathLike] = "hub",
use_auth_token: Union[bool, str] = True,
) -> str:
"""Pushes model and related files to the Hugging Face Hub ([hf.co](https://hf.co/))
# Parameters
repo_name: `str`
Name of the repository in the Hugging Face Hub.
serialization_dir : `Union[str, PathLike]`, optional (default = `None`)
Full path to a directory with the serialized model.
archive_path : `Union[str, PathLike]`, optional (default = `None`)
Full path to the zipped model (e.g. model/model.tar.gz). Use `serialization_dir` if possible.
organization : `Optional[str]`, optional (default = `None`)
Name of organization to which the model should be uploaded.
commit_message: `str` (default=`Update repository`)
Commit message to use for the push.
local_repo_path : `Union[str, Path]`, optional (default=`hub`)
Local directory where the repository will be saved.
use_auth_token (``str`` or ``bool``, `optional`, defaults ``True``):
huggingface_token can be extract from ``HfApi().login(username, password)`` and is used to authenticate
against the Hugging Face Hub (useful from Google Colab for instance). It's automatically retrieved
if you've done `huggingface-cli login` before.
"""
if serialization_dir is not None:
working_dir = Path(serialization_dir)
if archive_path is not None:
raise ValueError(
"serialization_dir and archive_path are mutually exclusive, please just use one."
)
if not working_dir.exists() or not working_dir.is_dir():
raise ValueError(
f"Can't find path: {serialization_dir}, please point"
"to a directory with the serialized model.")
elif archive_path is not None:
working_dir = Path(archive_path)
if (not working_dir.exists() or not zipfile.is_zipfile(working_dir)
and not tarfile.is_tarfile(working_dir)):
raise ValueError(
f"Can't find path: {archive_path}, please point to a .tar.gz archive"
"or to a directory with the serialized model.")
else:
logging.info(
"Using the archive_path is discouraged. Using the serialization_dir"
"will also upload metrics and TensorBoard traces to the Hugging Face Hub."
)
else:
raise ValueError(
"please specify either serialization_dir or archive_path")
info_msg = f"Preparing repository '{use_auth_token}'"
if isinstance(use_auth_token, str):
huggingface_token = use_auth_token
elif use_auth_token:
huggingface_token = HfFolder.get_token()
# Create the repo (or clone its content if it's nonempty)
api = HfApi()
repo_url = api.create_repo(
name=repo_name,
token=huggingface_token,
organization=organization,
private=False,
exist_ok=True,
)
repo_local_path = Path(local_repo_path) / repo_name
repo = Repository(repo_local_path,
clone_from=repo_url,
use_auth_token=use_auth_token)
repo.git_pull(rebase=True)
# Model file should be tracked with Git LFS
repo.lfs_track(["*.th"])
info_msg = f"Preparing repository '{repo_name}'"
if organization is not None:
info_msg += f" ({organization})"
logging.info(info_msg)
# Extract information from either serializable directory or a
# .tar.gz file
if serialization_dir is not None:
for filename in working_dir.iterdir():
_copy_allowed_file(Path(filename), repo_local_path)
else:
with tempfile.TemporaryDirectory() as temp_dir:
extracted_dir = Path(
cached_path(working_dir, temp_dir, extract_archive=True))
for filename in extracted_dir.iterdir():
_copy_allowed_file(Path(filename), repo_local_path)
_create_model_card(repo_local_path)
logging.info(f"Pushing repo {repo_name} to the Hugging Face Hub")
repo.push_to_hub(commit_message=commit_message)
logging.info(f"View your model in {repo_url}")
return repo_url
def push_to_huggingface_hub(self):
"""Creates a downstream repository on the Hub and pushes training artifacts to it."""
if self.args.hf_hub_org.lower() != "none":
organization = self.args.hf_hub_org
else:
organization = os.environ.get("HF_USERNAME")
huggingface_token = HfFolder.get_token()
print(f"[Runner] - Organisation to push fine-tuned model to: {organization}")
# Extract upstream repository metadata
if self.args.hub == "huggingface":
model_info = HfApi().model_info(self.args.upstream, token=huggingface_token)
downstream_model_id = model_info.sha
# Exclude "/" characters from downstream repo ID
upstream_model_id = model_info.modelId.replace("/", "__")
else:
upstream_model_id = self.args.upstream.replace("/", "__")
downstream_model_id = str(uuid.uuid4())[:8]
repo_name = f"{upstream_model_id}__{downstream_model_id}"
# Create downstream repo on the Hub
repo_url = HfApi().create_repo(
token=huggingface_token,
name=repo_name,
organization=organization,
exist_ok=True,
private=False,
)
print(f"[Runner] - Created Hub repo: {repo_url}")
# Download repo
HF_HUB_DIR = "hf_hub"
REPO_ROOT_DIR = os.path.join(self.args.expdir, HF_HUB_DIR, repo_name)
REPO_TASK_DIR = os.path.join(REPO_ROOT_DIR, self.args.downstream, self.args.expname)
print(f"[Runner] - Cloning Hub repo to {REPO_ROOT_DIR}")
model_repo = Repository(
local_dir=REPO_ROOT_DIR, clone_from=repo_url, use_auth_token=huggingface_token
)
# Pull latest changes if they exist
model_repo.git_pull()
# Copy checkpoints, tensorboard logs, and args / configs
# Note that this copies all files from the experiment directory,
# including those from multiple runs
shutil.copytree(self.args.expdir, REPO_TASK_DIR, dirs_exist_ok=True, ignore=shutil.ignore_patterns(HF_HUB_DIR))
# By default we use model.ckpt in the PreTrainedModel interface, so
# rename the best checkpoint to match this convention
checkpoints = list(Path(REPO_TASK_DIR).glob("*best*.ckpt"))
if len(checkpoints) == 0:
print("[Runner] - Did not find a best checkpoint! Using the final checkpoint instead ...")
CKPT_PATH = (
os.path.join(REPO_TASK_DIR, f"states-{self.config['runner']['total_steps']}.ckpt")
)
elif len(checkpoints) > 1:
print(f"[Runner] - More than one best checkpoint found! Using {checkpoints[0]} as default ...")
CKPT_PATH = checkpoints[0]
else:
print(f"[Runner] - Found best checkpoint {checkpoints[0]}!")
CKPT_PATH = checkpoints[0]
shutil.move(CKPT_PATH, os.path.join(REPO_TASK_DIR, "model.ckpt"))
model_repo.lfs_track("*.ckpt")
# Write model card
self._create_model_card(REPO_ROOT_DIR)
# Push everything to the Hub
print("[Runner] - Pushing model files to the Hub ...")
model_repo.push_to_hub()
print("[Runner] - Training run complete!")
