ReviewGuard / reviewguard /deploy.py
Emirinth's picture
Deploy ReviewGuard
db75deb verified
Raw History Blame Contribute Delete
4.54 kB
"""Publish ReviewGuard to Hugging Face: the app as a Space and the trained models as a model repo.
The Space runs the app in the cloud, and the model repo lets anyone skip retraining.
Uploads use the token saved by `hf auth login`.
"""
import shutil
import tempfile
from pathlib import Path
from huggingface_hub import HfApi, snapshot_download
from reviewguard import config
SPACE_ID = "Emirinth/ReviewGuard"
MODEL_REPO_ID = "Emirinth/ReviewGuard-models"
# Training leftovers that are not needed to run the models.
SKIPPED_MODEL_FILES = ["*training_args.bin", "roberta_checkpoints/*"]
# Spaces read their settings from the front matter of README.md.
SPACE_README = """---
title: ReviewGuard
emoji: 🛡️
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: {gradio_version}
python_version: "3.12"
app_file: app.py
pinned: false
short_description: AI triage for retail reviews and customer photos
---
# ReviewGuard
AI triage for retail reviews: sentiment, issue type and photo checks combined into one decision.
"""
# The CPU build of PyTorch is far smaller than the default GPU build.
SPACE_REQUIREMENTS = """--extra-index-url https://download.pytorch.org/whl/cpu
torch
transformers
accelerate
captum
sentence-transformers
textblob
vaderSentiment
joblib
pandas
pyarrow
scikit-learn
matplotlib
plotly
pillow
"""
ARTIFACT_FOLDERS = ["reports", "examples", "labels"]
def stage_space(folder: Path) -> None:
"""Copy everything the Space needs into one folder."""
import gradio
shutil.copy(config.CODE_DIR / "app.py", folder / "app.py")
shutil.copytree(
config.CODE_DIR / "reviewguard",
folder / "reviewguard",
ignore=shutil.ignore_patterns("__pycache__"),
)
(folder / "README.md").write_text(
SPACE_README.format(gradio_version=gradio.__version__), encoding="utf-8"
)
(folder / "requirements.txt").write_text(SPACE_REQUIREMENTS, encoding="utf-8")
artifacts = folder / "artifacts"
for name in ARTIFACT_FOLDERS:
shutil.copytree(config.ARTIFACTS_DIR / name, artifacts / name)
shutil.copytree(
config.MODELS_DIR,
artifacts / "models",
ignore=shutil.ignore_patterns("training_args.bin", "roberta_checkpoints"),
)
def deploy() -> str:
"""Create the Space if needed, upload the staged folder and return the Space URL."""
api = HfApi()
api.create_repo(SPACE_ID, repo_type="space", space_sdk="gradio", exist_ok=True)
with tempfile.TemporaryDirectory() as temp:
stage_space(Path(temp))
api.upload_folder(
folder_path=temp,
repo_id=SPACE_ID,
repo_type="space",
commit_message="Deploy ReviewGuard",
delete_patterns=["*"],
)
return f"https://huggingface.co/spaces/{SPACE_ID}"
# ---------------------------------------------------------------- trained models
MODEL_CARD = """---
license: mit
tags: [sentiment-analysis, retail, reviews, text-classification]
---
# ReviewGuard trained models
Models behind the ReviewGuard review-triage app.
- `roberta_sentiment/`: RoBERTa fine-tuned on 72,000 Amazon Reviews 2023 reviews.
It predicts negative, neutral or positive.
- `distilroberta_sentiment/`: a smaller, faster version trained the same way.
- `classical/`: TF-IDF with Naive Bayes and Logistic Regression baselines.
- `issue_student.joblib`: issue classifier distilled from a zero-shot teacher.
- `photo_calibration.json` and `lexicon_bands.json`: calibration and tuning settings.
Download them into the project with `python pipeline.py fetch-models`.
"""
def publish_models() -> str:
"""Upload the trained models and a short model card, then return the repo URL."""
api = HfApi()
api.create_repo(MODEL_REPO_ID, repo_type="model", exist_ok=True)
api.upload_file(
path_or_fileobj=MODEL_CARD.encode("utf-8"),
path_in_repo="README.md",
repo_id=MODEL_REPO_ID,
repo_type="model",
)
api.upload_folder(
folder_path=str(config.MODELS_DIR),
repo_id=MODEL_REPO_ID,
repo_type="model",
ignore_patterns=SKIPPED_MODEL_FILES,
commit_message="Upload trained ReviewGuard models",
)
return f"https://huggingface.co/{MODEL_REPO_ID}"
def fetch_models() -> str:
"""Download the published models into the local models folder.
Returns the folder they were saved to.
"""
config.MODELS_DIR.mkdir(parents=True, exist_ok=True)
return snapshot_download(MODEL_REPO_ID, repo_type="model", local_dir=str(config.MODELS_DIR))