app.py

import argparse
import logging
import os
import wandb
import gradio as gr

import zipfile
import pickle
from pathlib import Path
from typing import List, Any, Dict
from PIL import Image
from pathlib import Path

from transformers import AutoTokenizer
from sentence_transformers import SentenceTransformer, util
from multilingual_clip import pt_multilingual_clip
import torch

from pathlib import Path
from typing import Callable, Dict, List, Tuple
from PIL.Image import Image

print(__file__)

os.environ["CUDA_VISIBLE_DEVICES"] = ""  # do not use GPU

logging.basicConfig(level=logging.INFO)
DEFAULT_APPLICATION_NAME = "FashGen"

APP_DIR = Path(__file__).resolve().parent  # what is the directory for this application?
README = APP_DIR / "README.md"  # path to an app readme file in HTML/markdown

DEFAULT_PORT = 11700

EMBEDDINGS_DIR = "artifacts/img-embeddings"
EMBEDDINGS_FILE = os.path.join(EMBEDDINGS_DIR, "embeddings.pkl")
RAW_PHOTOS_DIR = "artifacts/raw-photos"

# Download image embeddings and raw photos
wandb.login(key="4b5a23a662b20fdd61f2aeb5032cf56fdce278a4")  # os.getenv('wandb')
api = wandb.Api()
artifact_embeddings = api.artifact("ryparmar/fashion-aggregator/unimoda-images:v1")
artifact_embeddings.download(EMBEDDINGS_DIR)
artifact_raw_photos = api.artifact("ryparmar/fashion-aggregator/unimoda-raw-images:v1")
artifact_raw_photos.download("artifacts")

with zipfile.ZipFile("artifacts/unimoda.zip", 'r') as zip_ref:
    zip_ref.extractall(RAW_PHOTOS_DIR)


class TextEncoder:
    """Encodes the given text"""

    def __init__(self, model_path="M-CLIP/XLM-Roberta-Large-Vit-B-32"):
        self.model = pt_multilingual_clip.MultilingualCLIP.from_pretrained(model_path)
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)

    @torch.no_grad()
    def encode(self, query: str) -> torch.Tensor:
        """Predict/infer text embedding for a given query."""
        query_emb = self.model.forward([query], self.tokenizer)
        return query_emb


class ImageEnoder:
    """Encodes the given image"""

    def __init__(self, model_path="clip-ViT-B-32"):
        self.model = SentenceTransformer(model_path)

    @torch.no_grad()
    def encode(self, image: Image) -> torch.Tensor:
        """Predict/infer text embedding for a given query."""
        image_emb = self.model.encode([image], convert_to_tensor=True, show_progress_bar=False)
        return image_emb


class Retriever:
    """Retrieves relevant images for a given text embedding."""

    def __init__(self, image_embeddings_path=None):
        self.text_encoder = TextEncoder()
        self.image_encoder = ImageEnoder()

        with open(image_embeddings_path, "rb") as file:
            self.image_names, self.image_embeddings = pickle.load(file)
            self.image_names = [
                img_name.replace("fashion-aggregator/fashion_aggregator/data/photos/", "")
                for img_name in self.image_names
            ]
        print("Images:", len(self.image_names))

    @torch.no_grad()
    def predict(self, text_query: str, k: int = 10) -> List[Any]:
        """Return top-k relevant items for a given embedding"""
        query_emb = self.text_encoder.encode(text_query)
        relevant_images = util.semantic_search(query_emb, self.image_embeddings, top_k=k)[0]
        return relevant_images

    @torch.no_grad()
    def search_images(self, text_query: str, k: int = 6) -> Dict[str, List[Any]]:
        """Return top-k relevant images for a given embedding"""
        images = self.predict(text_query, k)
        paths_and_scores = {"path": [], "score": []}
        for img in images:
            paths_and_scores["path"].append(os.path.join(RAW_PHOTOS_DIR, self.image_names[img["corpus_id"]]))
            paths_and_scores["score"].append(img["score"])
        return paths_and_scores


def main(args):
    predictor = PredictorBackend(url=args.model_url)
    frontend = make_frontend(predictor.run, flagging=args.flagging, gantry=args.gantry, app_name=args.application)
    frontend.launch(
        # server_name="0.0.0.0",  # make server accessible, binding all interfaces  # noqa: S104
        # server_port=args.port,  # set a port to bind to, failing if unavailable
        # share=False,  # should we create a (temporary) public link on https://gradio.app?
    )


def make_frontend(
    fn: Callable[[Image], str], flagging: bool = False, gantry: bool = False, app_name: str = "fashion-aggregator"
):
    """Creates a gradio.Interface frontend for text to image search function."""

    allow_flagging = "never"

    # build a basic browser interface to a Python function
    frontend = gr.Interface(
        fn=fn,  # which Python function are we interacting with?
        outputs=gr.Gallery(label="Relevant Items"),
        # what input widgets does it need? we configure an image widget
        inputs=gr.components.Textbox(label="Item Description"),
        title="FashGen",  # what should we display at the top of the page?
        description=__doc__,  # what should we display just above the interface?
        cache_examples=False,  # should we cache those inputs for faster inference? slows down start
        allow_flagging=allow_flagging,  # should we show users the option to "flag" outputs?
        flagging_options=["incorrect", "offensive", "other"],  # what options do users have for feedback?
    )
    return frontend


class PredictorBackend:
    """Interface to a backend that serves predictions.
    To communicate with a backend accessible via a URL, provide the url kwarg.
    Otherwise, runs a predictor locally.
    """

    def __init__(self, url=None):
        if url is not None:
            self.url = url
            self._predict = self._predict_from_endpoint
        else:
            model = Retriever(image_embeddings_path=EMBEDDINGS_FILE)
            self._predict = model.predict
            self._search_images = model.search_images

    def run(self, text: str):
        pred, metrics = self._predict_with_metrics(text)
        self._log_inference(pred, metrics)
        return pred

    def _predict_with_metrics(self, text: str) -> Tuple[List[str], Dict[str, float]]:
        paths_and_scores = self._search_images(text)
        metrics = {"mean_score": sum(paths_and_scores["score"]) / len(paths_and_scores["score"])}
        return paths_and_scores["path"], metrics

    def _log_inference(self, pred, metrics):
        for key, value in metrics.items():
            logging.info(f"METRIC {key} {value}")
        logging.info(f"PRED >begin\n{pred}\nPRED >end")


def _make_parser():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--model_url",
        default=None,
        type=str,
        help="Identifies a URL to which to send image data. Data is base64-encoded, converted to a utf-8 string, and then set via a POST request as JSON with the key 'image'. Default is None, which instead sends the data to a model running locally.",
    )
    parser.add_argument(
        "--port",
        default=DEFAULT_PORT,
        type=int,
        help=f"Port on which to expose this server. Default is {DEFAULT_PORT}.",
    )
    parser.add_argument(
        "--flagging",
        action="store_true",
        help="Pass this flag to allow users to 'flag' model behavior and provide feedback.",
    )
    parser.add_argument(
        "--gantry",
        action="store_true",
        help="Pass --flagging and this flag to log user feedback to Gantry. Requires GANTRY_API_KEY to be defined as an environment variable.",
    )
    parser.add_argument(
        "--application",
        default=DEFAULT_APPLICATION_NAME,
        type=str,
        help=f"Name of the Gantry application to which feedback should be logged, if --gantry and --flagging are passed. Default is {DEFAULT_APPLICATION_NAME}.",
    )
    return parser


if __name__ == "__main__":
    parser = _make_parser()
    args = parser.parse_args()
    main(args)