app.py

import multiprocessing
import random
from datasets import load_dataset
from sentence_transformers import SentenceTransformer
from PIL.Image import Image, ANTIALIAS
import gradio as gr
from faiss import METRIC_INNER_PRODUCT
import requests
import pandas as pd

import backoff
from functools import lru_cache

cpu_count = multiprocessing.cpu_count()

model = SentenceTransformer("clip-ViT-B-16")


def resize_image(image: Image, size: int = 224) -> Image:
    """Resizes an image retaining the aspect ratio."""
    w, h = image.size
    if w == h:
        image = image.resize((size, size), ANTIALIAS)
        return image
    if w > h:
        height_percent = size / float(h)
        width_size = int(float(w) * float(height_percent))
        image = image.resize((width_size, size), ANTIALIAS)
        return image
    if w < h:
        width_percent = size / float(w)
        height_size = int(float(w) * float(width_percent))
        image = image.resize((size, height_size), ANTIALIAS)
        return image


dataset = load_dataset("davanstrien/ia-loaded-embedded-gpu", split="train")
dataset = dataset.filter(lambda x: x["embedding"] is not None)
dataset.add_faiss_index("embedding", metric_type=METRIC_INNER_PRODUCT)


def get_nearest_k_examples(input, k):
    query = model.encode(input)
    # faiss_index = dataset.get_index("embedding").faiss_index # TODO maybe add range?
    # threshold = 0.95
    # limits, distances, indices = faiss_index.range_search(x=query, thresh=threshold)
    # images = dataset[indices]
    _, retrieved_examples = dataset.get_nearest_examples("embedding", query=query, k=k)
    images = retrieved_examples["image"][:k]
    last_modified = retrieved_examples["last_modified_date"]  # [:k]
    crawl_date = retrieved_examples["crawl_date"]  # [:k]
    metadata = [
        f"last_modified {modified}, crawl date:{crawl}"
        for modified, crawl in zip(last_modified, crawl_date)
    ]
    return list(zip(images, metadata))


def return_random_sample(k=27):
    sample = random.sample(range(len(dataset)), k)
    images = dataset[sample]["image"]
    return [resize_image(image).convert("RGB") for image in images]


def predict_subset(model_id, token):
    API_URL = f"https://api-inference.huggingface.co/models/{model_id}"
    headers = {"Authorization": f"Bearer {token}"}

    @backoff.on_predicate(backoff.expo, lambda x: x.status_code == 503, max_time=30)
    def _query(url):
        r = requests.post(API_URL, headers=headers, data=url)
        print(r)
        return r

    @lru_cache(maxsize=1000)
    def query(url):
        response = _query(url)
        try:
            data = response.json()
            argmax = data[0]
            return {"score": argmax["score"], "label": argmax["label"]}
        except Exception:
            return {"score": None, "label": None}

    # dataset2 = copy.deepcopy(dataset)
    # dataset2.drop_index("embedding")
    dataset = load_dataset("davanstrien/ia-loaded-embedded-gpu", split="train")
    sample = random.sample(range(len(dataset)), 10)
    sample = dataset.select(sample)
    print("predicting...")
    predictions = []
    for row in sample:
        url = row["url"]
        predictions.append(query(url))
    gallery = []
    for url, prediction in zip(sample["url"], predictions):
        gallery.append((url, f"{prediction['label'], prediction['score']}"))
    # sample = sample.map(lambda x:  query(x['url']))
    labels = [d["label"] for d in predictions]
    from toolz import frequencies

    df = pd.DataFrame(
        {"labels": frequencies(labels).keys(), "freqs": frequencies(labels).values()}
    )
    return gallery, df


with gr.Blocks() as demo:
    with gr.Tab("Random image gallery"):
        button = gr.Button("Refresh")
        gallery = gr.Gallery().style(grid=9, height="1400")
        button.click(return_random_sample, [], [gallery])
    with gr.Tab("image search"):
        text = gr.Textbox(label="Search for images")
        k = gr.Slider(minimum=3, maximum=18, step=1)
        button = gr.Button("search")
        gallery = gr.Gallery().style(grid=3)
        button.click(get_nearest_k_examples, [text, k], [gallery])
    # with gr.Tab("Export for label studio"):
    #     button = gr.Button("Export")
    #     dataset2 = copy.deepcopy(dataset)
    #     # dataset2 = dataset2.remove_columns('image')
    #     # dataset2 = dataset2.rename_column("url", "image")
    #     csv = dataset2.to_csv("label_studio.csv")
    #     csv_file = gr.File("label_studio.csv")
    #     button.click(dataset.save_to_disk, [], [csv_file])
    with gr.Tab("predict"):
        token = gr.Textbox(label="token", type="password")
        model_id = gr.Textbox(label="model_id")
        button = gr.Button("predict")
        plot = gr.BarPlot(x="labels", y="freqs", width=600, height=400, vertical=False)
        gallery = gr.Gallery()
        button.click(predict_subset, [model_id, token], [gallery, plot])

demo.launch(enable_queue=True)