app.py

from transformers import AutoFeatureExtractor, YolosForObjectDetection
import gradio as gr
from PIL import Image
import torch
import matplotlib.pyplot as plt
import io
import numpy as np
import os
os.system("pip -qq install yoloxdetect==0.0.7")
from yoloxdetect import YoloxDetector

# Images
torch.hub.download_url_to_file('https://tochkanews.ru/wp-content/uploads/2020/09/0.jpg', '1.jpg')
torch.hub.download_url_to_file('https://s.rdrom.ru/1/pubs/4/35893/1906770.jpg', '2.jpg')
torch.hub.download_url_to_file('https://static.mk.ru/upload/entities/2022/04/17/07/articles/detailPicture/5b/39/28/b6/ffb1aa636dd62c30e6ff670f84474f75.jpg', '3.jpg')

def yolox_inference(
    image_path: gr.inputs.Image = None,
    model_path: gr.inputs.Dropdown = 'kadirnar/yolox_s-v0.1.1',
    config_path: gr.inputs.Textbox = 'configs.yolox_s',
    image_size: gr.inputs.Slider = 640
):
    """
    YOLOX inference function
    Args:
        image: Input image
        model_path: Path to the model
        config_path: Path to the config file
        image_size: Image size
    Returns:
        Rendered image
    """

    model = YoloxDetector(model_path, config_path=config_path, device="cpu", hf_model=True)
    pred = model.predict(image_path=image_path, image_size=image_size)
    return pred
        

inputs = [
    gr.inputs.Image(type="filepath", label="Input Image"),
    gr.inputs.Dropdown(
        label="Model Path",
        choices=[
            "kadirnar/yolox_s-v0.1.1",
            "kadirnar/yolox_m-v0.1.1",
            "kadirnar/yolox_tiny-v0.1.1",
        ],
        default="kadirnar/yolox_s-v0.1.1",
    ),
    gr.inputs.Dropdown(
        label="Config Path",
        choices=[
            "configs.yolox_s",
            "configs.yolox_m",
            "configs.yolox_tiny",
        ],
        default="configs.yolox_s",
    ),
    gr.inputs.Slider(minimum=320, maximum=1280, default=640, step=32, label="Image Size"),
]

outputs = gr.outputs.Image(type="filepath", label="Output Image")
title = "YOLOX is a high-performance anchor-free YOLO."

examples = [
    ["1.jpg", "kadirnar/yolox_m-v0.1.1", "configs.yolox_m", 640],
    ["2.jpg", "kadirnar/yolox_s-v0.1.1", "configs.yolox_s", 640],
    ["3.jpg", "kadirnar/yolox_tiny-v0.1.1", "configs.yolox_tiny", 640],
]

demo_app = gr.Interface(
    fn=yolox_inference,
    inputs=inputs,
    outputs=outputs,
    title=title,
    examples=examples,
    cache_examples=True,
    theme='huggingface',
)
demo_app.launch(debug=True, enable_queue=True)


COLORS = [[0.000, 0.447, 0.741], [0.850, 0.325, 0.098], [0.929, 0.694, 0.125],
          [0.494, 0.184, 0.556], [0.466, 0.674, 0.188], [0.301, 0.745, 0.933]]


def get_class_list_from_input(classes_string: str):
    if classes_string == "":
        return []
    classes_list = classes_string.split(",")
    classes_list = [x.strip() for x in classes_list]
    return classes_list

def infer(img, model_name: str, prob_threshold: int, classes_to_show = str):
    feature_extractor = AutoFeatureExtractor.from_pretrained(f"hustvl/{model_name}")
    model = YolosForObjectDetection.from_pretrained(f"hustvl/{model_name}")

    img = Image.fromarray(img)
    
    pixel_values = feature_extractor(img, return_tensors="pt").pixel_values

    with torch.no_grad():
        outputs = model(pixel_values, output_attentions=True)
    
    probas = outputs.logits.softmax(-1)[0, :, :-1]
    keep = probas.max(-1).values > prob_threshold

    target_sizes = torch.tensor(img.size[::-1]).unsqueeze(0)
    postprocessed_outputs = feature_extractor.post_process(outputs, target_sizes)
    bboxes_scaled = postprocessed_outputs[0]['boxes']
    
    classes_list = get_class_list_from_input(classes_to_show)
    res_img = plot_results(img, probas[keep], bboxes_scaled[keep], model, classes_list)
    
    return res_img

def plot_results(pil_img, prob, boxes, model, classes_list):
    plt.figure(figsize=(16,10))
    plt.imshow(pil_img)
    ax = plt.gca()
    colors = COLORS * 100
    for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors):
        cl = p.argmax()
        object_class = model.config.id2label[cl.item()]
        
        if len(classes_list) > 0 :
            if object_class not in classes_list:
                continue
            
        ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
                                fill=False, color=c, linewidth=3))
        text = f'{object_class}: {p[cl]:0.2f}'
        ax.text(xmin, ymin, text, fontsize=15,
                bbox=dict(facecolor='yellow', alpha=0.5))
    plt.axis('off')
    return fig2img(plt.gcf())
    
def fig2img(fig):
    buf = io.BytesIO()
    fig.savefig(buf)
    buf.seek(0)
    img = Image.open(buf)
    return img

image_in = gr.components.Image()
image_out = gr.components.Image()
model_choice = gr.components.Dropdown(["yolos-tiny", "yolos-small", "yolos-base", "yolos-small-300", "yolos-small-dwr"], value="yolos-small", label="YOLOS Model")
prob_threshold_slider = gr.components.Slider(minimum=0, maximum=1.0, step=0.01, value=0.9, label="Probability Threshold")
classes_to_show = gr.components.Textbox(placeholder="e.g. car, truck, traffic light", label="Classes to use (empty means all classes)")

Iface = gr.Interface(
    fn=infer,
    inputs=[image_in,model_choice, prob_threshold_slider, classes_to_show],
    outputs=image_out,
    examples=examples
    title="Object Detection with YOLOS",
    description=description,
).launch()