app.py

# Gradio YOLOv5 Det v0.3
# author: Zeng Yifu（曾逸夫）
# creation time: 2022-05-09
# email: zyfiy1314@163.com
# project homepage: https://gitee.com/CV_Lab/gradio_yolov5_det

# import os

# os.system("pip install gradio==3.0.1")

import argparse
import csv
import json
import sys
from collections import Counter
from pathlib import Path
import pandas as pd

import gradio as gr
import torch
import yaml
from PIL import Image, ImageDraw, ImageFont

from util.fonts_opt import is_fonts
from util.pdf_opt import pdf_generate

ROOT_PATH = sys.path[0] # root directory

# model path
model_path = "ultralytics/yolov5"

# Gradio YOLOv5 Det version
GYD_VERSION = "Gradio YOLOv5 Det v0.3"

# model name temporary variable
model_name_tmp = ""

# Device temporary variables
device_tmp = ""

# File extension
suffix_list = [".csv", ".yaml"]

# font size
FONTSIZE = 25

# object style
obj_style = ["Small Object", "Medium Object", "Large Object"]


def parse_args(known=False):
    parser = argparse.ArgumentParser(description="Gradio YOLOv5 Det v0.3")
    parser.add_argument("--source", "-src", default="upload", type=str, help="input source")
    parser.add_argument("--img_tool", "-it", default="editor", type=str, help="input image tool")
    parser.add_argument("--model_name", "-mn", default="yolov5s", type=str, help="model name")
    parser.add_argument(
        "--model_cfg",
        "-mc",
        default="./model_config/model_name_p5_p6_all.yaml",
        type=str,
        help="model config",
    )
    parser.add_argument(
        "--cls_name",
        "-cls",
        default="./cls_name/cls_name_en.yaml",
        type=str,
        help="cls name",
    )
    parser.add_argument(
        "--nms_conf",
        "-conf",
        default=0.5,
        type=float,
        help="model NMS confidence threshold",
    )
    parser.add_argument("--nms_iou", "-iou", default=0.45, type=float, help="model NMS IoU threshold")
    parser.add_argument(
        "--device",
        "-dev",
        default="cpu",
        type=str,
        help="cuda or cpu",
    )
    parser.add_argument("--inference_size", "-isz", default=640, type=int, help="model inference size")
    parser.add_argument("--max_detnum", "-mdn", default="50", type=str, help="model max det num")

    args = parser.parse_known_args()[0] if known else parser.parse_args()
    return args


# yaml file parsing
def yaml_parse(file_path):
    return yaml.safe_load(open(file_path, encoding="utf-8").read())


# yaml csv file parsing
def yaml_csv(file_path, file_tag):
    file_suffix = Path(file_path).suffix
    if file_suffix == suffix_list[0]:
        # model name
        file_names = [i[0] for i in list(csv.reader(open(file_path)))] # csv version
    elif file_suffix == suffix_list[1]:
        # model name
        file_names = yaml_parse(file_path).get(file_tag) # yaml version
    else:
        print(f"{file_path} is not in the correct format! Program exits!")
        sys.exit()

    return file_names


# model loading
def model_loading(model_name, device):

    # load model
    model = torch.hub.load(
        model_path, model_name, force_reload=True, device=device, _verbose=False
    )

    return model


# check information
def export_json(results, model, img_size):

    return [
        [
            {
                "id": i,
                "class": int(result[i][5]),
                # "class_name": model.model.names[int(result[i][5])],
                "class_name": model_cls_name_cp[int(result[i][5])],
                "normalized_box": {
                    "x0": round(result[i][:4].tolist()[0], 6),
                    "y0": round(result[i][:4].tolist()[1], 6),
                    "x1": round(result[i][:4].tolist()[2], 6),
                    "y1": round(result[i][:4].tolist()[3], 6),},
                "confidence": round(float(result[i][4]), 2),
                "fps": round(1000 / float(results.t[1]), 2),
                "width": img_size[0],
                "height": img_size[1],} for i in range(len(result))] for result in results.xyxyn]


# frame conversion
def pil_draw(img, countdown_msg, textFont, xyxy, font_size, opt):

    img_pil = ImageDraw.Draw(img)

    img_pil.rectangle(xyxy, fill=None, outline="green") # bounding box

    if "label" in opt:
        text_w, text_h = textFont.getsize(countdown_msg) # Label size
        img_pil.rectangle(
            (xyxy[0], xyxy[1], xyxy[0] + text_w, xyxy[1] + text_h),
            fill="green",
            outline="green",
        ) # label background
        img_pil.multiline_text(
            (xyxy[0], xyxy[1]),
            countdown_msg,
            fill=(205, 250, 255),
            font=textFont,
            align="center",
        )

    return img


# YOLOv5 image detection function
def yolo_det(img, device, model_name, inference_size, conf, iou, max_num, model_cls, opt):

    global model, model_name_tmp, device_tmp

    # object size num
    s_obj, m_obj, l_obj = 0, 0, 0
    # object area list
    area_obj_all = []
    # cls num stat
    cls_det_stat = []

    if model_name_tmp != model_name:
        # Model judgment to avoid repeated loading
        model_name_tmp = model_name
        model = model_loading(model_name_tmp, device)
    elif device_tmp != device:
        device_tmp = device
        model = model_loading(model_name_tmp, device)

    # -------------Model tuning -------------
    model.conf = conf # NMS confidence threshold
    model.iou = iou # NMS IoU threshold
    model.max_det = int(max_num) # Maximum number of detection frames
    model.classes = model_cls # model classes
    
    img_size = img.size # frame size

    results = model(img, size=inference_size) # detection
    
    # Data Frame
    dataframe = results.pandas().xyxy[0].round(2)

    # ----------------Load fonts----------------
    yaml_index = cls_name.index(".yaml")
    cls_name_lang = cls_name[yaml_index - 2:yaml_index]

    if cls_name_lang == "zh":
        # Chinese
        textFont = ImageFont.truetype(str(f"{ROOT_PATH}/fonts/SimSun.ttf"), size=FONTSIZE)
    elif cls_name_lang in ["en", "ru", "es", "ar"]:
        # English, Russian, Spanish, Arabic
        textFont = ImageFont.truetype(str(f"{ROOT_PATH}/fonts/TimesNewRoman.ttf"), size=FONTSIZE)
    elif cls_name_lang == "ko":
        # Korean
        textFont = ImageFont.truetype(str(f"{ROOT_PATH}/fonts/malgun.ttf"), size=FONTSIZE)

    for result in results.xyxyn:
        for i in range(len(result)):
            id = int(i) # instance ID
            obj_cls_index = int(result[i][5]) # category index
            obj_cls = model_cls_name_cp[obj_cls_index] # category
            cls_det_stat.append(obj_cls)

            # ------------ border coordinates ------------
            x0 = float(result[i][:4].tolist()[0])
            y0 = float(result[i][:4].tolist()[1])
            x1 = float(result[i][:4].tolist()[2])
            y1 = float(result[i][:4].tolist()[3])

            # ------------ Actual coordinates of the border ------------
            x0 = int(img_size[0] * x0)
            y0 = int(img_size[1] * y0)
            x1 = int(img_size[0] * x1)
            y1 = int(img_size[1] * y1)

            conf = float(result[i][4]) # confidence
            # fps = f"{(1000 / float(results.t[1])):.2f}" # FPS

            det_img = pil_draw(
                img,
                f"{id}-{obj_cls}:{conf:.2f}",
                textFont,
                [x0, y0, x1, y1],
                FONTSIZE,
                opt,
            )

            # ----------add object size----------
            w_obj = x1 - x0
            h_obj = y1 - y0
            area_obj = w_obj * h_obj
            area_obj_all.append(area_obj)

    det_json = export_json(results, model, img.size)[0] # Detection information
    det_json_format = json.dumps(det_json, sort_keys=False, indent=4, separators=(",", ":"), ensure_ascii=False) # JSON formatting
    if "json" not in opt:
        det_json = None

    # -------pdf-------
    report = "./Det_Report.pdf"
    if "pdf" in opt:
        pdf_generate(f"{det_json_format}", report, GYD_VERSION)
    else:
        report = None

    # --------------object size compute--------------
    for i in range(len(area_obj_all)):
        if (0 < area_obj_all[i] <= 32 ** 2):
            s_obj = s_obj + 1
        elif (32 ** 2 < area_obj_all[i] <= 96 ** 2):
            m_obj = m_obj + 1
        elif (area_obj_all[i] > 96 ** 2):
            l_obj = l_obj + 1

    sml_obj_total = s_obj + m_obj + l_obj

    objSize_dict = {obj_style[i]: [s_obj, m_obj, l_obj][i] / sml_obj_total for i in range(3)}

    # ------------cls stat------------
    clsRatio_dict = {}
    clsDet_dict = Counter(cls_det_stat)
    clsDet_dict_sum = sum(clsDet_dict.values())

    for k, v in clsDet_dict.items():
        clsRatio_dict[k] = v / clsDet_dict_sum


    return det_img, objSize_dict, clsRatio_dict, det_json, report, dataframe


def main(args):
    gr.close_all()

    global model, model_cls_name_cp, cls_name

    slider_step = 0.05 # sliding step

    source = args.source
    img_tool = args.img_tool
    nms_conf = args.nms_conf
    nms_iou = args.nms_iou
    model_name = args.model_name
    model_cfg = args.model_cfg
    cls_name = args.cls_name
    device = args.device
    inference_size = args.inference_size
    max_detnum = args.max_detnum

    is_fonts(f"{ROOT_PATH}/fonts") # Check font files

    # model loading
    model = model_loading(model_name, device)

    model_names = yaml_csv(model_cfg, "model_names") # model names
    model_cls_name = yaml_csv(cls_name, "model_cls_name") # class name

    model_cls_name_cp = model_cls_name.copy() # class name

    # ------------------- Input Components -------------------
    inputs_img = gr.Image(image_mode="RGB", source=source, tool=img_tool, type="pil", label="original image")
    inputs_device = gr.Radio(choices=["cuda:0", "cpu"], value=device, label="device")
    inputs_model = gr.Dropdown(choices=model_names, value=model_name, type="value", label="model")
    inputs_size = gr.Radio(choices=[320, 640, 1280], value=inference_size, label="inference size")
    input_conf = gr.Slider(0, 1, step=slider_step, value=nms_conf, label="confidence threshold")
    inputs_iou = gr.Slider(0, 1, step=slider_step, value=nms_iou, label="IoU threshold")
    inputs_maxnum = gr.Textbox(lines=1, placeholder="Maximum number of detections", value=max_detnum, label="Maximum number of detections")
    inputs_clsName = gr.CheckboxGroup(choices=model_cls_name, value=model_cls_name, type="index", label="category")
    inputs_opt = gr.CheckboxGroup(choices=["label", "pdf", "json"],
                                         value=["label", "pdf"],
                                         type="value",
                                         label="operate")

    # Input parameters
    inputs = [
        inputs_img, # input image
        inputs_device, # device
        inputs_model, # model
        inputs_size, # inference size
        input_conf, # confidence threshold
        inputs_iou, # IoU threshold
        inputs_maxnum, # maximum number of detections
        inputs_clsName, # category
        inputs_opt, # detect operations
    ]

    # Output parameters
    outputs_img = gr.Image(type="pil", label="Detection image")
    outputs_json = gr.JSON(label="Detection information")
    outputs_pdf = gr.File(label="Download test report")
    outputs_df = gr.Dataframe(max_rows=5, overflow_row_behaviour="paginate", type="pandas", label="List of detection information")
    outputs_objSize = gr.Label(label="Object size ratio statistics")
    outputs_clsSize = gr.Label(label="Category detection proportion statistics")

    outputs = [outputs_img, outputs_objSize, outputs_clsSize, outputs_json, outputs_pdf, outputs_df]

    # title
    title = "Gradio YOLOv5 Det v0.3"

    # describe
    description = "<div align='center'>Customizable target detection model, easy to install, easy to use</div>"

    # example image
    examples = [
        [
            "./img_example/bus.jpg",
            "cpu",
            "yolov5s",
            640,
            0.6,
            0.5,
            10,
            ["person", "bus"],
            ["label", "pdf"],],
        [
            "./img_example/giraffe.jpg",
            "cpu",
            "yolov5l",
            320,
            0.5,
            0.45,
            12,
            ["giraffe"],
            ["label", "pdf"],],
        [
            "./img_example/zidane.jpg",
            "cpu",
            "yolov5m",
            640,
            0.25,
            0.5,
            15,
            ["person", "tie"],
            ["pdf", "json"],],
        [
            "./img_example/Millenial-at-work.jpg",
            "cpu",
            "yolov5s6",
            1280,
            0.5,
            0.5,
            20,
            ["person", "chair", "cup", "laptop"],
            ["label", "pdf"],],]

    # interface
    gr.Interface(
        fn=yolo_det,
        inputs=inputs,
        outputs=outputs,
        title=title,
        description=description,
        article="",
        # examples=examples,
        # theme="seafoam",
        # flagging_dir="run", # output directory
    ).launch(
        inbrowser=True, # Automatically open default browser
        show_tips=True, # Automatically display the latest features of gradio
    )


if __name__ == "__main__":
    args = parse_args()
    main(args)