app.py

# Gradio YOLOv5 Det v0.4
# author: Zeng Yifu（曾逸夫）
# creation time: 2022-05-28
# email: zyfiy1314@163.com
# project homepage: https://gitee.com/CV_Lab/gradio_yolov5_det

import argparse
import csv
import gc
import json
import os
import sys
from collections import Counter
from pathlib import Path

import cv2
import gradio as gr
import numpy as np
import pandas as pd
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 v0.4"

# 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 v0.4")
    parser.add_argument("--source", "-src", default="upload", type=str, help="input source")
    parser.add_argument("--source_video", "-src_v", default="webcam", type=str, help="video 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=float, help="model max det num")
    parser.add_argument("--slider_step", "-ss", default=0.05, type=float, help="slider step")
    parser.add_argument(
        "--is_login",
        "-isl",
        action="store_true",
        default=False,
        help="is login",
    )
    parser.add_argument('--usr_pwd',
                        "-up",
                        nargs='+',
                        type=str,
                        default=["admin", "admin"],
                        help="user & password for login")
    parser.add_argument(
        "--is_share",
        "-is",
        action="store_true",
        default=False,
        help="is login",
    )

    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, img_size):

    return [[{
        "ID": i,
        "CLASS": int(result[i][5]),
        "CLASS_NAME": model_cls_name_cp[int(result[i][5])],
        "BOUNDING_BOX": {
            "XMIN": round(result[i][:4].tolist()[0], 6),
            "YMIN": round(result[i][:4].tolist()[1], 6),
            "XMAX": round(result[i][:4].tolist()[2], 6),
            "YMAX": round(result[i][:4].tolist()[3], 6),},
        "CONF": round(float(result[i][4]), 2),
        "FPS": round(1000 / float(results.t[1]), 2),
        "IMG_WIDTH": img_size[0],
        "IMG_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, obj_cls_index, color_list):

    img_pil = ImageDraw.Draw(img)

    img_pil.rectangle(xyxy, fill=None, outline=color_list[obj_cls_index])  # 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=color_list[obj_cls_index],
            outline=color_list[obj_cls_index],
        )  # label background
        
        img_pil.multiline_text(
            (xyxy[0], xyxy[1]),
            countdown_msg,
            fill=(255, 255, 255),
            font=textFont,
            align="center",
        )

    return img


def color_set(cls_num):
    color_list = []
    for i in range(cls_num):
        color = tuple(np.random.choice(range(256), size=3))
        # color = ["#"+''.join([random.choice('0123456789ABCDEF') for j in range(6)])]
        color_list.append(color)
    
    return color_list


# YOLOv5 image detection function
def yolo_det_img(img, device, model_name, infer_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

    color_list = color_set(len(model_cls_name_cp)) # 设置颜色

    img_size = img.size  # frame size

    results = model(img, size=infer_size)  # detection
    
    # ----------------目标裁剪----------------
    crops = results.crop(save=False)
    img_crops = []
    for i in range(len(crops)):
        img_crops.append(crops[i]["im"][..., ::-1])

    # 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,
                obj_cls_index,
                color_list,
            )

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

    # ------------JSON generate------------
    det_json = export_json(results, 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 generate-------
    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, img_crops, objSize_dict, clsRatio_dict, dataframe, det_json, report


# YOLOv5 video detection function
def yolo_det_video(video, device, model_name, infer_size, conf, iou, max_num, model_cls, opt):

    global model, model_name_tmp, device_tmp

    os.system("""
        if [ -e './output.mp4' ]; then
        rm ./output.mp4
        fi
        """)

    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

    color_list = color_set(len(model_cls_name_cp)) # 设置颜色

    # ----------------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)

    # video->frame
    gc.collect()
    output_video_path = "./output.avi"
    cap = cv2.VideoCapture(video)
    fourcc = cv2.VideoWriter_fourcc(*"I420")  # encoder

    out = cv2.VideoWriter(output_video_path, fourcc, 30.0, (int(cap.get(3)), int(cap.get(4))))
    while cap.isOpened():
        ret, frame = cap.read()
        # Determine empty frame
        if not ret:
            break

        frame2 = frame.copy()
        results = model(frame2, size=infer_size)  # detection
        h, w, _ = frame.shape  # frame size
        img_size = (w, h)  # frame size

        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

                # ------------ 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

                frame = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
                frame = pil_draw(
                    frame,
                    f"{id}-{obj_cls}:{conf:.2f}",
                    textFont,
                    [x0, y0, x1, y1],
                    FONTSIZE,
                    opt,
                    obj_cls_index,
                    color_list,
                )

                frame = cv2.cvtColor(np.asarray(frame), cv2.COLOR_RGB2BGR)

            # frame->video
            out.write(frame)
    out.release()
    cap.release()
    # cv2.destroyAllWindows()

    return output_video_path


def main(args):
    gr.close_all()

    global model, model_cls_name_cp, cls_name

    source = args.source
    source_video = args.source_video
    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
    slider_step = args.slider_step
    is_login = args.is_login
    usr_pwd = args.usr_pwd
    is_share = args.is_share

    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_device01 = gr.Radio(choices=["cuda:0", "cpu"], value=device, label="device")
    inputs_model01 = gr.Dropdown(choices=model_names, value=model_name, type="value", label="model")
    inputs_size01 = gr.Radio(choices=[320, 640, 1280], value=inference_size, label="inference size")
    input_conf01 = gr.Slider(0, 1, step=slider_step, value=nms_conf, label="confidence threshold")
    inputs_iou01 = gr.Slider(0, 1, step=slider_step, value=nms_iou, label="IoU threshold")
    inputs_maxnum01 = gr.Number(value=max_detnum, label="Maximum number of detections")
    inputs_clsName01 = gr.CheckboxGroup(choices=model_cls_name, value=model_cls_name, type="index", label="category")
    inputs_opt01 = gr.CheckboxGroup(choices=["label", "pdf", "json"],
                                    value=["label", "pdf"],
                                    type="value",
                                    label="operate")

    # ------------------- Input Components -------------------
    inputs_video = gr.Video(format="mp4", source=source_video, label="original video")  # webcam
    inputs_device02 = gr.Radio(choices=["cuda:0", "cpu"], value=device, label="device")
    inputs_model02 = gr.Dropdown(choices=model_names, value=model_name, type="value", label="model")
    inputs_size02 = gr.Radio(choices=[320, 640, 1280], value=inference_size, label="inference size")
    input_conf02 = gr.Slider(0, 1, step=slider_step, value=nms_conf, label="confidence threshold")
    inputs_iou02 = gr.Slider(0, 1, step=slider_step, value=nms_iou, label="IoU threshold")
    inputs_maxnum02 = gr.Number(value=max_detnum, label="Maximum number of detections")
    inputs_clsName02 = gr.CheckboxGroup(choices=model_cls_name, value=model_cls_name, type="index", label="category")
    inputs_opt02 = gr.CheckboxGroup(choices=["label"], value=["label"], type="value", label="operate")

    # Input parameters
    inputs_img_list = [
        inputs_img,  # input image
        inputs_device01,  # device
        inputs_model01,  # model
        inputs_size01,  # inference size
        input_conf01,  # confidence threshold
        inputs_iou01,  # IoU threshold
        inputs_maxnum01,  # maximum number of detections
        inputs_clsName01,  # category
        inputs_opt01,  # detect operations
    ]

    inputs_video_list = [
        inputs_video,  # input image
        inputs_device02,  # device
        inputs_model02,  # model
        inputs_size02,  # inference size
        input_conf02,  # confidence threshold
        inputs_iou02,  # IoU threshold
        inputs_maxnum02,  # maximum number of detections
        inputs_clsName02,  # category
        inputs_opt02,  # detect operation
    ]

    # -------------------output component-------------------
    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_crops = gr.Gallery(label="Object crop")

    # -------------------output component-------------------
    outputs_video = gr.Video(format='mp4', label="Detection video")

    # output parameters
    outputs_img_list = [outputs_img, outputs_crops, outputs_objSize, outputs_clsSize, outputs_df, outputs_json, outputs_pdf]
    outputs_video_list = [outputs_video]

    # title
    title = "Gradio YOLOv5 Det v0.4"

    # describe
    description = "<div align='center'>Customizable target detection model, easy to install, easy to use</div>"
    # article="https://gitee.com/CV_Lab/gradio_yolov5_det"

    # 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,
            0.6,
            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
    gyd_img = gr.Interface(
        fn=yolo_det_img,
        inputs=inputs_img_list,
        outputs=outputs_img_list,
        title=title,
        description=description,
        # article=article,
        # examples=examples,
        # theme="seafoam",
        # live=True, # Change output in real time
        flagging_dir="run",  # output directory
        # allow_flagging="manual",
        # flagging_options=["good", "generally", "bad"],
    )

    gyd_video = gr.Interface(
        # fn=yolo_det_video_test,
        fn=yolo_det_video,
        inputs=inputs_video_list,
        outputs=outputs_video_list,
        title=title,
        description=description,
        # article=article,
        # examples=examples,
        # theme="seafoam",
        # live=True, # Change output in real time
        flagging_dir="run",  # output directory
        allow_flagging="never",
        # flagging_options=["good", "generally", "bad"],
    )

    gyd = gr.TabbedInterface(interface_list=[gyd_img, gyd_video], tab_names=["Image Mode", "Video Mode"])

    if not is_login:
        gyd.launch(
            inbrowser=True,  # Automatically open default browser
            show_tips=True,  # Automatically display the latest features of gradio
            share=is_share,  # Project sharing, other devices can access
            favicon_path="./icon/logo.ico",  # web icon
            show_error=True,  # Display error message in browser console
            quiet=True,  # Suppress most print statements
        )
    else:
        gyd.launch(
            inbrowser=True,  # Automatically open default browser
            show_tips=True,  # Automatically display the latest features of gradio
            auth=usr_pwd,  # login interface
            share=is_share,  # Project sharing, other devices can access
            favicon_path="./icon/logo.ico",  # web icon
            show_error=True,  # Display error message in browser console
            quiet=True,  # Suppress most print statements
        )


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