hubconf.py

import os
import cv2
import math
import pathlib
import torch
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

from yolov6.layers.common import DetectBackend
from yolov6.utils.nms import non_max_suppression
from yolov6.data.data_augment import letterbox
from yolov6.core.inferer import Inferer
from yolov6.utils.events import LOGGER
from yolov6.utils.events import load_yaml

PATH_YOLOv6 = pathlib.Path(__file__).parent
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
CLASS_NAMES = load_yaml(str(PATH_YOLOv6/"data/coco.yaml"))['names']


def visualize_detections(image,
                         boxes,
                         classes,
                         scores,
                         min_score=0.4,
                         figsize=(16, 16),
                         linewidth=2,
                         color='lawngreen'
                         ):
    image = np.array(image, dtype=np.uint8)
    fig = plt.figure(figsize=figsize)
    plt.axis("off")
    plt.imshow(image)
    ax = plt.gca()
    for box, name, score in zip(boxes, classes, scores):
        if score >= min_score:
            text = "{}: {:.2f}".format(name, score)
            x1, y1, x2, y2 = box
            w, h = x2 - x1, y2 - y1
            patch = plt.Rectangle(
                [x1, y1], w, h, fill=False, edgecolor=color, linewidth=linewidth
            )
            ax.add_patch(patch)
            ax.text(
                x1,
                y1,
                text,
                bbox={"facecolor": color, "alpha": 0.8},
                clip_box=ax.clipbox,
                clip_on=True,
            )
    plt.show()


def check_img_size(img_size, s=32, floor=0):
    def make_divisible(x, divisor):
        return math.ceil(x / divisor) * divisor
    if isinstance(img_size, int):  # integer i.e. img_size=640
        new_size = max(make_divisible(img_size, int(s)), floor)
    elif isinstance(img_size, list):  # list i.e. img_size=[640, 480]
        new_size = [max(make_divisible(x, int(s)), floor) for x in img_size]
    else:
        raise Exception(f"Unsupported type of img_size: {type(img_size)}")

    if new_size != img_size:
        LOGGER.info(
            f'WARNING: --img-size {img_size} must be multiple of max stride {s}, updating to {new_size}')
    return new_size if isinstance(img_size, list) else [new_size] * 2


def process_image(path, img_size, stride):
    '''Preprocess image before inference.'''
    try:
            img_src = cv2.imread(path)
            img_src = cv2.cvtColor(img_src, cv2.COLOR_RGB2BGR)
            assert img_src is not None, f"opencv cannot read image correctly or {path} not exists"
    except:
            img_src = np.asarray(Image.open(path))
            assert img_src is not None, f"Image Not Found {path}, workdir: {os.getcwd()}"

    image = letterbox(img_src, img_size, stride=stride)[0]
    image = image.transpose((2, 0, 1)) # HWC to CHW
    image = torch.from_numpy(np.ascontiguousarray(image))
    image = image.float()
    image /= 255
    return image, img_src


class Detector(DetectBackend):
    def __init__(self,
                 ckpt_path,
                 class_names,
                 device,
                 img_size=640,
                 conf_thres=0.25,
                 iou_thres=0.45,
                 max_det=1000):
        super().__init__(ckpt_path, device)
        self.class_names = class_names
        self.model.float()
        self.device = device
        self.img_size = check_img_size(img_size)
        self.conf_thres = conf_thres
        self.iou_thres = iou_thres
        self.max_det = max_det

    def forward(self, x, src_shape):
        pred_results = super().forward(x)
        classes = None  # the classes to keep
        det = non_max_suppression(pred_results, self.conf_thres, self.iou_thres,
                                  classes, agnostic=False, max_det=self.max_det)[0]

        det[:, :4] = Inferer.rescale(
            x.shape[2:], det[:, :4], src_shape).round()
        boxes = det[:, :4]
        scores = det[:, 4]
        labels = det[:, 5].long()
        prediction = {'boxes': boxes, 'scores': scores, 'labels': labels}
        return prediction

    def predict(self, img_path):
        img, img_src = process_image(img_path, self.img_size, 32)
        img = img.to(self.device)
        if len(img.shape) == 3:
            img = img[None]

        prediction = self.forward(img, img_src.shape)
        out = {k: v.cpu().numpy() for k, v in prediction.items()}
        out['classes'] = [self.class_names[i] for i in out['labels']]
        return out

    def show_predict(self,
                     img_path,
                     min_score=0.5,
                     figsize=(16, 16),
                     color='lawngreen',
                     linewidth=2):
        prediction = self.predict(img_path)
        boxes, scores, classes = prediction['boxes'], prediction['scores'], prediction['classes']
        visualize_detections(Image.open(img_path),
                             boxes, classes, scores,
                             min_score=min_score, figsize=figsize,  color=color, linewidth=linewidth
                             )


def create_model(model_name, class_names=CLASS_NAMES, device=DEVICE,
                 img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    if not os.path.exists(str(PATH_YOLOv6/'weights')):
        os.mkdir(str(PATH_YOLOv6/'weights'))
    if not os.path.exists(str(PATH_YOLOv6/'weights') + f'/{model_name}.pt'):
        torch.hub.load_state_dict_from_url(
            f"https://github.com/meituan/YOLOv6/releases/download/0.3.0/{model_name}.pt",
            str(PATH_YOLOv6/'weights'))
    return Detector(str(PATH_YOLOv6/'weights') + f'/{model_name}.pt',
                    class_names, device, img_size=img_size, conf_thres=conf_thres,
                    iou_thres=iou_thres, max_det=max_det)


def yolov6n(class_names=CLASS_NAMES, device=DEVICE, img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    return create_model('yolov6n', class_names, device, img_size=img_size, conf_thres=conf_thres,
                        iou_thres=iou_thres, max_det=max_det)


def yolov6s(class_names=CLASS_NAMES, device=DEVICE, img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    return create_model('yolov6s', class_names, device, img_size=img_size, conf_thres=conf_thres,
                        iou_thres=iou_thres, max_det=max_det)


def yolov6m(class_names=CLASS_NAMES, device=DEVICE, img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    return create_model('yolov6m', class_names, device, img_size=img_size, conf_thres=conf_thres,
                        iou_thres=iou_thres, max_det=max_det)


def yolov6l(class_names=CLASS_NAMES, device=DEVICE, img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    return create_model('yolov6l', class_names, device, img_size=img_size, conf_thres=conf_thres,
                        iou_thres=iou_thres, max_det=max_det)


def custom(ckpt_path, class_names, device=DEVICE, img_size=640, conf_thres=0.25, iou_thres=0.45, max_det=1000):
    return Detector(ckpt_path, class_names, device, img_size=img_size, conf_thres=conf_thres,
                    iou_thres=iou_thres, max_det=max_det)