Aquarium Combined.v2-raw-1024.voc
├── test [126 entries exceeds filelimit, not opening dir]
├── train [894 entries exceeds filelimit, not opening dir]
├── valid [254 entries exceeds filelimit, not opening dir]
├── README.dataset.txt
└── README.roboflow.txt

git clone https://github.com/sovit-123/vision_transformers.git
cd vision_transformers

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

pip install -r requirements.txt

pip install vision_transformers

├── input
│   ├── Aquarium Combined.v2-raw-1024.voc
│   └── inference_data
└── vision_transformers
    ├── data
    ├── examples
    ├── example_test_data
    ├── readme_images
    ├── runs
    ├── tools
    ├── vision_transformers
    ├── README.md
    ├── requirements.txt
    └── setup.py

# 圖像和標(biāo)簽?zāi)夸浵鄬?duì)于train.py腳本的相對(duì)路徑
TRAIN_DIR_IMAGES: '../input/Aquarium Combined.v2-raw-1024.voc/train'
TRAIN_DIR_LABELS: '../input/Aquarium Combined.v2-raw-1024.voc/train'
VALID_DIR_IMAGES: '../input/Aquarium Combined.v2-raw-1024.voc/valid'
VALID_DIR_LABELS: '../input/Aquarium Combined.v2-raw-1024.voc/valid'
# 類名
CLASSES: [
    '__background__',
    'fish', 'jellyfish', 'penguin',
    'shark', 'puffin', 'stingray',
    'starfish'
]
# 類別數(shù)
NC: 8
# 是否在訓(xùn)練期間保存驗(yàn)證集的預(yù)測(cè)結(jié)果
SAVE_VALID_PREDICTION_IMAGES: True

python tools/train_detector.py --epochs 20 --batch 2 --data data/aquarium.yaml --model detr_resnet50 --name detr_resnet50

IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.172
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.383
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.126
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.094
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.107
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.247
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.088
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.250
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.337
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.235
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.330
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.344
BEST VALIDATION mAP: 0.17192136022687962
SAVING BEST MODEL FOR EPOCH: 20

python tools/train_detector.py --epochs 20 --batch 2 --data data/aquarium.yaml --model detr_resnet50_dc5 --name detr_resnet50_dc5

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.161
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.360
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.123
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.141
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.155
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.233
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.096
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.248
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.345
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.379
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.373
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.340
BEST VALIDATION mAP: 0.16066837142161672
SAVING BEST MODEL FOR EPOCH: 20

python tools/train_detector.py --epochs 20 --batch 2 --data data/aquarium.yaml --model detr_resnet101 --name detr_resnet101

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.175
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.381
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.132
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.089
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.113
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.260
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.095
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.269
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.362
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.298
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.451
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.351
BEST VALIDATION mAP: 0.17489964894400944
SAVING BEST MODEL FOR EPOCH: 17

python tools/train_detector.py --epochs 20 --batch 2 --data data/aquarium.yaml --model detr_resnet101_dc5 --name detr_resnet101_dc5

IoU metric: bbox
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.206
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.438
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.178
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.110
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.093
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.303
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.099
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.287
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.391
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.317
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.394
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.394
BEST VALIDATION mAP: 0.20588343074278573
SAVING BEST MODEL FOR EPOCH: 20

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.239
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.501
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.186
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.119
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.143
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.328
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.109
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.290
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.394
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.349
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.369
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.398
BEST VALIDATION mAP: 0.23894132553612263
SAVING BEST MODEL FOR EPOCH: 48

python tools/inference_video_detect.py --weights runs/training/detr_resnet101_dc5_60e/best_model.pth --input ../input/inference_data/video_1.mp4 --show

import torch
import cv2
import numpy as np
import argparse
import yaml
import os
import time
import torchinfo

from vision_transformers.detection.detr.model import DETRModel
from utils.detection.detr.general import (
    set_infer_dir,
    load_weights
)
from utils.detection.detr.transforms import infer_transforms, resize
from utils.detection.detr.annotations import (
    convert_detections,
    inference_annotations,
    annotate_fps,
    convert_pre_track,
    convert_post_track
)
from deep_sort_realtime.deepsort_tracker import DeepSort
from utils.detection.detr.viz_attention import visualize_attention

# NumPy隨機(jī)數(shù)生成器的種子值為2023
np.random.seed(2023)

# 命令行參數(shù)配置選項(xiàng)
def parse_opt():
    parser = argparse.ArgumentParser()
    # 模型權(quán)重文件的路徑
    parser.add_argument(
        '-w', 
        '--weights',
    )
    # 輸入圖像或圖像文件夾的路徑
    parser.add_argument(
        '-i', '--input', 
        help='folder path to input input image (one image or a folder path)',
    )
    # 數(shù)據(jù)配置文件的路徑
    parser.add_argument(
        '--data', 
        default=None,
        help='(optional) path to the data config file'
    )
    # 模型名稱，默認(rèn)為'detr_resnet50'
    parser.add_argument(
        '--model', 
        default='detr_resnet50',
        help='name of the model'
    )
    # 計(jì)算和訓(xùn)練設(shè)備，默認(rèn)使用GPU（如果可用）否則使用CPU
    parser.add_argument(
        '--device', 
        default=torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'),
        help='computation/training device, default is GPU if GPU present'
    )
    # 圖像的尺寸，默認(rèn)為640
    parser.add_argument(
        '--imgsz', 
        '--img-size', 
        default=640,
        dest='imgsz',
        type=int,
        help='resize image to, by default use the original frame/image size'
    )
    # 可視化時(shí)的置信度閾值，默認(rèn)為0.5
    parser.add_argument(
        '-t', 
        '--threshold',
        type=float,
        default=0.5,
        help='confidence threshold for visualization'
    )
    # 訓(xùn)練結(jié)果存放目錄
    parser.add_argument(
        '--name', 
        default=None, 
        type=str, 
        help='training result dir name in outputs/training/, (default res_#)'
    )
    # 不顯示邊界框上的標(biāo)簽
    parser.add_argument(
        '--hide-labels',
        dest='hide_labels',
        action='store_true',
        help='do not show labels during on top of bounding boxes'
    )
    # 只有傳遞該選項(xiàng)時(shí)才會(huì)顯示輸出
    parser.add_argument(
        '--show', 
        dest='show', 
        action='store_true',
        help='visualize output only if this argument is passed'
    )
    # 開啟跟蹤功能
    parser.add_argument(
        '--track',
        action='store_true'
    )
    # 過濾要可視化的類別，如--classes 1 2 3
    parser.add_argument(
        '--classes',
        nargs='+',
        type=int,
        default=None,
        help='filter classes by visualization, --classes 1 2 3'
    )
    # 可視化檢測(cè)框的注意力圖
    parser.add_argument(
        '--viz-atten',
        dest='vis_atten',
        action='store_true',
        help='visualize attention map of detected boxes'
    )
    args = parser.parse_args()
    return args

# 讀取并處理視頻文件相關(guān)信息
def read_return_video_data(video_path):
    # 打開指定路徑的視頻文件
    cap = cv2.VideoCapture(video_path)
    # 獲取視頻幀的寬度和高度
    frame_width = int(cap.get(3))
    frame_height = int(cap.get(4))
    # 獲取視頻的幀率
    fps = int(cap.get(5))
    # 檢查視頻的寬度和高度是否不為零。如果它們都是零，那么會(huì)拋出一個(gè)錯(cuò)誤消息，提示用戶檢查視頻路徑是否正確
    assert (frame_width != 0 and frame_height !=0), 'Please check video path...'
    # 函數(shù)返回一個(gè)元組，包含VideoCapture對(duì)象cap以及視頻的寬度、高度和幀率fps
    return cap, frame_width, frame_height, fps

def main(args):
    # 如果args.track為真，初始化DeepSORT追蹤器
    if args.track:
        tracker = DeepSort(max_age=30)
    # 根據(jù)args.data加載數(shù)據(jù)配置（如果存在）以獲取類別數(shù)量和類別列表
    NUM_CLASSES = None
    CLASSES = None
    data_configs = None
    if args.data is not None:
        with open(args.data) as file:
            data_configs = yaml.safe_load(file)
        NUM_CLASSES = data_configs['NC']
        CLASSES = data_configs['CLASSES']
    # 獲取設(shè)備類型
    DEVICE = args.device
    # 設(shè)置輸出目錄
    OUT_DIR = set_infer_dir(args.name)
    # 加載模型權(quán)重
    model, CLASSES, data_path = load_weights(
        args, 
        # 設(shè)備類型
        DEVICE, 
        # 模型類
        DETRModel, 
        # 數(shù)據(jù)配置
        data_configs, 
        # 類別數(shù)量
        NUM_CLASSES, 
        # 類別列表
        CLASSES, 
        video=True
    )
    # 將模型移動(dòng)到指定的設(shè)備（如GPU或CPU）并將其設(shè)置為評(píng)估模式（.eval()）
    _ = model.to(DEVICE).eval()
    # 使用torchinfo.summary來打印模型的詳細(xì)結(jié)構(gòu)和參數(shù)統(tǒng)計(jì)
    try:
        torchinfo.summary(
            model, 
            device=DEVICE, 
            input_size=(1, 3, args.imgsz, args.imgsz), 
            row_settings=["var_names"]
        )
    # 如果此過程出現(xiàn)異常，代碼會(huì)打印模型的完整結(jié)構(gòu)，并計(jì)算模型的總參數(shù)數(shù)和可訓(xùn)練參數(shù)數(shù)
    except:
        print(model)
        # 計(jì)算模型的所有參數(shù)總數(shù)
        total_params = sum(p.numel() for p in model.parameters())
        print(f"{total_params:,} total parameters.")
        # 只計(jì)算那些需要在訓(xùn)練過程中更新的參數(shù)（即requires_grad屬性為True的參數(shù)）
        total_trainable_params = sum(
            p.numel() for p in model.parameters() if p.requires_grad)
        print(f"{total_trainable_params:,} training parameters.")

    # 生成一個(gè)隨機(jī)分布的顏色數(shù)組，每個(gè)元素的值在0到255之間，這是標(biāo)準(zhǔn)的8位RGB色彩空間中的每個(gè)通道的取值范圍
    COLORS = np.random.uniform(0, 255, size=(len(CLASSES), 3))
    # 獲取視頻的路徑
    VIDEO_PATH = args.input
    # 如果用戶沒有通過命令行參數(shù)--input指定視頻路徑，則將VIDEO_PATH設(shè)置為data_path
    if VIDEO_PATH == None:
        VIDEO_PATH = data_path
    # cap: 一個(gè)cv2.VideoCapture對(duì)象，用于讀取和處理視頻文件
    # frame_width: 視頻的幀寬度（寬度像素?cái)?shù)）
    # frame_height: 視頻的幀高度（高度像素?cái)?shù)）
    # fps: 視頻的幀率（每秒幀數(shù)）
    cap, frame_width, frame_height, fps = read_return_video_data(VIDEO_PATH)
    # 生成輸出文件的名稱
    # [-1]：選取列表中的最后一個(gè)元素，即文件名（包括擴(kuò)展名）
    # .split('.')[0]：再次分割文件名，這次是基于點(diǎn)號(hào)（.）來分隔，然后選取第一個(gè)元素，即文件的基本名稱，不包括擴(kuò)展名
    save_name = VIDEO_PATH.split(os.path.sep)[-1].split('.')[0]
    # 將處理后的幀寫入輸出視頻文件
    # 輸出文件路徑：f"{OUT_DIR}/{save_name}.mp4"
    # 編碼器（codec）：cv2.VideoWriter_fourcc(*'mp4v')
    # 幀率（fps）
    # 視頻尺寸：(frame_width, frame_height)
    out = cv2.VideoWriter(f"{OUT_DIR}/{save_name}.mp4", 
                        cv2.VideoWriter_fourcc(*'mp4v'), fps, 
                        (frame_width, frame_height))
    # 檢查args.imgsz是否已設(shè)置（即用戶是否通過命令行參數(shù)指定了圖像大小）
    # 如果args.imgsz有值，說明用戶想要將輸入圖像（或視頻幀）縮放到指定的大小，那么RESIZE_TO將被設(shè)置為這個(gè)值
    if args.imgsz != None:
        RESIZE_TO = args.imgsz
    # 如果args.imgsz沒有設(shè)置或者為None，則默認(rèn)使用視頻幀的原始寬度frame_width作為縮放尺寸
    else:
        RESIZE_TO = frame_width
    # 記錄總的幀數(shù)
    frame_count = 0
    # 計(jì)算最終的幀率
    total_fps = 0

    # 檢查視頻是否已經(jīng)結(jié)束
    while(cap.isOpened()):
        # 讀取下一幀，并返回一個(gè)布爾值ret表示是否成功讀取
        ret, frame = cap.read()
        if ret:
            # 復(fù)制原始幀以保留未處理的版本
            orig_frame = frame.copy()
            # 使用resize函數(shù)將幀調(diào)整到指定的大小（如果args.imgsz已設(shè)置，否則保持原大?。?            frame = resize(frame, RESIZE_TO, square=True)
            image = frame.copy()
            # 將BGR圖像轉(zhuǎn)換為RGB
            image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
            # 將圖像歸一化到0-1范圍
            image = image / 255.0
            # 預(yù)處理
            image = infer_transforms(image)
            # 將圖像轉(zhuǎn)換為PyTorch張量，設(shè)置數(shù)據(jù)類型為torch.float32
            image = torch.tensor(image, dtype=torch.float32)
            # 調(diào)整張量維度，使通道維度成為第一個(gè)維度，以便于模型輸入（模型通常期望輸入張量的形狀為(batch_size, channels, height, width)）
            image = torch.permute(image, (2, 0, 1))
            # 在張量前面添加一個(gè)維度以表示批次大小（batch_size=1）
            image = image.unsqueeze(0)

            # 計(jì)算模型前向傳播的時(shí)間（start_time和forward_end_time）以衡量處理單幀的速度
            start_time = time.time()
            with torch.no_grad():
                outputs = model(image.to(args.device))
            forward_end_time = time.time()

            forward_pass_time = forward_end_time - start_time

            # 計(jì)算當(dāng)前幀的處理速度
            fps = 1 / (forward_pass_time)
            # Add `fps` to `total_fps`.
            total_fps += fps
            # Increment frame count.
            frame_count += 1
            # 如果啟用了注意力可視化（args.vis_atten），則將注意力圖保存為圖像文件
            if args.vis_atten:
                visualize_attention(
                    model,
                    image, 
                    args.threshold, 
                    orig_frame,
                    f"{OUT_DIR}/frame_{str(frame_count)}.png",
                    DEVICE
                )
            # 如果模型檢測(cè)到了物體（outputs['pred_boxes'][0]非空）
            if len(outputs['pred_boxes'][0]) != 0:
                # 轉(zhuǎn)換預(yù)測(cè)結(jié)果
                draw_boxes, pred_classes, scores = convert_detections(
                    outputs, 
                    args.threshold,
                    CLASSES,
                    orig_frame,
                    args 
                )
                # 使用tracker更新跟蹤狀態(tài)，并將結(jié)果轉(zhuǎn)換回檢測(cè)框（convert_pre_track和convert_post_track）
                if args.track:
                    tracker_inputs = convert_pre_track(
                        draw_boxes, pred_classes, scores
                    )
                    # Update tracker with detections.
                    tracks = tracker.update_tracks(
                        tracker_inputs, frame=frame
                    )
                    draw_boxes, pred_classes, scores = convert_post_track(tracks) 
                # 將預(yù)測(cè)結(jié)果應(yīng)用到原始幀上（inference_annotations），包括繪制邊界框、類別標(biāo)簽和置信度
                orig_frame = inference_annotations(
                    draw_boxes,
                    pred_classes,
                    scores,
                    CLASSES,
                    COLORS,
                    orig_frame,
                    args
                )
            # 在幀上添加實(shí)時(shí)FPS信息
            orig_frame = annotate_fps(orig_frame, fps)
            # 將處理后的幀寫入輸出視頻文件
            out.write(orig_frame)
            if args.show:
                cv2.imshow('Prediction', orig_frame)
                # Press `q` to exit
                if cv2.waitKey(1) & 0xFF == ord('q'):
                    break
        else:
            break
    if args.show:
        # Release VideoCapture().
        cap.release()
        # Close all frames and video windows.
        cv2.destroyAllWindows()

    # Calculate and print the average FPS.
    avg_fps = total_fps / frame_count
    print(f"Average FPS: {avg_fps:.3f}")

if __name__ == '__main__':
    args = parse_opt()
    main(args)

python tools/inference_image_detect.py --weights runs/training/detr_resnet101_dc5_60e/best_model.pth --input "../input/Aquarium Combined.v2-raw-1024.voc/test"

import torch
import cv2
import numpy as np
import argparse
import yaml
import glob
import os
import time
import torchinfo

from vision_transformers.detection.detr.model import DETRModel
from utils.detection.detr.general import (
    set_infer_dir,
    load_weights
)
from utils.detection.detr.transforms import infer_transforms, resize
from utils.detection.detr.annotations import (
    convert_detections,
    inference_annotations, 
)
from utils.detection.detr.viz_attention import visualize_attention

np.random.seed(2023)

def parse_opt():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '-w', 
        '--weights',
    )
    parser.add_argument(
        '-i', '--input', 
        help='folder path to input input image (one image or a folder path)',
    )
    parser.add_argument(
        '--data', 
        default=None,
        help='(optional) path to the data config file'
    )
    parser.add_argument(
        '--model', 
        default='detr_resnet50',
        help='name of the model'
    )
    parser.add_argument(
        '--device', 
        default=torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'),
        help='computation/training device, default is GPU if GPU present'
    )
    parser.add_argument(
        '--imgsz', 
        '--img-size',
        default=640,
        dest='imgsz',
        type=int,
        help='resize image to, by default use the original frame/image size'
    )
    parser.add_argument(
        '-t', 
        '--threshold',
        type=float,
        default=0.5,
        help='confidence threshold for visualization'
    )
    parser.add_argument(
        '--name', 
        default=None, 
        type=str, 
        help='training result dir name in outputs/training/, (default res_#)'
    )
    parser.add_argument(
        '--hide-labels',
        dest='hide_labels',
        action='store_true',
        help='do not show labels during on top of bounding boxes'
    )
    parser.add_argument(
        '--show', 
        dest='show', 
        action='store_true',
        help='visualize output only if this argument is passed'
    )
    parser.add_argument(
        '--track',
        action='store_true'
    )
    parser.add_argument(
        '--classes',
        nargs='+',
        type=int,
        default=None,
        help='filter classes by visualization, --classes 1 2 3'
    )
    parser.add_argument(
        '--viz-atten',
        dest='vis_atten',
        action='store_true',
        help='visualize attention map of detected boxes'
    )
    args = parser.parse_args()
    return args

def collect_all_images(dir_test):
    """
    Function to return a list of image paths.
    :param dir_test: Directory containing images or single image path.
    Returns:
        test_images: List containing all image paths.
    """
    test_images = []
    if os.path.isdir(dir_test):
        image_file_types = ['*.jpg', '*.jpeg', '*.png', '*.ppm']
        for file_type in image_file_types:
            test_images.extend(glob.glob(f"{dir_test}/{file_type}"))
    else:
        test_images.append(dir_test)
    return test_images   

def main(args):
    NUM_CLASSES = None
    CLASSES = None
    data_configs = None
    if args.data is not None:
        with open(args.data) as file:
            data_configs = yaml.safe_load(file)
        NUM_CLASSES = data_configs['NC']
        CLASSES = data_configs['CLASSES']
    
    DEVICE = args.device
    OUT_DIR = set_infer_dir(args.name)

    model, CLASSES, data_path = load_weights(
        args, DEVICE, DETRModel, data_configs, NUM_CLASSES, CLASSES
    )
    _ = model.to(DEVICE).eval()
    try:
        torchinfo.summary(
            model, 
            device=DEVICE, 
            input_size=(1, 3, args.imgsz, args.imgsz),
            row_settings=["var_names"]
        )
    except:
        print(model)
        # Total parameters and trainable parameters.
        total_params = sum(p.numel() for p in model.parameters())
        print(f"{total_params:,} total parameters.")
        total_trainable_params = sum(
            p.numel() for p in model.parameters() if p.requires_grad)
        print(f"{total_trainable_params:,} training parameters.")

    # Colors for visualization.
    COLORS = np.random.uniform(0, 255, size=(len(CLASSES), 3))
    DIR_TEST = args.input
    if DIR_TEST == None:
        DIR_TEST = data_path
    test_images = collect_all_images(DIR_TEST)
    print(f"Test instances: {len(test_images)}")

    # To count the total number of frames iterated through.
    frame_count = 0
    # To keep adding the frames' FPS.
    total_fps = 0
    for image_num in range(len(test_images)):
        image_name = test_images[image_num].split(os.path.sep)[-1].split('.')[0]
        orig_image = cv2.imread(test_images[image_num])
        frame_height, frame_width, _ = orig_image.shape
        if args.imgsz != None:
            RESIZE_TO = args.imgsz
        else:
            RESIZE_TO = frame_width
        
        image_resized = resize(orig_image, RESIZE_TO, square=True)
        image = cv2.cvtColor(image_resized, cv2.COLOR_BGR2RGB)
        image = image / 255.0
        image = infer_transforms(image)
        input_tensor = torch.tensor(image, dtype=torch.float32)
        input_tensor = torch.permute(input_tensor, (2, 0, 1))
        input_tensor = input_tensor.unsqueeze(0)
        h, w, _ = orig_image.shape

        start_time = time.time()
        with torch.no_grad():
            outputs = model(input_tensor.to(DEVICE))
        end_time = time.time()
        # Get the current fps.
        fps = 1 / (end_time - start_time)
        # Add `fps` to `total_fps`.
        total_fps += fps
        # Increment frame count.
        frame_count += 1

        if args.vis_atten:
            visualize_attention(
                model,
                input_tensor, 
                args.threshold, 
                orig_image,
                f"{OUT_DIR}/{image_name}.png",
                DEVICE
            )

        if len(outputs['pred_boxes'][0]) != 0:
            draw_boxes, pred_classes, scores = convert_detections(
                outputs, 
                args.threshold,
                CLASSES,
                orig_image,
                args 
            )
            orig_image = inference_annotations(
                draw_boxes,
                pred_classes,
                scores,
                CLASSES,
                COLORS,
                orig_image,
                args
            )
            if args.show:
                cv2.imshow('Prediction', orig_image)
                cv2.waitKey(1)
            
        cv2.imwrite(f"{OUT_DIR}/{image_name}.jpg", orig_image)
        print(f"Image {image_num+1} done...")
        print('-'*50)

    print('TEST PREDICTIONS COMPLETE')
    if args.show:
        cv2.destroyAllWindows()
        # Calculate and print the average FPS.
    avg_fps = total_fps / frame_count
    print(f"Average FPS: {avg_fps:.3f}")

if __name__ == '__main__':
    args = parse_opt()
    main(args)

python tools/inference_image_detect.py \
    --weights /path/to/best/weights.pth \
    --input /path/to/test/images/directory \
    --threshold 0.5