├── .idea ├── .gitignore ├── DRENet.iml ├── deployment.xml ├── inspectionProfiles │ ├── Project_Default.xml │ └── profiles_settings.xml ├── misc.xml ├── modules.xml └── vcs.xml ├── DegradeGenerate.py ├── LICENSE ├── NetworkStructure.png ├── README.md ├── data ├── hyp.finetune.yaml ├── hyp.scratch.yaml └── ship.yaml ├── detect.py ├── hubconf.py ├── models ├── Ablation_Only_CRMA.yaml ├── Ablation_Only_CRMA_Remove_PAN.yaml ├── Ablation_Only_CRMA_and_Scale.yaml ├── Ablation_Only_Detector.yaml ├── Ablation_Only_Enhancer.yaml ├── Ablation_Only_Scale.yaml ├── DRENet.yaml ├── __init__.py ├── common.py ├── experimental.py ├── export.py ├── yolo.py └── yolov5s.yaml ├── requirements.txt ├── test.py ├── train.py └── utils ├── __init__.py ├── activations.py ├── autoanchor.py ├── datasets.py ├── general.py ├── google_utils.py ├── loss.py ├── metrics.py ├── plots.py └── torch_utils.py /.idea/.gitignore: -------------------------------------------------------------------------------- 1 | # Default ignored files 2 | /shelf/ 3 | /workspace.xml 4 | # Datasource local storage ignored files 5 | /dataSources/ 6 | /dataSources.local.xml 7 | # Editor-based HTTP Client requests 8 | /httpRequests/ 9 | -------------------------------------------------------------------------------- /.idea/DRENet.iml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | -------------------------------------------------------------------------------- /.idea/deployment.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 | 39 | 40 | 41 | 42 | -------------------------------------------------------------------------------- /.idea/inspectionProfiles/Project_Default.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 17 | -------------------------------------------------------------------------------- /.idea/inspectionProfiles/profiles_settings.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 6 | -------------------------------------------------------------------------------- /.idea/misc.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 6 | 7 | -------------------------------------------------------------------------------- /.idea/modules.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | -------------------------------------------------------------------------------- /.idea/vcs.xml: -------------------------------------------------------------------------------- 1 | 2 | 3 | 4 | 5 | 6 | -------------------------------------------------------------------------------- /DegradeGenerate.py: -------------------------------------------------------------------------------- 1 | import warnings 2 | import skimage 3 | from skimage import io 4 | import os 5 | from glob import glob 6 | from tqdm import tqdm 7 | import numpy as np 8 | import cv2 9 | from numba import jit 10 | import concurrent.futures 11 | 12 | warnings.filterwarnings('ignore') 13 | 14 | # @jit(nopython=True) 15 | def customBlur(imgFile): 16 | size = 512 17 | targetPath = r'.\train\degrade' # save path 18 | labelFile = imgFile.replace('png', 'txt').replace('images', 'labels') 19 | if os.path.exists(labelFile): 20 | img = io.imread(imgFile) 21 | label = np.loadtxt(labelFile, ndmin=2) 22 | dst = img.copy() 23 | centers = label.copy() 24 | for j in range(len(centers)): 25 | centers[j, 1] = label[j, 1] * size 26 | centers[j, 2] = label[j, 2] * size 27 | for i in range(img.shape[0]): 28 | for j in range(img.shape[1]): 29 | minDis = 130 * 130 30 | for center in centers: 31 | distance = (i - center[2]) ** 2 + (j - center[1]) ** 2 32 | if distance < minDis: 33 | minDis = distance 34 | # boxSize = (int(0.05* (minDis ** 0.5))) // 2 + 1 35 | boxSize = (int(1.03** (minDis ** 0.5))) // 2 # You can change the degradation configuration here 36 | # boxSize = int(((minDis**0.5)/5+1))//2 37 | dst[i, j,0] = img[max(i - boxSize, 0):min(i + boxSize + 1, size), max(j - boxSize, 0):min(j + boxSize + 1, size),0].mean() 38 | dst[i, j, 1] = img[max(i - boxSize, 0):min(i + boxSize + 1, size), 39 | max(j - boxSize, 0):min(j + boxSize + 1, size), 1].mean() 40 | dst[i, j, 2] = img[max(i - boxSize, 0):min(i + boxSize + 1, size), 41 | max(j - boxSize, 0):min(j + boxSize + 1, size), 2].mean() 42 | io.imsave(os.path.join(targetPath, os.path.basename(imgFile)), dst) 43 | else: 44 | img = cv2.imread(imgFile) 45 | dst = cv2.blur(img, (20, 20)) 46 | cv2.imwrite(os.path.join(targetPath, os.path.basename(imgFile)), dst) 47 | 48 | if __name__=='__main__': 49 | sourcePath = r'.\train\images' # source path 50 | with concurrent.futures.ProcessPoolExecutor(1) as executor: # set 1 to other number for speedup 51 | imgfiles=glob(os.path.join(sourcePath, '*.png')) 52 | for i in tqdm(zip(imgfiles, executor.map(customBlur, imgfiles)),total=1): 53 | pass 54 | -------------------------------------------------------------------------------- /NetworkStructure.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/WindVChen/DRENet/a187dbe0f623b521a62c6176c7cafaa7322f5f66/NetworkStructure.png -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # DRENet: Fast and Accurate Tiny Ship detection Method 2 | 3 | ![](https://komarev.com/ghpvc/?username=windvchendrenet&label=visitors) 4 | ![GitHub stars](https://badgen.net/github/stars/windvchen/DRENet) 5 | [![](https://img.shields.io/badge/license-GPL--3.0-brightgreen)](#License) 6 | 7 | ## Share us a :star: if this repo does help 8 | 9 | This repo is the official implementation of the DRENet in "A Degraded Reconstruction Enhancement-based Method for Tiny Ship Detection in Remote Sensing Images with A New Large-scale Dataset". The paper can be accessed in [[IEEE](https://ieeexplore.ieee.org/document/9791363) | [Lab Server](http://levir.buaa.edu.cn/publications/DRENet.pdf) | [ResearchGate](https://www.researchgate.net/profile/Keyan-Chen-6/publication/361178478_A_Degraded_Reconstruction_Enhancement-based_Method_for_Tiny_Ship_Detection_in_Remote_Sensing_Images_with_A_New_Large-scale_Dataset/links/62f47d69b8dc8b4403d4ce5e/A-Degraded-Reconstruction-Enhancement-Based-Method-for-Tiny-Ship-Detection-in-Remote-Sensing-Images-With-a-New-Large-Scale-Dataset.pdf)]. ***(Accepted by TGRS 2022)*** 10 | 11 | If you encounter any question, please feel free to contact us. You can create an issue or just send email to me windvchen@gmail.com. Also welcome for any idea exchange and discussion. 12 | 13 | ## Updates 14 | 15 | ***05/30/2023*** 16 | 17 | Thanks for the discussions in [#10](https://github.com/WindVChen/DRENet/issues/10). The current code now supports processing different resolutions without manually modifying the yaml configurations. To achieve this, you can comment out L139-141 in [common.py](./models/common.py). (Note that this feature has not well been tested, if there is a drop in performance, you can get back to the manual modification mentioned in [#4](https://github.com/WindVChen/DRENet/issues/4)). 18 | 19 | ***11/19/2022*** 20 | 21 | We have improved the previous codes, and the current repository support directly detect the images without generating degraded images, which correspond to the file **detect.py**. 22 | 23 | ***06/10/2022*** 24 | 25 | The code cleanup is finished and the complete codes are provided, also the weights of our model on LEVIR-Ship dataset. 26 | 27 | ***06/06/2022*** 28 | 29 | We will finish the code cleanup within a week, and make both the code and dataset fully public. Please be patient. 30 | 31 | 32 | ## Table of Contents 33 | 34 | - [Introduction](#Introduction) 35 | - [Results and Trained Model](#Results-and-Trained-Model) 36 | - [Preliminaries](#Preliminaries) 37 | - [Environments](#Environments) 38 | - [Run Details](#Run-Details) 39 | - [Train Process](#Train-Process) 40 | - [Valid Process](#Valid-Process) 41 | - [Detect Process](#Detect-Process) 42 | - [Visualization](#Visualization) 43 | - [Citation](#Citation) 44 | - [License](#License) 45 | 46 | ## Introduction 47 | ![Our Network Structure](NetworkStructure.png) 48 | 49 | We focus on **tiny ship detection** task in **medium-resolution (MR, about 16m/pixel)** remote sensing (RS) images . Compared with high-resolution (HR) RS image, an MR image covers a much wider area, thus facilitating quick ship detection. This direction has **great research significance**, and can greatly benefit the **rapid ship detection** under massive RS images. 50 | 51 | For the task, we propose an effective **Degraded Reconstruction Enhancement Network (DRENet)**, where a degraded reconstruction enhancer is designed to learn to regress an object-aware blurred version of the input image. Our method achieves both **great effectiveness and efficiency**, and outperforms many recent methods. 52 | 53 | ## Results and Trained Model 54 | ### Models trained on LEVIR-Ship dataset 55 | | Methods | Params(M) |FLOPs(G) | AP | FPS | 56 | |:---|:---:|:---:|:---:| :---:| 57 | | YOLOv3 | 61.52 | 99.2 | 69.9 | 61 | 58 | | YOLOv5s | 7.05 | 10.4 | 75.6 [[Google Drive](https://drive.google.com/file/d/10AQA_ynjvmVD8XSiOhM_9A64EDm5lxfS/view?usp=sharing)
| [Baidu Pan](https://pan.baidu.com/s/1AffKx_gChABQiicJv2zjtg) (code:ogdm)] | **95** | 59 | | Retinanet | 36.33 | 104.4 | 74.9 | 12 | 60 | | SSD | 24.39 | 175.2 | 52.6 | 25 | 61 | | FasterRCNN | 136.70 | 299.2 | 70.8 | 10 | 62 | | EfficientDet-D0 | **3.84** | **4.6** | 71.3 | 32 | 63 | | EfficientDet-D2 | 8.01 | 20.0 | 80.9 | 21 | 64 | | FCOS | 5.92 | 51.8 | 75.5 | 37 | 65 | | CenterNet | 191.24 | 584.6 | 77.7 | 25 | 66 | | HSFNet | 157.59 | 538.1 | 73.6 | 7 | 67 | | ImYOLOv3 | 62.86 | 101.9 | 72.6 | 51 | 68 | | MaskRCNN+DFR+RFE | 24.99 | 237.8 | 76.2 | 6 | 69 | | **DRENet** | 4.79 | 8.3 | **82.4** [[Google Drive](https://drive.google.com/file/d/1ApAejwSNYQDvROM1yRtltQOGdYAwYyF3/view?usp=sharing)
| [Baidu Pan](https://pan.baidu.com/s/1tBxhGOhmxc-L5ioHSqSjEQ) (code:x710)] | 85| 70 | 71 | 72 | ## Preliminaries 73 | Please at first download dataset [LEVIR-Ship](https://github.com/WindVChen/LEVIR-Ship), then prepare the dataset as the following structure: 74 | ``` 75 | ├── train 76 | ├── images 77 | ├── img_1.png 78 | ├── img_2.png 79 | ├── ... 80 | ├── degrade 81 | # images processed by Selective Degradation (refer to our paper for detals) 82 | ├── degraded_img_1.png 83 | ├── degraded_img_2.png 84 | ├── ... 85 | ├── labels 86 | ├── label_1.txt 87 | ├── label_2.txt 88 | ├── ... 89 | ├── val 90 | ├── test 91 | ``` 92 | Note that apart from the images and labels in LEVIR-Ship dataset, you should also generate the **degraded images**, which are the supervision of the enhancer (see details in our paper). Here, we provide [DegradeGenerate.py](DegradeGenerate.py) to easily generate the degraded images. 93 | 94 | After preparing the dataset as above, change the paths in [ship.yaml](data/ship.yaml). 95 | 96 | ***(The partitioned dataset, including the degraded images, can all be accessed [here](https://github.com/WindVChen/LEVIR-Ship))*** 97 | 98 | ## Environments 99 | 100 | - Windows/Linux Support 101 | - python 3.8 102 | - pytorch 1.9.0 103 | - torchvision 104 | - wandb (Suggested, a good tool to visualize the training process. If not want to use it, you should comment out the related codes.) 105 | - ...... (See more details in [requirements.txt](requirements.txt)) 106 | 107 | *(The code is constructed based on YOLOv5s, for more details about YOLOv5, please refer to their repo [here](https://github.com/ultralytics/yolov5).)* 108 | 109 | 110 | ## Run Details 111 | ### Train Process 112 | To train our `DRENet`, run: 113 | ``` 114 | python train.py --cfg "./models/DRENet.yaml" --epochs 1000 --workers 8 --batch-size 16 --device 0 --project "./LEVIR-Ship" --data "./data/ship.yaml" 115 | ``` 116 | **Parameters Description** 117 | - `cfg`: You can change it to use different network structures. More structure configurations can be found in [models](models) directory, where we provide the baseline YOLOv5s, and the ablation structures of DRENet. You can try them if you are interested 118 | - `epochs`: A longer training time is suggested, and 1,000 epochs are enough. 119 | - `project`: The path where you want to save your experiments. Also the name of the project in wandb. 120 | 121 | **Others** 122 | 123 | The current codes use **fixed weight balance**, which can also achieve a good result. 124 | 125 | If you want to make use of **automatic weight balance**, please search the key word `weightOptimizer` in [train.py](train.py) and uncomment the code lines, also the code lines with the key word `ForAuto` in [loss.py](utils/loss.py) be uncommented and the other lines be commented out. 126 | 127 | ### Valid Process 128 | 129 | To evaluate our `DRENet`, you should first train the network or download [our provided weights](#Models-trained-on-LEVIR-Ship-dataset), then run: 130 | ``` 131 | python test.py --weights "./DRENet.pt" --project "runs/test" --device 0 --batch-size 16 --data "./data/ship.yaml" 132 | ``` 133 | You can set how many detected results to plot by changing the value of `plot_batch_num` in [test.py](test.py). Also ensure that you have changed the val path in [ship.yaml](data/ship.yaml) into your test path. 134 | 135 | Please ensure that there are corresponding degraded images in the **degrade** folder. (See [#issue 4](https://github.com/WindVChen/DRENet/issues/4) for more details.) 136 | 137 | ### Detect Process 138 | To directly output the detect results **without the need of the degraded images**, please run the following command: 139 | ``` 140 | python detect.py --weights "./DRENet.pt" --source "images/" --device 0 141 | ``` 142 | where "--source" is the path that the images need detection in. 143 | 144 | 145 | ## Citation 146 | If you find this paper useful in your research, please consider citing: 147 | ``` 148 | @ARTICLE{9791363, 149 | author={Chen, Jianqi and Chen, Keyan and Chen, Hao and Zou, Zhengxia and Shi, Zhenwei}, 150 | journal={IEEE Transactions on Geoscience and Remote Sensing}, 151 | title={A Degraded Reconstruction Enhancement-based Method for Tiny Ship Detection in Remote Sensing Images with A New Large-scale Dataset}, 152 | year={2022}, 153 | volume={60}, 154 | number={}, 155 | pages={1-14}, 156 | doi={10.1109/TGRS.2022.3180894}} 157 | ``` 158 | 159 | 160 | ## License 161 | This project is licensed under the GPL-3.0 License. See [LICENSE](LICENSE) for details 162 | -------------------------------------------------------------------------------- /data/hyp.finetune.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for VOC finetuning 2 | # python train.py --batch 64 --weights yolov5m.pt --data voc.yaml --img 512 --epochs 50 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | # Hyperparameter Evolution Results 7 | # Generations: 306 8 | # P R mAP.5 mAP.5:.95 box obj cls 9 | # Metrics: 0.6 0.936 0.896 0.684 0.0115 0.00805 0.00146 10 | 11 | lr0: 0.0032 12 | lrf: 0.12 13 | momentum: 0.843 14 | weight_decay: 0.00036 15 | warmup_epochs: 2.0 16 | warmup_momentum: 0.5 17 | warmup_bias_lr: 0.05 18 | box: 0.0296 19 | cls: 0.243 20 | cls_pw: 0.631 21 | obj: 0.301 22 | obj_pw: 0.911 23 | iou_t: 0.2 24 | anchor_t: 2.91 25 | # anchors: 3.63 26 | fl_gamma: 0.0 27 | hsv_h: 0.0138 28 | hsv_s: 0.664 29 | hsv_v: 0.464 30 | degrees: 0.373 31 | translate: 0.245 32 | scale: 0.898 33 | shear: 0.602 34 | perspective: 0.0 35 | flipud: 0.00856 36 | fliplr: 0.5 37 | mosaic: 1.0 38 | mixup: 0.243 39 | -------------------------------------------------------------------------------- /data/hyp.scratch.yaml: -------------------------------------------------------------------------------- 1 | # Hyperparameters for COCO training from scratch 2 | # python train.py --batch 40 --cfg yolov5m.yaml --weights '' --data coco.yaml --img 640 --epochs 300 3 | # See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials 4 | 5 | 6 | lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) 7 | lrf: 0.2 # final OneCycleLR learning rate (lr0 * lrf) 8 | momentum: 0.937 # SGD momentum/Adam beta1 9 | weight_decay: 0.0005 # optimizer weight decay 5e-4 10 | warmup_epochs: 3.0 # warmup epochs (fractions ok) 11 | warmup_momentum: 0.8 # warmup initial momentum 12 | warmup_bias_lr: 0.1 # warmup initial bias lr 13 | box: 0.05 # box loss gain 14 | cls: 0.5 # cls loss gain 15 | cls_pw: 1.0 # cls BCELoss positive_weight 16 | obj: 1.0 # obj loss gain (scale with pixels) 17 | obj_pw: 1.0 # obj BCELoss positive_weight 18 | iou_t: 0.20 # IoU training threshold 19 | anchor_t: 4.0 # anchor-multiple threshold 20 | # anchors: 3 # anchors per output layer (0 to ignore) 21 | fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) 22 | hsv_h: 0.015 # image HSV-Hue augmentation (fraction) 23 | hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) 24 | hsv_v: 0.4 # image HSV-Value augmentation (fraction) 25 | degrees: 0.0 # image rotation (+/- deg) 26 | translate: 0.1 # image translation (+/- fraction) 27 | scale: 0.5 # image scale (+/- gain) 28 | shear: 0.0 # image shear (+/- deg) 29 | perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 30 | flipud: 0.0 # image flip up-down (probability) 31 | fliplr: 0.5 # image flip left-right (probability) 32 | mosaic: 1.0 # image mosaic (probability) 33 | mixup: 0.0 # image mixup (probability) 34 | -------------------------------------------------------------------------------- /data/ship.yaml: -------------------------------------------------------------------------------- 1 | train: ../../ForDRENet/finalDataSet/yolo/train/images/ 2 | val: ../../ForDRENet/finalDataSet/yolo/val/images/ # when test, change "val" to "test" 3 | 4 | # number of classes 5 | nc: 1 6 | 7 | # class names 8 | names: [ 'ship' ] 9 | -------------------------------------------------------------------------------- /detect.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import time 3 | from pathlib import Path 4 | 5 | import cv2 6 | import torch 7 | import torch.backends.cudnn as cudnn 8 | from numpy import random 9 | 10 | from models.experimental import attempt_load 11 | from utils.datasets import LoadStreams, LoadImages 12 | from utils.general import check_img_size, check_requirements, check_imshow, non_max_suppression, apply_classifier, \ 13 | scale_coords, xyxy2xywh, strip_optimizer, set_logging, increment_path 14 | from utils.plots import plot_one_box 15 | from utils.torch_utils import select_device, load_classifier, time_synchronized 16 | 17 | 18 | def detect(save_img=False): 19 | source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size 20 | webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith( 21 | ('rtsp://', 'rtmp://', 'http://')) 22 | 23 | # Directories 24 | save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run 25 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 26 | 27 | # Initialize 28 | set_logging() 29 | device = select_device(opt.device) 30 | half = device.type != 'cpu' # half precision only supported on CUDA 31 | 32 | # Load model 33 | model = attempt_load(weights, map_location=device) # load FP32 model 34 | stride = int(model.stride.max()) # model stride 35 | imgsz = check_img_size(imgsz, s=stride) # check img_size 36 | if half: 37 | model.half() # to FP16 38 | 39 | # Second-stage classifier 40 | classify = False 41 | if classify: 42 | modelc = load_classifier(name='resnet101', n=2) # initialize 43 | modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']).to(device).eval() 44 | 45 | # Set Dataloader 46 | vid_path, vid_writer = None, None 47 | if webcam: 48 | view_img = check_imshow() 49 | cudnn.benchmark = True # set True to speed up constant image size inference 50 | dataset = LoadStreams(source, img_size=imgsz) 51 | else: 52 | save_img = True 53 | dataset = LoadImages(source, img_size=imgsz) 54 | 55 | # Get names and colors 56 | names = model.module.names if hasattr(model, 'module') else model.names 57 | colors = [[random.randint(0, 255) for _ in range(3)] for _ in names] 58 | 59 | # Run inference 60 | if device.type != 'cpu': 61 | model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters()))) # run once 62 | t0 = time.time() 63 | for path, img, im0s, vid_cap in dataset: 64 | img = torch.from_numpy(img).to(device) 65 | img = img.half() if half else img.float() # uint8 to fp16/32 66 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 67 | if img.ndimension() == 3: 68 | img = img.unsqueeze(0) 69 | 70 | # Inference 71 | t1 = time_synchronized() 72 | pred = model(img, augment=opt.augment)[0][0] 73 | 74 | # Apply NMS 75 | pred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes, agnostic=opt.agnostic_nms) 76 | t2 = time_synchronized() 77 | 78 | # Apply Classifier 79 | if classify: 80 | pred = apply_classifier(pred, modelc, img, im0s) 81 | 82 | # Process detections 83 | for i, det in enumerate(pred): # detections per image 84 | if webcam: # batch_size >= 1 85 | p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.count 86 | else: 87 | p, s, im0, frame = path, '', im0s, getattr(dataset, 'frame', 0) 88 | 89 | p = Path(p) # to Path 90 | save_path = str(save_dir / p.name) # img.jpg 91 | txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # img.txt 92 | s += '%gx%g ' % img.shape[2:] # print string 93 | gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh 94 | if len(det): 95 | # Rescale boxes from img_size to im0 size 96 | det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round() 97 | 98 | # Print results 99 | for c in det[:, -1].unique(): 100 | n = (det[:, -1] == c).sum() # detections per class 101 | s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string 102 | 103 | # Write results 104 | for *xyxy, conf, cls in reversed(det): 105 | if save_txt: # Write to file 106 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 107 | line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh) # label format 108 | with open(txt_path + '.txt', 'a') as f: 109 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 110 | 111 | if save_img or view_img: # Add bbox to image 112 | label = f'{names[int(cls)]} {conf:.2f}' 113 | plot_one_box(xyxy, im0, label=label, color=colors[int(cls)], line_thickness=1) 114 | 115 | # Print time (inference + NMS) 116 | print(f'{s}Done. ({t2 - t1:.3f}s)') 117 | 118 | # Stream results 119 | if view_img: 120 | cv2.imshow(str(p), im0) 121 | cv2.waitKey(1) # 1 millisecond 122 | 123 | # Save results (image with detections) 124 | if save_img: 125 | if dataset.mode == 'image': 126 | cv2.imwrite(save_path, im0) 127 | else: # 'video' 128 | if vid_path != save_path: # new video 129 | vid_path = save_path 130 | if isinstance(vid_writer, cv2.VideoWriter): 131 | vid_writer.release() # release previous video writer 132 | 133 | fourcc = 'mp4v' # output video codec 134 | fps = vid_cap.get(cv2.CAP_PROP_FPS) 135 | w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) 136 | h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) 137 | vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (w, h)) 138 | vid_writer.write(im0) 139 | 140 | if save_txt or save_img: 141 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 142 | print(f"Results saved to {save_dir}{s}") 143 | 144 | print(f'Done. ({time.time() - t0:.3f}s)') 145 | 146 | 147 | if __name__ == '__main__': 148 | parser = argparse.ArgumentParser() 149 | parser.add_argument('--weights', nargs='+', type=str, default=r'.\DRENet.pt', help='model.pt path(s)') 150 | parser.add_argument('--source', type=str, default='images/', help='source') # file/folder, 0 for webcam 151 | parser.add_argument('--img-size', type=int, default=512, help='inference size (pixels)') 152 | parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold') 153 | parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS') 154 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 155 | parser.add_argument('--view-img', action='store_true', help='display results') 156 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 157 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 158 | parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3') 159 | parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') 160 | parser.add_argument('--augment', action='store_true', help='augmented inference') 161 | parser.add_argument('--update', action='store_true', help='update all models') 162 | parser.add_argument('--project', default='runs/detect', help='save results to project/name') 163 | parser.add_argument('--name', default='exp', help='save results to project/name') 164 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 165 | opt = parser.parse_args() 166 | print(opt) 167 | check_requirements() 168 | 169 | with torch.no_grad(): 170 | if opt.update: # update all models (to fix SourceChangeWarning) 171 | for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 172 | detect() 173 | strip_optimizer(opt.weights) 174 | else: 175 | detect() 176 | -------------------------------------------------------------------------------- /hubconf.py: -------------------------------------------------------------------------------- 1 | """File for accessing YOLOv5 via PyTorch Hub https://pytorch.org/hub/ 2 | 3 | Usage: 4 | import torch 5 | model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True, channels=3, classes=80) 6 | """ 7 | 8 | from pathlib import Path 9 | 10 | import torch 11 | 12 | from models.yolo import Model 13 | from utils.general import set_logging 14 | from utils.google_utils import attempt_download 15 | 16 | dependencies = ['torch', 'yaml'] 17 | set_logging() 18 | 19 | 20 | def create(name, pretrained, channels, classes, autoshape): 21 | """Creates a specified YOLOv5 model 22 | 23 | Arguments: 24 | name (str): name of model, i.e. 'yolov5s' 25 | pretrained (bool): load pretrained weights into the model 26 | channels (int): number of input channels 27 | classes (int): number of model classes 28 | 29 | Returns: 30 | pytorch model 31 | """ 32 | config = Path(__file__).parent / 'models' / f'{name}.yaml' # model.yaml path 33 | try: 34 | model = Model(config, channels, classes) 35 | if pretrained: 36 | fname = f'{name}.pt' # checkpoint filename 37 | attempt_download(fname) # download if not found locally 38 | ckpt = torch.load(fname, map_location=torch.device('cpu')) # load 39 | state_dict = ckpt['model'].float().state_dict() # to FP32 40 | state_dict = {k: v for k, v in state_dict.items() if model.state_dict()[k].shape == v.shape} # filter 41 | model.load_state_dict(state_dict, strict=False) # load 42 | if len(ckpt['model'].names) == classes: 43 | model.names = ckpt['model'].names # set class names attribute 44 | if autoshape: 45 | model = model.autoshape() # for file/URI/PIL/cv2/np inputs and NMS 46 | return model 47 | 48 | except Exception as e: 49 | help_url = 'https://github.com/ultralytics/yolov5/issues/36' 50 | s = 'Cache maybe be out of date, try force_reload=True. See %s for help.' % help_url 51 | raise Exception(s) from e 52 | 53 | 54 | def yolov5s(pretrained=False, channels=3, classes=80, autoshape=True): 55 | """YOLOv5-small model from https://github.com/ultralytics/yolov5 56 | 57 | Arguments: 58 | pretrained (bool): load pretrained weights into the model, default=False 59 | channels (int): number of input channels, default=3 60 | classes (int): number of model classes, default=80 61 | 62 | Returns: 63 | pytorch model 64 | """ 65 | return create('yolov5s', pretrained, channels, classes, autoshape) 66 | 67 | 68 | def yolov5m(pretrained=False, channels=3, classes=80, autoshape=True): 69 | """YOLOv5-medium model from https://github.com/ultralytics/yolov5 70 | 71 | Arguments: 72 | pretrained (bool): load pretrained weights into the model, default=False 73 | channels (int): number of input channels, default=3 74 | classes (int): number of model classes, default=80 75 | 76 | Returns: 77 | pytorch model 78 | """ 79 | return create('yolov5m', pretrained, channels, classes, autoshape) 80 | 81 | 82 | def yolov5l(pretrained=False, channels=3, classes=80, autoshape=True): 83 | """YOLOv5-large model from https://github.com/ultralytics/yolov5 84 | 85 | Arguments: 86 | pretrained (bool): load pretrained weights into the model, default=False 87 | channels (int): number of input channels, default=3 88 | classes (int): number of model classes, default=80 89 | 90 | Returns: 91 | pytorch model 92 | """ 93 | return create('yolov5l', pretrained, channels, classes, autoshape) 94 | 95 | 96 | def yolov5x(pretrained=False, channels=3, classes=80, autoshape=True): 97 | """YOLOv5-xlarge model from https://github.com/ultralytics/yolov5 98 | 99 | Arguments: 100 | pretrained (bool): load pretrained weights into the model, default=False 101 | channels (int): number of input channels, default=3 102 | classes (int): number of model classes, default=80 103 | 104 | Returns: 105 | pytorch model 106 | """ 107 | return create('yolov5x', pretrained, channels, classes, autoshape) 108 | 109 | 110 | def custom(path_or_model='path/to/model.pt', autoshape=True): 111 | """YOLOv5-custom model from https://github.com/ultralytics/yolov5 112 | 113 | Arguments (3 options): 114 | path_or_model (str): 'path/to/model.pt' 115 | path_or_model (dict): torch.load('path/to/model.pt') 116 | path_or_model (nn.Module): torch.load('path/to/model.pt')['model'] 117 | 118 | Returns: 119 | pytorch model 120 | """ 121 | model = torch.load(path_or_model) if isinstance(path_or_model, str) else path_or_model # load checkpoint 122 | if isinstance(model, dict): 123 | model = model['ema' if model.get('ema') else 'model'] # load model 124 | 125 | hub_model = Model(model.yaml).to(next(model.parameters()).device) # create 126 | hub_model.load_state_dict(model.float().state_dict()) # load state_dict 127 | hub_model.names = model.names # class names 128 | return hub_model.autoshape() if autoshape else hub_model 129 | 130 | 131 | if __name__ == '__main__': 132 | model = create(name='yolov5s', pretrained=True, channels=3, classes=80, autoshape=True) # pretrained example 133 | # model = custom(path_or_model='path/to/model.pt') # custom example 134 | 135 | # Verify inference 136 | import numpy as np 137 | from PIL import Image 138 | 139 | imgs = [Image.open('data/images/bus.jpg'), # PIL 140 | 'data/images/zidane.jpg', # filename 141 | 'https://github.com/ultralytics/yolov5/raw/master/data/images/bus.jpg', # URI 142 | np.zeros((640, 480, 3))] # numpy 143 | 144 | results = model(imgs) # batched inference 145 | results.print() 146 | results.save() 147 | -------------------------------------------------------------------------------- /models/Ablation_Only_CRMA.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3ResAtnMHSA, [256, 64, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/Ablation_Only_CRMA_Remove_PAN.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3ResAtnMHSA, [256, 64, False]], # 17 (P3/8-small) 38 | 39 | [14, 3, C3ResAtnMHSA, [512, 32, False]], # 18 (P4/16-medium) 40 | 41 | [10, 3, C3ResAtnMHSA, [1024, 16, False]], # 19 (P5/32-large) 42 | 43 | [[17, 18, 19], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 44 | ] 45 | -------------------------------------------------------------------------------- /models/Ablation_Only_CRMA_and_Scale.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, ConcatFusionFactor, [1]], # cat backbone P4 32 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, ConcatFusionFactor, [1]], # cat backbone P3 37 | [-1, 3, C3ResAtnMHSA, [256, 64, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, ConcatFusionFactor, [1]], # cat head P4 41 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, ConcatFusionFactor, [1]], # cat head P5 45 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/Ablation_Only_Detector.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, ConcatFusionFactor, [1]], # cat backbone P4 32 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, ConcatFusionFactor, [1]], # cat backbone P3 37 | [-1, 3, C3ResAtnMHSA, [256, 64, False]], # 17 (P3/8-small) 38 | 39 | [14, 3, C3ResAtnMHSA, [512, 32, False]], # 18 (P4/16-medium) 40 | 41 | [10, 3, C3ResAtnMHSA, [1024, 16, False]], # 19 (P5/32-large) 42 | 43 | [[17, 18, 19], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 44 | ] 45 | -------------------------------------------------------------------------------- /models/Ablation_Only_Enhancer.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [4, 1, RCAN, []], 48 | 49 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 50 | ] 51 | -------------------------------------------------------------------------------- /models/Ablation_Only_Scale.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, ConcatFusionFactor, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, ConcatFusionFactor, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, ConcatFusionFactor, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, ConcatFusionFactor, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /models/DRENet.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3ResAtnMHSA, [1024, 16, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, ConcatFusionFactor, [1]], # cat backbone P4 32 | [-1, 3, C3ResAtnMHSA, [512, 32, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, ConcatFusionFactor, [1]], # cat backbone P3 37 | [-1, 3, C3ResAtnMHSA, [256, 64, False]], # 17 (P3/8-small) 38 | 39 | [14, 3, C3ResAtnMHSA, [512, 32, False]], # 18 (P4/16-medium) 40 | 41 | [10, 3, C3ResAtnMHSA, [1024, 16, False]], # 19 (P5/32-large) 42 | 43 | [4, 1, RCAN, []], 44 | [[17, 18, 19], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 45 | ] 46 | -------------------------------------------------------------------------------- /models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/WindVChen/DRENet/a187dbe0f623b521a62c6176c7cafaa7322f5f66/models/__init__.py -------------------------------------------------------------------------------- /models/common.py: -------------------------------------------------------------------------------- 1 | # This file contains modules common to various models 2 | 3 | import math 4 | 5 | import numpy as np 6 | import requests 7 | import torch 8 | import torch.nn as nn 9 | from PIL import Image 10 | 11 | from utils.datasets import letterbox 12 | from utils.general import non_max_suppression, make_divisible, scale_coords, xyxy2xywh 13 | from utils.plots import color_list 14 | 15 | 16 | def autopad(k, p=None): # kernel, padding 17 | # Pad to 'same' 18 | if p is None: 19 | p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad 20 | return p 21 | 22 | 23 | def DWConv(c1, c2, k=1, s=1, act=True): 24 | # Depthwise convolution 25 | return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act) 26 | 27 | class Conv(nn.Module): 28 | # Standard convolution 29 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, bias=False): # ch_in, ch_out, kernel, stride, padding, groups 30 | super(Conv, self).__init__() 31 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) 32 | self.bn = nn.BatchNorm2d(c2) 33 | self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) 34 | 35 | def forward(self, x): 36 | return self.act(self.bn(self.conv(x))) 37 | 38 | def fuseforward(self, x): 39 | return self.act(self.conv(x)) 40 | 41 | class inv(nn.Module): 42 | # Standard convolution 43 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, bias=False): # ch_in, ch_out, kernel, stride, padding, groups 44 | super(inv, self).__init__() 45 | self.INV = False 46 | self.inChannel = c1 47 | if self.inChannel<4 or self.inChannel<16 or not self.INV: 48 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=bias) 49 | self.bn = nn.BatchNorm2d(c2) 50 | self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) 51 | else: 52 | kernel_size = k 53 | stride = s 54 | channels = c1 55 | channelsOut = c2 56 | self.kernel_size = k 57 | self.stride = s 58 | self.channels = c1 59 | reduction_ratio = 4 60 | self.group_channels = 16 61 | self.groups = self.channels // self.group_channels 62 | self.conv1 = nn.Sequential(nn.Conv2d(channels, channels // reduction_ratio, 1, groups=g, bias=bias), 63 | nn.BatchNorm2d(channels // reduction_ratio), 64 | nn.ReLU()) 65 | self.conv2 = nn.Conv2d(channels // reduction_ratio, kernel_size ** 2 * self.groups, 1, groups=g, bias=bias) 66 | if stride > 1: 67 | self.avgpool = nn.AvgPool2d(stride, stride) 68 | self.unfold = nn.Unfold(kernel_size, 1, (kernel_size - 1) // 2, stride) 69 | self.bn = nn.BatchNorm2d(channels) 70 | self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) 71 | self.conv3 = nn.Conv2d(channels, channelsOut, 1, groups=g, bias=bias) 72 | self.bn2 = nn.BatchNorm2d(channelsOut) 73 | self.act2 = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) 74 | 75 | def forward(self, x): 76 | if self.inChannel < 4 or self.inChannel < 16 or not self.INV: 77 | return self.act(self.bn(self.conv(x))) 78 | else: 79 | weight = self.conv2(self.conv1(x if self.stride == 1 else self.avgpool(x))) 80 | b, c, h, w = weight.shape 81 | weight = weight.view(b, self.groups, self.kernel_size ** 2, h, w).unsqueeze(2) 82 | out = self.unfold(x).view(b, self.groups, self.group_channels, self.kernel_size ** 2, h, w) 83 | out = (weight * out).sum(dim=3).view(b, self.channels, h, w) 84 | out = self.act(self.bn(out)) 85 | return self.act2(self.bn2(self.conv3(out))) 86 | def fuseforward(self, x): 87 | if self.inChannel < 4 or self.inChannel < 16 or not self.INV: 88 | return self.act(self.conv(x)) 89 | else: 90 | return self.act2(self.conv3(x)) 91 | 92 | 93 | class Bottleneck(nn.Module): 94 | # Standard bottleneck 95 | def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion 96 | super(Bottleneck, self).__init__() 97 | c_ = int(c2 * e) # hidden channels 98 | self.cv1 = Conv(c1, c_, 1, 1) 99 | self.cv2 = Conv(c_, c2, 3, 1, g=g) 100 | self.add = shortcut and c1 == c2 101 | 102 | def forward(self, x): 103 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 104 | 105 | class BottleneckResAtnMHSA(nn.Module): 106 | # Standard bottleneck 107 | def __init__(self, n_dims, size, shortcut=True): # ch_in, ch_out, shortcut, groups, expansion 108 | super(BottleneckResAtnMHSA, self).__init__() 109 | 110 | height=size 111 | width=size 112 | self.cv1 = Conv(n_dims, n_dims//2, 1, 1) 113 | self.cv2 = Conv(n_dims//2, n_dims, 1, 1) 114 | '''MHSA PARAGRAMS''' 115 | self.query = nn.Conv2d(n_dims//2, n_dims//2, kernel_size=1) 116 | self.key = nn.Conv2d(n_dims//2, n_dims//2, kernel_size=1) 117 | self.value = nn.Conv2d(n_dims//2, n_dims//2, kernel_size=1) 118 | 119 | self.rel_h = nn.Parameter(torch.randn([1, n_dims//2, height, 1]), requires_grad=True) 120 | self.rel_w = nn.Parameter(torch.randn([1, n_dims//2, 1, width]), requires_grad=True) 121 | 122 | self.softmax = nn.Softmax(dim=-1) 123 | self.add = shortcut 124 | 125 | def forward(self, x): 126 | x1=self.cv1(x) 127 | n_batch, C, width, height = x1.size() 128 | q = self.query(x1).view(n_batch, C, -1) 129 | k = self.key(x1).view(n_batch, C, -1) 130 | v = self.value(x1).view(n_batch, C, -1) 131 | 132 | content_content = torch.bmm(q.permute(0, 2, 1), k) 133 | 134 | content_position = (self.rel_h + self.rel_w).view(1, C, -1).permute(0, 2, 1) 135 | 136 | # If you want to use resolution-agnostic positional encoding, you can uncomment the following lines. 137 | # See details in https://github.com/WindVChen/DRENet/issues/10. 138 | # Note that the performance of this resolution-agnostic positional encoding is not tested. 139 | # content_position = (self.rel_h + self.rel_w) 140 | # content_position = nn.functional.interpolate(content_position, (int(content_content.shape[-1]**0.5), int(content_content.shape[-1]**0.5)), mode='bilinear') 141 | # content_position = content_position.view(1, C, -1).permute(0, 2, 1) 142 | 143 | content_position = torch.matmul(content_position, q) 144 | 145 | energy = content_content + content_position 146 | attention = self.softmax(energy) 147 | 148 | out = torch.bmm(v, attention.permute(0, 2, 1)) 149 | out = out.view(n_batch, C, width, height) 150 | 151 | 152 | return x + self.cv2(out) if self.add else self.cv2(out) 153 | 154 | 155 | class BottleneckCSP(nn.Module): 156 | # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks 157 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 158 | super(BottleneckCSP, self).__init__() 159 | c_ = int(c2 * e) # hidden channels 160 | self.cv1 = Conv(c1, c_, 1, 1) 161 | self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) 162 | self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) 163 | self.cv4 = Conv(2 * c_, c2, 1, 1) 164 | self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) 165 | self.act = nn.LeakyReLU(0.1, inplace=True) 166 | self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) 167 | 168 | def forward(self, x): 169 | y1 = self.cv3(self.m(self.cv1(x))) 170 | y2 = self.cv2(x) 171 | return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1)))) 172 | 173 | 174 | class C3(nn.Module): 175 | # CSP Bottleneck with 3 convolutions 176 | def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 177 | super(C3, self).__init__() 178 | c_ = int(c2 * e) # hidden channels 179 | self.cv1 = Conv(c1, c_, 1, 1) 180 | self.cv2 = Conv(c1, c_, 1, 1) 181 | self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2) 182 | self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)]) 183 | # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)]) 184 | 185 | def forward(self, x): 186 | return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1)) 187 | 188 | class C3ResAtnMHSA(nn.Module): 189 | # CSP Bottleneck with 3 convolutions 190 | def __init__(self, c1, c2, n=1, size=14, shortcut=True, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion 191 | super(C3ResAtnMHSA, self).__init__() 192 | c_ = int(c2 * e) # hidden channels 193 | self.cv1 = Conv(c1, c_, 1, 1) 194 | self.cv2 = Conv(c1, c_, 1, 1) 195 | self.cv3 = Conv(2 * c_, c2, 1) # act=FReLU(c2) 196 | self.m = nn.Sequential(*[BottleneckResAtnMHSA(c_, size, shortcut=True) for _ in range(n)]) 197 | # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)]) 198 | 199 | def forward(self, x): 200 | return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1)) 201 | 202 | 203 | class SPP(nn.Module): 204 | # Spatial pyramid pooling layer used in YOLOv3-SPP 205 | def __init__(self, c1, c2, k=(5, 9, 13)): 206 | super(SPP, self).__init__() 207 | c_ = c1 // 2 # hidden channels 208 | self.cv1 = Conv(c1, c_, 1, 1) 209 | self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) 210 | self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) 211 | 212 | def forward(self, x): 213 | x = self.cv1(x) 214 | return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) 215 | 216 | 217 | class Focus(nn.Module): 218 | # Focus wh information into c-space 219 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups 220 | super(Focus, self).__init__() 221 | self.conv = Conv(c1 * 4, c2, k, s, p, g, act) 222 | # self.contract = Contract(gain=2) 223 | 224 | def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) 225 | return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)) 226 | # return self.conv(self.contract(x)) 227 | 228 | 229 | class Contract(nn.Module): 230 | # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40) 231 | def __init__(self, gain=2): 232 | super().__init__() 233 | self.gain = gain 234 | 235 | def forward(self, x): 236 | N, C, H, W = x.size() # assert (H / s == 0) and (W / s == 0), 'Indivisible gain' 237 | s = self.gain 238 | x = x.view(N, C, H // s, s, W // s, s) # x(1,64,40,2,40,2) 239 | x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40) 240 | return x.view(N, C * s * s, H // s, W // s) # x(1,256,40,40) 241 | 242 | 243 | class Expand(nn.Module): 244 | # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160) 245 | def __init__(self, gain=2): 246 | super().__init__() 247 | self.gain = gain 248 | 249 | def forward(self, x): 250 | N, C, H, W = x.size() # assert C / s ** 2 == 0, 'Indivisible gain' 251 | s = self.gain 252 | x = x.view(N, s, s, C // s ** 2, H, W) # x(1,2,2,16,80,80) 253 | x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2) 254 | return x.view(N, C // s ** 2, H * s, W * s) # x(1,16,160,160) 255 | 256 | 257 | class Concat(nn.Module): 258 | # Concatenate a list of tensors along dimension 259 | def __init__(self, dimension=1, selectPos=None): 260 | super(Concat, self).__init__() 261 | self.d = dimension 262 | self.p=selectPos 263 | 264 | def forward(self, x): 265 | if isinstance(self.p, int): 266 | return torch.cat([x[0][self.p],x[1]], self.d) 267 | else: 268 | return torch.cat(x, self.d) 269 | 270 | class ConcatFusionFactor(nn.Module): 271 | # Concatenate a list of tensors along dimension 272 | def __init__(self, dimension=1): 273 | super(ConcatFusionFactor, self).__init__() 274 | self.d = dimension 275 | self.factor=torch.nn.Parameter(torch.FloatTensor([1])) 276 | 277 | def forward(self, x): 278 | x[0]=x[0]*self.factor 279 | return torch.cat(x, self.d) 280 | 281 | 282 | class NMS(nn.Module): 283 | # Non-Maximum Suppression (NMS) module 284 | conf = 0.25 # confidence threshold 285 | iou = 0.45 # IoU threshold 286 | classes = None # (optional list) filter by class 287 | 288 | def __init__(self): 289 | super(NMS, self).__init__() 290 | 291 | def forward(self, x): 292 | return non_max_suppression(x[0], conf_thres=self.conf, iou_thres=self.iou, classes=self.classes) 293 | 294 | 295 | class autoShape(nn.Module): 296 | # input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS 297 | img_size = 640 # inference size (pixels) 298 | conf = 0.25 # NMS confidence threshold 299 | iou = 0.45 # NMS IoU threshold 300 | classes = None # (optional list) filter by class 301 | 302 | def __init__(self, model): 303 | super(autoShape, self).__init__() 304 | self.model = model.eval() 305 | 306 | def autoshape(self): 307 | print('autoShape already enabled, skipping... ') # model already converted to model.autoshape() 308 | return self 309 | 310 | def forward(self, imgs, size=640, augment=False, profile=False): 311 | # Inference from various sources. For height=720, width=1280, RGB images example inputs are: 312 | # filename: imgs = 'data/samples/zidane.jpg' 313 | # URI: = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/zidane.jpg' 314 | # OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(720,1280,3) 315 | # PIL: = Image.open('image.jpg') # HWC x(720,1280,3) 316 | # numpy: = np.zeros((720,1280,3)) # HWC 317 | # torch: = torch.zeros(16,3,720,1280) # BCHW 318 | # multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images 319 | 320 | p = next(self.model.parameters()) # for device and type 321 | if isinstance(imgs, torch.Tensor): # torch 322 | return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference 323 | 324 | # Pre-process 325 | n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs]) # number of images, list of images 326 | shape0, shape1 = [], [] # image and inference shapes 327 | for i, im in enumerate(imgs): 328 | if isinstance(im, str): # filename or uri 329 | im = Image.open(requests.get(im, stream=True).raw if im.startswith('http') else im) # open 330 | im = np.array(im) # to numpy 331 | if im.shape[0] < 5: # image in CHW 332 | im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1) 333 | im = im[:, :, :3] if im.ndim == 3 else np.tile(im[:, :, None], 3) # enforce 3ch input 334 | s = im.shape[:2] # HWC 335 | shape0.append(s) # image shape 336 | g = (size / max(s)) # gain 337 | shape1.append([y * g for y in s]) 338 | imgs[i] = im # update 339 | shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)] # inference shape 340 | x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs] # pad 341 | x = np.stack(x, 0) if n > 1 else x[0][None] # stack 342 | x = np.ascontiguousarray(x.transpose((0, 3, 1, 2))) # BHWC to BCHW 343 | x = torch.from_numpy(x).to(p.device).type_as(p) / 255. # uint8 to fp16/32 344 | 345 | # Inference 346 | with torch.no_grad(): 347 | y = self.model(x, augment, profile)[0] # forward 348 | y = non_max_suppression(y, conf_thres=self.conf, iou_thres=self.iou, classes=self.classes) # NMS 349 | 350 | # Post-process 351 | for i in range(n): 352 | scale_coords(shape1, y[i][:, :4], shape0[i]) 353 | 354 | return Detections(imgs, y, self.names) 355 | 356 | 357 | class Detections: 358 | # detections class for YOLOv5 inference results 359 | def __init__(self, imgs, pred, names=None): 360 | super(Detections, self).__init__() 361 | d = pred[0].device # device 362 | gn = [torch.tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.], device=d) for im in imgs] # normalizations 363 | self.imgs = imgs # list of images as numpy arrays 364 | self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) 365 | self.names = names # class names 366 | self.xyxy = pred # xyxy pixels 367 | self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels 368 | self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized 369 | self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized 370 | self.n = len(self.pred) 371 | 372 | def display(self, pprint=False, show=False, save=False, render=False): 373 | colors = color_list() 374 | for i, (img, pred) in enumerate(zip(self.imgs, self.pred)): 375 | str = f'Image {i + 1}/{len(self.pred)}: {img.shape[0]}x{img.shape[1]} ' 376 | if pred is not None: 377 | for c in pred[:, -1].unique(): 378 | n = (pred[:, -1] == c).sum() # detections per class 379 | str += f'{n} {self.names[int(c)]}s, ' # add to string 380 | if show or save or render: 381 | img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img # from np 382 | for *box, conf, cls in pred: # xyxy, confidence, class 383 | # str += '%s %.2f, ' % (names[int(cls)], conf) # label 384 | ImageDraw.Draw(img).rectangle(box, width=4, outline=colors[int(cls) % 10]) # plot 385 | if pprint: 386 | print(str) 387 | if show: 388 | img.show(f'Image {i}') # show 389 | if save: 390 | f = f'results{i}.jpg' 391 | str += f"saved to '{f}'" 392 | img.save(f) # save 393 | if render: 394 | self.imgs[i] = np.asarray(img) 395 | 396 | def print(self): 397 | self.display(pprint=True) # print results 398 | 399 | def show(self): 400 | self.display(show=True) # show results 401 | 402 | def save(self): 403 | self.display(save=True) # save results 404 | 405 | def render(self): 406 | self.display(render=True) # render results 407 | return self.imgs 408 | 409 | def __len__(self): 410 | return self.n 411 | 412 | def tolist(self): 413 | # return a list of Detections objects, i.e. 'for result in results.tolist():' 414 | x = [Detections([self.imgs[i]], [self.pred[i]], self.names) for i in range(self.n)] 415 | for d in x: 416 | for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']: 417 | setattr(d, k, getattr(d, k)[0]) # pop out of list 418 | return x 419 | 420 | 421 | class Classify(nn.Module): 422 | # Classification head, i.e. x(b,c1,20,20) to x(b,c2) 423 | def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups 424 | super(Classify, self).__init__() 425 | self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1) 426 | self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1) 427 | self.flat = nn.Flatten() 428 | 429 | def forward(self, x): 430 | z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list 431 | return self.flat(self.conv(z)) # flatten to x(b,c2) 432 | 433 | class LocalReconstruct(nn.Module): 434 | def __init__(self, c1, c2): 435 | super(LocalReconstruct, self).__init__() 436 | self.reconstruct = nn.Sequential( 437 | Conv(c1, c2, 1, 1), 438 | Conv(c2, c2//4, 3, 1), 439 | Conv(c2//4, c2, 1, 1) 440 | ) 441 | 442 | def forward(self, x): 443 | x0=self.reconstruct(x[0]) 444 | x1=self.reconstruct(x[1]) 445 | x2=self.reconstruct(x[2]) 446 | return x0,x1,x2 447 | 448 | class SEAtn(nn.Module): 449 | def __init__(self, channel, reduction=16): 450 | super(SEAtn, self).__init__() 451 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 452 | self.fc = nn.Sequential( 453 | nn.Linear(channel, channel // reduction, bias=False), 454 | nn.ReLU(inplace=True), 455 | nn.Linear(channel // reduction, channel, bias=False), 456 | nn.Sigmoid() 457 | ) 458 | 459 | def forward(self, x): 460 | b, c, _, _ = x.size() 461 | y = self.avg_pool(x).view(b, c) 462 | y = self.fc(y).view(b, c, 1, 1) 463 | return y 464 | 465 | class AtnMut(nn.Module): 466 | def __init__(self, start, end): 467 | super(AtnMut, self).__init__() 468 | self.start=start 469 | self.end=end 470 | 471 | def forward(self, x): 472 | atn= x[0][:, self.start:self.end, :, :] 473 | obj= x[1] 474 | out= obj*atn.expand_as(obj) 475 | return out 476 | 477 | class MHSA(nn.Module): 478 | def __init__(self, n_dims, size): 479 | super(MHSA, self).__init__() 480 | 481 | height = size 482 | width = size 483 | self.query = nn.Conv2d(n_dims, n_dims, kernel_size=1) 484 | self.key = nn.Conv2d(n_dims, n_dims, kernel_size=1) 485 | self.value = nn.Conv2d(n_dims, n_dims, kernel_size=1) 486 | 487 | self.rel_h = nn.Parameter(torch.randn([1, n_dims, height, 1]), requires_grad=True) 488 | self.rel_w = nn.Parameter(torch.randn([1, n_dims, 1, width]), requires_grad=True) 489 | 490 | self.softmax = nn.Softmax(dim=-1) 491 | 492 | def forward(self, x): 493 | n_batch, C, width, height = x.size() 494 | q = self.query(x).view(n_batch, C, -1) 495 | k = self.key(x).view(n_batch, C, -1) 496 | v = self.value(x).view(n_batch, C, -1) 497 | 498 | content_content = torch.bmm(q.permute(0, 2, 1), k) 499 | 500 | content_position = (self.rel_h + self.rel_w).view(1, C, -1).permute(0, 2, 1) 501 | content_position = torch.matmul(content_position, q) 502 | 503 | energy = content_content + content_position 504 | attention = self.softmax(energy) 505 | 506 | out = torch.bmm(v, attention.permute(0, 2, 1)) 507 | out = out.view(n_batch, C, width, height) 508 | 509 | return out 510 | 511 | ## Residual Channel Attention Network (RCAN) 512 | class RCAN(nn.Module): 513 | def __init__(self, c1, conv=Conv): 514 | super(RCAN, self).__init__() 515 | 516 | n_resgroups = 1 517 | n_resblocks = 1 518 | n_feats = c1 519 | kernel_size = 3 520 | reduction = 16 521 | scale = 2 522 | act = nn.SiLU() 523 | 524 | # define body module 525 | modules_body = [ 526 | ResidualGroup( 527 | conv, n_feats, kernel_size, reduction, act=act, res_scale=1, n_resblocks=n_resblocks) \ 528 | for _ in range(n_resgroups)] 529 | 530 | modules_body.append(conv(n_feats, n_feats, kernel_size)) 531 | 532 | # define tail module 533 | modules_tail = [ 534 | Upsampler(conv, scale, n_feats, act=False), 535 | conv(n_feats, 3, kernel_size)] 536 | 537 | self.body = nn.Sequential(*modules_body) 538 | self.tail = nn.Sequential(*modules_tail) 539 | 540 | def forward(self, x): 541 | res = self.body(x) 542 | res += x 543 | 544 | x = self.tail(res) 545 | 546 | return x 547 | 548 | 549 | class CALayer(nn.Module): 550 | def __init__(self, channel, reduction=16): 551 | super(CALayer, self).__init__() 552 | # global average pooling: feature --> point 553 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 554 | # feature channel downscale and upscale --> channel weight 555 | self.conv_du = nn.Sequential( 556 | nn.Conv2d(channel, channel // reduction, 1, padding=0, bias=True), 557 | nn.ReLU(inplace=True), 558 | nn.Conv2d(channel // reduction, channel, 1, padding=0, bias=True), 559 | nn.Sigmoid() 560 | ) 561 | 562 | def forward(self, x): 563 | y = self.avg_pool(x) 564 | y = self.conv_du(y) 565 | return x * y 566 | 567 | 568 | ## Residual Channel Attention Block (RCAB) 569 | class RCAB(nn.Module): 570 | def __init__( 571 | self, conv, n_feat, kernel_size, reduction, 572 | bias=True, bn=False, act=nn.ReLU(True), res_scale=1): 573 | 574 | super(RCAB, self).__init__() 575 | modules_body = [] 576 | for i in range(2): 577 | modules_body.append(conv(n_feat, n_feat, kernel_size, bias=bias)) 578 | if bn: modules_body.append(nn.BatchNorm2d(n_feat)) 579 | if i == 0: modules_body.append(act) 580 | modules_body.append(CALayer(n_feat, reduction)) 581 | self.body = nn.Sequential(*modules_body) 582 | self.res_scale = res_scale 583 | 584 | def forward(self, x): 585 | res = self.body(x) 586 | # res = self.body(x).mul(self.res_scale) 587 | res += x 588 | return res 589 | 590 | 591 | ## Residual Group (RG) 592 | class ResidualGroup(nn.Module): 593 | def __init__(self, conv, n_feat, kernel_size, reduction, act, res_scale, n_resblocks): 594 | super(ResidualGroup, self).__init__() 595 | modules_body = [] 596 | modules_body = [ 597 | RCAB( 598 | conv, n_feat, kernel_size, reduction, bias=True, bn=False, act=nn.ReLU(True), res_scale=1) \ 599 | for _ in range(n_resblocks)] 600 | modules_body.append(conv(n_feat, n_feat, kernel_size)) 601 | self.body = nn.Sequential(*modules_body) 602 | 603 | def forward(self, x): 604 | res = self.body(x) 605 | res += x 606 | return res 607 | 608 | class Upsampler(nn.Sequential): 609 | def __init__(self, conv, scale, n_feat, bn=False, act=False, bias=True): 610 | 611 | m = [] 612 | for _ in range(int(math.log(scale, 2))): 613 | m.append(conv(n_feat, 4 * n_feat, 3, bias = bias)) 614 | m.append(nn.PixelShuffle(2)) 615 | if bn: m.append(nn.BatchNorm2d(n_feat)) 616 | if act: m.append(act()) 617 | 618 | super(Upsampler, self).__init__(*m) -------------------------------------------------------------------------------- /models/experimental.py: -------------------------------------------------------------------------------- 1 | # This file contains experimental modules 2 | 3 | import numpy as np 4 | import torch 5 | import torch.nn as nn 6 | 7 | from models.common import Conv, DWConv, RCAN 8 | from utils.google_utils import attempt_download 9 | 10 | 11 | class CrossConv(nn.Module): 12 | # Cross Convolution Downsample 13 | def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): 14 | # ch_in, ch_out, kernel, stride, groups, expansion, shortcut 15 | super(CrossConv, self).__init__() 16 | c_ = int(c2 * e) # hidden channels 17 | self.cv1 = Conv(c1, c_, (1, k), (1, s)) 18 | self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) 19 | self.add = shortcut and c1 == c2 20 | 21 | def forward(self, x): 22 | return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) 23 | 24 | 25 | class Sum(nn.Module): 26 | # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 27 | def __init__(self, n, weight=False): # n: number of inputs 28 | super(Sum, self).__init__() 29 | self.weight = weight # apply weights boolean 30 | self.iter = range(n - 1) # iter object 31 | if weight: 32 | self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights 33 | 34 | def forward(self, x): 35 | y = x[0] # no weight 36 | if self.weight: 37 | w = torch.sigmoid(self.w) * 2 38 | for i in self.iter: 39 | y = y + x[i + 1] * w[i] 40 | else: 41 | for i in self.iter: 42 | y = y + x[i + 1] 43 | return y 44 | 45 | 46 | class GhostConv(nn.Module): 47 | # Ghost Convolution https://github.com/huawei-noah/ghostnet 48 | def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups 49 | super(GhostConv, self).__init__() 50 | c_ = c2 // 2 # hidden channels 51 | self.cv1 = Conv(c1, c_, k, s, None, g, act) 52 | self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) 53 | 54 | def forward(self, x): 55 | y = self.cv1(x) 56 | return torch.cat([y, self.cv2(y)], 1) 57 | 58 | 59 | class GhostBottleneck(nn.Module): 60 | # Ghost Bottleneck https://github.com/huawei-noah/ghostnet 61 | def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride 62 | super(GhostBottleneck, self).__init__() 63 | c_ = c2 // 2 64 | self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw 65 | DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw 66 | GhostConv(c_, c2, 1, 1, act=False)) # pw-linear 67 | self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), 68 | Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity() 69 | 70 | def forward(self, x): 71 | return self.conv(x) + self.shortcut(x) 72 | 73 | 74 | class MixConv2d(nn.Module): 75 | # Mixed Depthwise Conv https://arxiv.org/abs/1907.09595 76 | def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): 77 | super(MixConv2d, self).__init__() 78 | groups = len(k) 79 | if equal_ch: # equal c_ per group 80 | i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices 81 | c_ = [(i == g).sum() for g in range(groups)] # intermediate channels 82 | else: # equal weight.numel() per group 83 | b = [c2] + [0] * groups 84 | a = np.eye(groups + 1, groups, k=-1) 85 | a -= np.roll(a, 1, axis=1) 86 | a *= np.array(k) ** 2 87 | a[0] = 1 88 | c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b 89 | 90 | self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)]) 91 | self.bn = nn.BatchNorm2d(c2) 92 | self.act = nn.LeakyReLU(0.1, inplace=True) 93 | 94 | def forward(self, x): 95 | return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) 96 | 97 | 98 | class Ensemble(nn.ModuleList): 99 | # Ensemble of models 100 | def __init__(self): 101 | super(Ensemble, self).__init__() 102 | 103 | def forward(self, x, augment=False): 104 | y = [] 105 | for module in self: 106 | y.append(module(x, augment)[0]) 107 | # y = torch.stack(y).max(0)[0] # max ensemble 108 | # y = torch.stack(y).mean(0) # mean ensemble 109 | y = torch.cat(y, 1) # nms ensemble 110 | return y, None # inference, train output 111 | 112 | 113 | def attempt_load(weights, map_location=None): 114 | # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a 115 | model = Ensemble() 116 | for w in weights if isinstance(weights, list) else [weights]: 117 | attempt_download(w) 118 | ckpt = torch.load(w, map_location=map_location) # load 119 | model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model 120 | 121 | # Compatibility updates 122 | for m in model.modules(): 123 | if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: 124 | m.inplace = True # pytorch 1.7.0 compatibility 125 | elif type(m) is Conv: 126 | m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility 127 | 128 | if len(model) == 1: 129 | return model[-1] # return model 130 | else: 131 | print('Ensemble created with %s\n' % weights) 132 | for k in ['names', 'stride']: 133 | setattr(model, k, getattr(model[-1], k)) 134 | return model # return ensemble 135 | -------------------------------------------------------------------------------- /models/export.py: -------------------------------------------------------------------------------- 1 | """Exports a YOLOv5 *.pt model to ONNX and TorchScript formats 2 | 3 | Usage: 4 | $ export PYTHONPATH="$PWD" && python models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1 5 | """ 6 | 7 | import argparse 8 | import sys 9 | import time 10 | 11 | sys.path.append('./') # to run '$ python *.py' files in subdirectories 12 | 13 | import torch 14 | import torch.nn as nn 15 | 16 | import models 17 | from models.experimental import attempt_load 18 | from utils.activations import Hardswish, SiLU 19 | from utils.general import set_logging, check_img_size 20 | 21 | if __name__ == '__main__': 22 | parser = argparse.ArgumentParser() 23 | parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path') # from yolov5/models/ 24 | parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size') # height, width 25 | parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes') 26 | parser.add_argument('--batch-size', type=int, default=1, help='batch size') 27 | opt = parser.parse_args() 28 | opt.img_size *= 2 if len(opt.img_size) == 1 else 1 # expand 29 | print(opt) 30 | set_logging() 31 | t = time.time() 32 | 33 | # Load PyTorch model 34 | model = attempt_load(opt.weights, map_location=torch.device('cpu')) # load FP32 model 35 | labels = model.names 36 | 37 | # Checks 38 | gs = int(max(model.stride)) # grid size (max stride) 39 | opt.img_size = [check_img_size(x, gs) for x in opt.img_size] # verify img_size are gs-multiples 40 | 41 | # Input 42 | img = torch.zeros(opt.batch_size, 3, *opt.img_size) # image size(1,3,320,192) iDetection 43 | 44 | # Update model 45 | for k, m in model.named_modules(): 46 | m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility 47 | if isinstance(m, models.common.Conv): # assign export-friendly activations 48 | if isinstance(m.act, nn.Hardswish): 49 | m.act = Hardswish() 50 | elif isinstance(m.act, nn.SiLU): 51 | m.act = SiLU() 52 | # elif isinstance(m, models.yolo.Detect): 53 | # m.forward = m.forward_export # assign forward (optional) 54 | model.model[-1].export = True # set Detect() layer export=True 55 | y = model(img) # dry run 56 | 57 | # TorchScript export 58 | try: 59 | print('\nStarting TorchScript export with torch %s...' % torch.__version__) 60 | f = opt.weights.replace('.pt', '.torchscript.pt') # filename 61 | ts = torch.jit.trace(model, img) 62 | ts.save(f) 63 | print('TorchScript export success, saved as %s' % f) 64 | except Exception as e: 65 | print('TorchScript export failure: %s' % e) 66 | 67 | # ONNX export 68 | try: 69 | import onnx 70 | 71 | print('\nStarting ONNX export with onnx %s...' % onnx.__version__) 72 | f = opt.weights.replace('.pt', '.onnx') # filename 73 | torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'], 74 | output_names=['classes', 'boxes'] if y is None else ['output'], 75 | dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'}, # size(1,3,640,640) 76 | 'output': {0: 'batch', 2: 'y', 3: 'x'}} if opt.dynamic else None) 77 | 78 | # Checks 79 | onnx_model = onnx.load(f) # load onnx model 80 | onnx.checker.check_model(onnx_model) # check onnx model 81 | # print(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable model 82 | print('ONNX export success, saved as %s' % f) 83 | except Exception as e: 84 | print('ONNX export failure: %s' % e) 85 | 86 | # CoreML export 87 | try: 88 | import coremltools as ct 89 | 90 | print('\nStarting CoreML export with coremltools %s...' % ct.__version__) 91 | # convert model from torchscript and apply pixel scaling as per detect.py 92 | model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])]) 93 | f = opt.weights.replace('.pt', '.mlmodel') # filename 94 | model.save(f) 95 | print('CoreML export success, saved as %s' % f) 96 | except Exception as e: 97 | print('CoreML export failure: %s' % e) 98 | 99 | # Finish 100 | print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t)) 101 | -------------------------------------------------------------------------------- /models/yolo.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import logging 3 | import sys 4 | from copy import deepcopy 5 | from pathlib import Path 6 | 7 | sys.path.append('./') # to run '$ python *.py' files in subdirectories 8 | logger = logging.getLogger(__name__) 9 | 10 | from models.common import * 11 | from models.experimental import MixConv2d, CrossConv 12 | from utils.autoanchor import check_anchor_order 13 | from utils.general import make_divisible, check_file, set_logging 14 | from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \ 15 | select_device, copy_attr 16 | 17 | try: 18 | import thop # for FLOPS computation 19 | except ImportError: 20 | thop = None 21 | 22 | 23 | class Detect(nn.Module): 24 | stride = None # strides computed during build 25 | export = False # onnx export 26 | 27 | def __init__(self, nc=80, anchors=(), ch=()): # detection layer 28 | super(Detect, self).__init__() 29 | self.nc = nc # number of classes 30 | self.no = nc + 5 # number of outputs per anchor 31 | self.nl = len(anchors) # number of detection layers 32 | self.na = len(anchors[0]) // 2 # number of anchors 33 | self.grid = [torch.zeros(1)] * self.nl # init grid 34 | a = torch.tensor(anchors).float().view(self.nl, -1, 2) 35 | self.register_buffer('anchors', a) # shape(nl,na,2) 36 | self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2) 37 | # self.m0=nn.ModuleList(MHSA(x, y, y) for x,y in zip(ch,[32,16,8])) 38 | # self.m1=nn.ModuleList(nn.BatchNorm2d(x) for x in ch) 39 | # self.m2 = nn.ModuleList(nn.Sigmoid() for _ in ch) 40 | self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv 41 | 42 | def forward(self, x): 43 | # x = x.copy() # for profiling 44 | z = [] # inference output 45 | self.training |= self.export 46 | for i in range(self.nl): 47 | # x[i] = self.m0[i](x[i]) + x[i] 48 | # x[i] = self.m1[i](x[i]) 49 | # x[i] = self.m2[i](x[i]) 50 | x[i] = self.m[i](x[i]) # conv 51 | bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) 52 | x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() 53 | 54 | if not self.training: # inference 55 | if self.grid[i].shape[2:4] != x[i].shape[2:4]: 56 | self.grid[i] = self._make_grid(nx, ny).to(x[i].device) 57 | 58 | y = x[i].sigmoid() 59 | y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy 60 | y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh 61 | z.append(y.view(bs, -1, self.no)) 62 | 63 | return x if self.training else (torch.cat(z, 1), x) 64 | 65 | @staticmethod 66 | def _make_grid(nx=20, ny=20): 67 | yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) 68 | return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() 69 | 70 | 71 | class Model(nn.Module): 72 | def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes 73 | super(Model, self).__init__() 74 | 75 | if isinstance(cfg, dict): 76 | self.yaml = cfg # model dict 77 | else: # is *.yaml 78 | import yaml # for torch hub 79 | self.yaml_file = Path(cfg).name 80 | with open(cfg) as f: 81 | self.yaml = yaml.load(f, Loader=yaml.SafeLoader) # model dict 82 | 83 | # Define model 84 | ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels 85 | if nc and nc != self.yaml['nc']: 86 | logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") 87 | self.yaml['nc'] = nc # override yaml value 88 | if anchors: 89 | logger.info(f'Overriding model.yaml anchors with anchors={anchors}') 90 | self.yaml['anchors'] = round(anchors) # override yaml value 91 | self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist 92 | self.names = [str(i) for i in range(self.yaml['nc'])] # default names 93 | # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))]) 94 | 95 | # Build strides, anchors 96 | m = self.model[-1] # Detect() 97 | if isinstance(m, Detect): 98 | s = 512 # 2x min stride 99 | m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))[0]]) # forward 100 | m.anchors /= m.stride.view(-1, 1, 1) 101 | check_anchor_order(m) 102 | self.stride = m.stride 103 | self._initialize_biases() # only run once 104 | # print('Strides: %s' % m.stride.tolist()) 105 | 106 | # Init weights, biases 107 | initialize_weights(self) 108 | self.info() 109 | logger.info('') 110 | 111 | 112 | 113 | def forward(self, x, augment=False, profile=False): 114 | if augment: 115 | img_size = x.shape[-2:] # height, width 116 | s = [1, 0.83, 0.67] # scales 117 | f = [None, 3, None] # flips (2-ud, 3-lr) 118 | y = [] # outputs 119 | for si, fi in zip(s, f): 120 | xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) 121 | yi = self.forward_once(xi)[0][0] # forward 122 | # cv2.imwrite('img%g.jpg' % s, 255 * xi[0].numpy().transpose((1, 2, 0))[:, :, ::-1]) # save 123 | yi[..., :4] /= si # de-scale 124 | if fi == 2: 125 | yi[..., 1] = img_size[0] - yi[..., 1] # de-flip ud 126 | elif fi == 3: 127 | yi[..., 0] = img_size[1] - yi[..., 0] # de-flip lr 128 | y.append(yi) 129 | return torch.cat(y, 1), None # augmented inference, train 130 | else: 131 | return self.forward_once(x, profile) # single-scale inference, train 132 | 133 | def forward_once(self, x, profile=False): 134 | multi = None 135 | y, dt = [], [] # outputs 136 | for m in self.model: 137 | if m.f != -1: # if not from previous layer 138 | x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers 139 | 140 | if profile: 141 | o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS 142 | t = time_synchronized() 143 | for _ in range(10): 144 | _ = m(x) 145 | dt.append((time_synchronized() - t) * 100) 146 | print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type)) 147 | 148 | x = m(x) # run 149 | 150 | if 'RCAN' in m.type: 151 | multi = x 152 | 153 | y.append(x if m.i in self.save else None) # save output 154 | 155 | if profile: 156 | print('%.1fms total' % sum(dt)) 157 | return x, multi 158 | 159 | def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency 160 | # https://arxiv.org/abs/1708.02002 section 3.3 161 | # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. 162 | m = self.model[-1] # Detect() module 163 | for mi, s in zip(m.m, m.stride): # from 164 | b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85) 165 | b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) 166 | b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls 167 | mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) 168 | 169 | def _print_biases(self): 170 | m = self.model[-1] # Detect() module 171 | for mi in m.m: # from 172 | b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85) 173 | print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean())) 174 | 175 | # def _print_weights(self): 176 | # for m in self.model.modules(): 177 | # if type(m) is Bottleneck: 178 | # print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights 179 | 180 | def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers 181 | print('Fusing layers... ') 182 | for m in self.model.modules(): 183 | if type(m) is Conv and hasattr(m, 'bn'): 184 | m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv 185 | delattr(m, 'bn') # remove batchnorm 186 | m.forward = m.fuseforward # update forward 187 | elif type(m) is RCAN: 188 | m.body = nn.Sequential() 189 | m.tail = nn.Sequential() 190 | self.info() 191 | return self 192 | 193 | def nms(self, mode=True): # add or remove NMS module 194 | present = type(self.model[-1]) is NMS # last layer is NMS 195 | if mode and not present: 196 | print('Adding NMS... ') 197 | m = NMS() # module 198 | m.f = -1 # from 199 | m.i = self.model[-1].i + 1 # index 200 | self.model.add_module(name='%s' % m.i, module=m) # add 201 | self.eval() 202 | elif not mode and present: 203 | print('Removing NMS... ') 204 | self.model = self.model[:-1] # remove 205 | return self 206 | 207 | def autoshape(self): # add autoShape module 208 | print('Adding autoShape... ') 209 | m = autoShape(self) # wrap model 210 | copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=()) # copy attributes 211 | return m 212 | 213 | def info(self, verbose=False, img_size=512): # print model information 214 | model_info(self, verbose, img_size) 215 | 216 | 217 | def parse_model(d, ch): # model_dict, input_channels(3) 218 | logger.info('\n%3s%18s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments')) 219 | anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] 220 | na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors 221 | no = na * (nc + 5) # number of outputs = anchors * (classes + 5) 222 | 223 | layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out 224 | for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args 225 | m = eval(m) if isinstance(m, str) else m # eval strings 226 | for j, a in enumerate(args): 227 | try: 228 | args[j] = eval(a) if isinstance(a, str) else a # eval strings 229 | except: 230 | pass 231 | 232 | n = max(round(n * gd), 1) if n > 1 else n # depth gain 233 | if m in [Conv, Bottleneck, SPP, DWConv, MixConv2d, Focus, CrossConv, BottleneckCSP, C3, C3ResAtnMHSA]: 234 | if f<0: 235 | c1, c2 = ch[f], args[0] 236 | else: 237 | c1, c2 = ch[f+1], args[0] 238 | 239 | # Normal 240 | # if i > 0 and args[0] != no: # channel expansion factor 241 | # ex = 1.75 # exponential (default 2.0) 242 | # e = math.log(c2 / ch[1]) / math.log(2) 243 | # c2 = int(ch[1] * ex ** e) 244 | # if m != Focus: 245 | 246 | c2 = make_divisible(c2 * gw, 8) if c2 != no else c2 247 | 248 | # Experimental 249 | # if i > 0 and args[0] != no: # channel expansion factor 250 | # ex = 1 + gw # exponential (default 2.0) 251 | # ch1 = 32 # ch[1] 252 | # e = math.log(c2 / ch1) / math.log(2) # level 1-n 253 | # c2 = int(ch1 * ex ** e) 254 | # if m != Focus: 255 | # c2 = make_divisible(c2, 8) if c2 != no else c2 256 | 257 | args = [c1, c2, *args[1:]] 258 | if m in [BottleneckCSP, C3, C3ResAtnMHSA]: 259 | args.insert(2, n) 260 | n = 1 261 | elif m in [nn.BatchNorm2d, SEAtn]: 262 | args = [ch[f]] 263 | c2=ch[f if f < 0 else f + 1] 264 | elif m in [Concat,ConcatFusionFactor]: 265 | c2 = sum([ch[x if x < 0 else x + 1] for x in f]) 266 | elif m is Detect: 267 | args.append([ch[x + 1] for x in f]) 268 | if isinstance(args[1], int): # number of anchors 269 | args[1] = [list(range(args[1] * 2))] * len(f) 270 | elif m is Contract: 271 | c2 = ch[f if f < 0 else f + 1] * args[0] ** 2 272 | elif m is Expand: 273 | c2 = ch[f if f < 0 else f + 1] // args[0] ** 2 274 | elif m is MHSA: 275 | args=[ch[f],*args[:]] 276 | c2=ch[f if f < 0 else f + 1] 277 | elif m is RCAN: 278 | args = [ch[f if f < 0 else f + 1]] 279 | c2 = ch[f if f < 0 else f + 1] 280 | elif m is AtnMut: 281 | args[0]=args[0]//2 282 | args[1] = args[1] // 2 283 | c2=ch[f[1] if f[1] < 0 else f[1] + 1] 284 | elif m is LocalReconstruct: 285 | args[0] = args[0] // 2 286 | args[1] = args[1] // 2 287 | c2 = ch[f[1] if f[1] < 0 else f[1] + 1] 288 | else: 289 | c2 = ch[f if f < 0 else f + 1] 290 | 291 | m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module 292 | t = str(m)[8:-2].replace('__main__.', '') # module type 293 | np = sum([x.numel() for x in m_.parameters()]) # number params 294 | m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params 295 | logger.info('%3s%18s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print 296 | save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist 297 | layers.append(m_) 298 | ch.append(c2) 299 | return nn.Sequential(*layers), sorted(save) 300 | 301 | 302 | if __name__ == '__main__': 303 | parser = argparse.ArgumentParser() 304 | parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml') 305 | parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 306 | opt = parser.parse_args() 307 | opt.cfg = check_file(opt.cfg) # check file 308 | set_logging() 309 | device = select_device(opt.device) 310 | 311 | # Create model 312 | model = Model(opt.cfg).to(device) 313 | model.train() 314 | 315 | # Profile 316 | # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device) 317 | # y = model(img, profile=True) 318 | 319 | # Tensorboard 320 | # from torch.utils.tensorboard import SummaryWriter 321 | # tb_writer = SummaryWriter() 322 | # print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/") 323 | # tb_writer.add_graph(model.model, img) # add model to tensorboard 324 | # tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard 325 | -------------------------------------------------------------------------------- /models/yolov5s.yaml: -------------------------------------------------------------------------------- 1 | # parameters 2 | nc: 80 # number of classes 3 | depth_multiple: 0.33 # model depth multiple 4 | width_multiple: 0.50 # layer channel multiple 5 | 6 | # anchors 7 | anchors: 8 | - [10,13, 16,30, 33,23] # P3/8 9 | - [30,61, 62,45, 59,119] # P4/16 10 | - [116,90, 156,198, 373,326] # P5/32 11 | 12 | # YOLOv5 backbone 13 | backbone: 14 | # [from, number, module, args] 15 | [[-1, 1, Focus, [64, 3]], # 0-P1/2 16 | [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 17 | [-1, 3, C3, [128]], 18 | [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 19 | [-1, 9, C3, [256]], 20 | [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 21 | [-1, 9, C3, [512]], 22 | [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 23 | [-1, 1, SPP, [1024, [5, 9, 13]]], 24 | [-1, 3, C3, [1024, False]], # 9 25 | ] 26 | 27 | # YOLOv5 head 28 | head: 29 | [[-1, 1, Conv, [512, 1, 1]], 30 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 31 | [[-1, 6], 1, Concat, [1]], # cat backbone P4 32 | [-1, 3, C3, [512, False]], # 13 33 | 34 | [-1, 1, Conv, [256, 1, 1]], 35 | [-1, 1, nn.Upsample, [None, 2, 'nearest']], 36 | [[-1, 4], 1, Concat, [1]], # cat backbone P3 37 | [-1, 3, C3, [256, False]], # 17 (P3/8-small) 38 | 39 | [-1, 1, Conv, [256, 3, 2]], 40 | [[-1, 14], 1, Concat, [1]], # cat head P4 41 | [-1, 3, C3, [512, False]], # 20 (P4/16-medium) 42 | 43 | [-1, 1, Conv, [512, 3, 2]], 44 | [[-1, 10], 1, Concat, [1]], # cat head P5 45 | [-1, 3, C3, [1024, False]], # 23 (P5/32-large) 46 | 47 | [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) 48 | ] 49 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | # pip install -r requirements.txt 2 | 3 | # base ---------------------------------------- 4 | Cython 5 | matplotlib>=3.2.2 6 | numpy>=1.18.5 7 | opencv-python>=4.1.2 8 | Pillow 9 | PyYAML>=5.3 10 | scipy>=1.4.1 11 | tensorboard>=2.2 12 | torch>=1.7.0 13 | torchvision>=0.8.1 14 | tqdm>=4.41.0 15 | 16 | # logging ------------------------------------- 17 | wandb 18 | 19 | # plotting ------------------------------------ 20 | seaborn>=0.11.0 21 | pandas 22 | 23 | # export -------------------------------------- 24 | # coremltools==4.0 25 | # onnx>=1.8.0 26 | # scikit-learn==0.19.2 # for coreml quantization 27 | 28 | # extras -------------------------------------- 29 | thop # FLOPS computation 30 | pycocotools>=2.0 # COCO mAP 31 | -------------------------------------------------------------------------------- /test.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import json 3 | import os 4 | from pathlib import Path 5 | from threading import Thread 6 | 7 | import numpy as np 8 | import torch 9 | import yaml 10 | from tqdm import tqdm 11 | 12 | from models.experimental import attempt_load 13 | from utils.datasets import create_dataloader 14 | from utils.general import coco80_to_coco91_class, check_dataset, check_file, check_img_size, check_requirements, \ 15 | box_iou, non_max_suppression, scale_coords, xyxy2xywh, xywh2xyxy, set_logging, increment_path, colorstr 16 | from utils.metrics import ap_per_class, ConfusionMatrix 17 | from utils.plots import plot_images, output_to_target, plot_study_txt, plot_images_each, plot_images_together 18 | from utils.torch_utils import select_device, time_synchronized 19 | import time 20 | 21 | def test(data, 22 | weights=None, 23 | batch_size=32, 24 | imgsz=640, 25 | conf_thres=0.001, 26 | iou_thres=0.6, # for NMS 27 | save_json=False, 28 | single_cls=False, 29 | augment=False, 30 | verbose=False, 31 | model=None, 32 | dataloader=None, 33 | save_dir=Path(''), # for saving images 34 | save_txt=False, # for auto-labelling 35 | save_hybrid=False, # for hybrid auto-labelling 36 | save_conf=False, # save auto-label confidences 37 | plots=True, 38 | log_imgs=0, # number of logged images 39 | compute_loss=None): 40 | 41 | plot_batch_num = 3 # How many batches you want to plot in test phase. 42 | 43 | # Initialize/load model and set device 44 | training = model is not None 45 | if training: # called by train.py 46 | device = next(model.parameters()).device # get model device 47 | 48 | else: # called directly 49 | set_logging() 50 | device = select_device(opt.device, batch_size=batch_size) 51 | 52 | # Directories 53 | save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) # increment run 54 | (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir 55 | 56 | # Load model 57 | model = attempt_load(weights, map_location=device) # load FP32 model 58 | imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size 59 | 60 | # Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99 61 | # if device.type != 'cpu' and torch.cuda.device_count() > 1: 62 | # model = nn.DataParallel(model) 63 | 64 | # Half 65 | half = device.type != 'cpu' # half precision only supported on CUDA 66 | if half: 67 | model.half() 68 | 69 | # Configure 70 | model.eval() 71 | is_coco = data.endswith('coco.yaml') # is COCO dataset 72 | with open(data) as f: 73 | data = yaml.load(f, Loader=yaml.SafeLoader) # model dict 74 | check_dataset(data) # check 75 | nc = 1 if single_cls else int(data['nc']) # number of classes 76 | iouv = torch.linspace(0.5, 0.95, 10).to(device) # iou vector for mAP@0.5:0.95 77 | niou = iouv.numel() 78 | 79 | # Logging 80 | log_imgs, wandb = min(log_imgs, 100), None # ceil 81 | try: 82 | import wandb # Weights & Biases 83 | except ImportError: 84 | log_imgs = 0 85 | 86 | # Dataloader 87 | if not training: 88 | img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img 89 | _ = model(img.half() if half else img) if device.type != 'cpu' else None # run once 90 | path = data['test'] if opt.task == 'test' else data['val'] # path to val/test images 91 | dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, pad=0.0, rect=True, 92 | prefix=colorstr('test: ' if opt.task == 'test' else 'val: '))[0] 93 | 94 | seen = 0 95 | confusion_matrix = ConfusionMatrix(nc=nc) 96 | names = {k: v for k, v in enumerate(model.names if hasattr(model, 'names') else model.module.names)} 97 | coco91class = coco80_to_coco91_class() 98 | s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R', 'mAP@.5', 'mAP@.5:.95') 99 | p, r, f1, mp, mr, map50, map, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0. 100 | loss = torch.zeros(4, device=device) 101 | jdict, stats, ap, ap_class, wandb_images = [], [], [], [], [] 102 | alltime=0 103 | for batch_i, (img, dgimgs, targets, paths, shapes) in enumerate(tqdm(dataloader, desc=s)): 104 | img = img.to(device, non_blocking=True) 105 | img = img.half() if half else img.float() # uint8 to fp16/32 106 | img /= 255.0 # 0 - 255 to 0.0 - 1.0 107 | targets = targets.to(device) 108 | dgimgs = dgimgs.to(device) 109 | nb, _, height, width = img.shape # batch size, channels, height, width 110 | 111 | with torch.no_grad(): 112 | # Run model 113 | t = time_synchronized() 114 | start = time.time() 115 | (out, train_out), pdg = model(img, augment=augment) # inference and training outputs 116 | alltime+=time.time()-start 117 | t0 += time_synchronized() - t 118 | 119 | # Compute loss 120 | if compute_loss: 121 | loss += compute_loss(([x.float() for x in train_out], pdg), dgimgs, targets)[1][:4] # box, obj, cls 122 | 123 | # Run NMS 124 | targets[:, 2:] *= torch.Tensor([width, height, width, height]).to(device) # to pixels 125 | lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling 126 | t = time_synchronized() 127 | out = non_max_suppression(out, conf_thres=conf_thres, iou_thres=iou_thres, labels=lb, multi_label=True) 128 | t1 += time_synchronized() - t 129 | 130 | # Statistics per image 131 | for si, pred in enumerate(out): 132 | labels = targets[targets[:, 0] == si, 1:] 133 | nl = len(labels) 134 | tcls = labels[:, 0].tolist() if nl else [] # target class 135 | path = Path(paths[si]) 136 | seen += 1 137 | 138 | if len(pred) == 0: 139 | if nl: 140 | stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls)) 141 | continue 142 | 143 | # Predictions 144 | predn = pred.clone() 145 | scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred 146 | 147 | # Append to text file 148 | if save_txt: 149 | gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0]] # normalization gain whwh 150 | for *xyxy, conf, cls in predn.tolist(): 151 | xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh 152 | line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format 153 | with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f: 154 | f.write(('%g ' * len(line)).rstrip() % line + '\n') 155 | 156 | # W&B logging 157 | if plots and len(wandb_images) < log_imgs: 158 | box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]}, 159 | "class_id": int(cls), 160 | "box_caption": "%s %.3f" % (names[cls], conf), 161 | "scores": {"class_score": conf}, 162 | "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] 163 | boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space 164 | if img.shape[1]!=3: 165 | wandb_images.append(wandb.Image(img[si][:3,:,:], boxes=boxes, caption=path.name)) 166 | else: 167 | wandb_images.append(wandb.Image(img[si], boxes=boxes, caption=path.name)) 168 | 169 | # Append to pycocotools JSON dictionary 170 | if save_json: 171 | # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ... 172 | image_id = int(path.stem) if path.stem.isnumeric() else path.stem 173 | box = xyxy2xywh(predn[:, :4]) # xywh 174 | box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner 175 | for p, b in zip(pred.tolist(), box.tolist()): 176 | jdict.append({'image_id': image_id, 177 | 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]), 178 | 'bbox': [round(x, 3) for x in b], 179 | 'score': round(p[4], 5)}) 180 | 181 | # Assign all predictions as incorrect 182 | correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device) 183 | if nl: 184 | detected = [] # target indices 185 | tcls_tensor = labels[:, 0] 186 | 187 | # target boxes 188 | tbox = xywh2xyxy(labels[:, 1:5]) 189 | scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels 190 | if plots: 191 | confusion_matrix.process_batch(predn, torch.cat((labels[:, 0:1], tbox), 1)) 192 | 193 | # Per target class 194 | for cls in torch.unique(tcls_tensor): 195 | ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # prediction indices 196 | pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # target indices 197 | 198 | # Search for detections 199 | if pi.shape[0]: 200 | # Prediction to target ious 201 | ious, i = box_iou(predn[pi, :4], tbox[ti]).max(1) # best ious, indices 202 | 203 | # Append detections 204 | detected_set = set() 205 | for j in (ious > iouv[0]).nonzero(as_tuple=False): 206 | d = ti[i[j]] # detected target 207 | if d.item() not in detected_set: 208 | detected_set.add(d.item()) 209 | detected.append(d) 210 | correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn 211 | if len(detected) == nl: # all targets already located in image 212 | break 213 | 214 | # Append statistics (correct, conf, pcls, tcls) 215 | stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls)) 216 | # Plot images 217 | if plots and batch_i < plot_batch_num: 218 | f = save_dir / f'test_batch{batch_i}_labels_pred.jpg' #altogether 219 | Thread(target=plot_images_together, args=(img, targets, output_to_target(out), paths, f, names), daemon=True).start() 220 | # f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels 221 | # Thread(target=plot_images_each, args=(img, targets, paths, f, names), daemon=True).start() 222 | # f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions 223 | # if pdg!=None: 224 | # Thread(target=plot_images_each, args=(img, output_to_target(out), paths, f, names), daemon=True).start() 225 | # else: 226 | # Thread(target=plot_images_each, args=(img, output_to_target(out), paths, f, names), daemon=True).start() 227 | print(alltime) 228 | 229 | # Compute statistics 230 | stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy 231 | if len(stats) and stats[0].any(): 232 | p, r, ap, f1, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) 233 | p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1) # [P, R, AP@0.5, AP@0.5:0.95] 234 | mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() 235 | nt = np.bincount(stats[3].astype(np.int64), minlength=nc) # number of targets per class 236 | else: 237 | nt = torch.zeros(1) 238 | 239 | # Print results 240 | pf = '%20s' + '%12.3g' * 6 # print format 241 | print(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) 242 | 243 | # Print results per class 244 | if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): 245 | for i, c in enumerate(ap_class): 246 | print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) 247 | 248 | # Print speeds 249 | t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple 250 | if not training: 251 | print('Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g' % t) 252 | 253 | # Plots 254 | if plots: 255 | confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) 256 | if wandb and wandb.run: 257 | wandb.log({"Images": wandb_images}) 258 | wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in sorted(save_dir.glob('test*.jpg'))]}) 259 | 260 | # Save JSON 261 | if save_json and len(jdict): 262 | w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights 263 | anno_json = '../coco/annotations/instances_val2017.json' # annotations json 264 | pred_json = str(save_dir / f"{w}_predictions.json") # predictions json 265 | print('\nEvaluating pycocotools mAP... saving %s...' % pred_json) 266 | with open(pred_json, 'w') as f: 267 | json.dump(jdict, f) 268 | 269 | try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb 270 | from pycocotools.coco import COCO 271 | from pycocotools.cocoeval import COCOeval 272 | 273 | anno = COCO(anno_json) # init annotations api 274 | pred = anno.loadRes(pred_json) # init predictions api 275 | eval = COCOeval(anno, pred, 'bbox') 276 | if is_coco: 277 | eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.img_files] # image IDs to evaluate 278 | eval.evaluate() 279 | eval.accumulate() 280 | eval.summarize() 281 | map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) 282 | except Exception as e: 283 | print(f'pycocotools unable to run: {e}') 284 | 285 | # Return results 286 | if not training: 287 | s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' 288 | print(f"Results saved to {save_dir}{s}") 289 | model.float() # for training 290 | maps = np.zeros(nc) + map 291 | for i, c in enumerate(ap_class): 292 | maps[c] = ap[i] 293 | return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t 294 | 295 | 296 | if __name__ == '__main__': 297 | parser = argparse.ArgumentParser(prog='test.py') 298 | parser.add_argument('--weights', nargs='+', type=str, default=r'.\DRENet.pt', help='model.pt path(s)') 299 | parser.add_argument('--data', type=str, default='data/ship.yaml', help='*.data path') 300 | parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch') 301 | parser.add_argument('--img-size', type=int, default=512, help='inference size (pixels)') 302 | parser.add_argument('--conf-thres', type=float, default=0.001, help='object confidence threshold') 303 | parser.add_argument('--iou-thres', type=float, default=0.6, help='IOU threshold for NMS') 304 | parser.add_argument('--task', default='val', help="'val', 'test', 'study'") 305 | parser.add_argument('--device', default='0', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') 306 | parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset') 307 | parser.add_argument('--augment', action='store_true', help='augmented inference') 308 | parser.add_argument('--verbose', action='store_true', help='report mAP by class') 309 | parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') 310 | parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt') 311 | parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') 312 | parser.add_argument('--save-json', action='store_true', help='save a cocoapi-compatible JSON results file') 313 | parser.add_argument('--project', default='runs/test', help='save to project/name') 314 | parser.add_argument('--name', default='test', help='save to project/name') 315 | parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') 316 | opt = parser.parse_args() 317 | opt.save_json |= opt.data.endswith('coco.yaml') 318 | opt.data = check_file(opt.data) # check file 319 | print(opt) 320 | check_requirements() 321 | 322 | if opt.task in ['val', 'test']: # run normally 323 | test(opt.data, 324 | opt.weights, 325 | opt.batch_size, 326 | opt.img_size, 327 | opt.conf_thres, 328 | opt.iou_thres, 329 | opt.save_json, 330 | opt.single_cls, 331 | opt.augment, 332 | opt.verbose, 333 | save_txt=opt.save_txt | opt.save_hybrid, 334 | save_hybrid=opt.save_hybrid, 335 | save_conf=opt.save_conf, 336 | ) 337 | 338 | elif opt.task == 'study': # run over a range of settings and save/plot 339 | for weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']: 340 | f = 'study_%s_%s.txt' % (Path(opt.data).stem, Path(weights).stem) # filename to save to 341 | x = list(range(320, 800, 64)) # x axis 342 | y = [] # y axis 343 | for i in x: # img-size 344 | print('\nRunning %s point %s...' % (f, i)) 345 | r, _, t = test(opt.data, weights, opt.batch_size, i, opt.conf_thres, opt.iou_thres, opt.save_json, 346 | plots=False) 347 | y.append(r + t) # results and times 348 | np.savetxt(f, y, fmt='%10.4g') # save 349 | os.system('zip -r study.zip study_*.txt') 350 | plot_study_txt(f, x) # plot 351 | -------------------------------------------------------------------------------- /utils/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/WindVChen/DRENet/a187dbe0f623b521a62c6176c7cafaa7322f5f66/utils/__init__.py -------------------------------------------------------------------------------- /utils/activations.py: -------------------------------------------------------------------------------- 1 | # Activation functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | import torch.nn.functional as F 6 | 7 | 8 | # SiLU https://arxiv.org/pdf/1606.08415.pdf ---------------------------------------------------------------------------- 9 | class SiLU(nn.Module): # export-friendly version of nn.SiLU() 10 | @staticmethod 11 | def forward(x): 12 | return x * torch.sigmoid(x) 13 | 14 | 15 | class Hardswish(nn.Module): # export-friendly version of nn.Hardswish() 16 | @staticmethod 17 | def forward(x): 18 | # return x * F.hardsigmoid(x) # for torchscript and CoreML 19 | return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX 20 | 21 | 22 | class MemoryEfficientSwish(nn.Module): 23 | class F(torch.autograd.Function): 24 | @staticmethod 25 | def forward(ctx, x): 26 | ctx.save_for_backward(x) 27 | return x * torch.sigmoid(x) 28 | 29 | @staticmethod 30 | def backward(ctx, grad_output): 31 | x = ctx.saved_tensors[0] 32 | sx = torch.sigmoid(x) 33 | return grad_output * (sx * (1 + x * (1 - sx))) 34 | 35 | def forward(self, x): 36 | return self.F.apply(x) 37 | 38 | 39 | # Mish https://github.com/digantamisra98/Mish -------------------------------------------------------------------------- 40 | class Mish(nn.Module): 41 | @staticmethod 42 | def forward(x): 43 | return x * F.softplus(x).tanh() 44 | 45 | 46 | class MemoryEfficientMish(nn.Module): 47 | class F(torch.autograd.Function): 48 | @staticmethod 49 | def forward(ctx, x): 50 | ctx.save_for_backward(x) 51 | return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) 52 | 53 | @staticmethod 54 | def backward(ctx, grad_output): 55 | x = ctx.saved_tensors[0] 56 | sx = torch.sigmoid(x) 57 | fx = F.softplus(x).tanh() 58 | return grad_output * (fx + x * sx * (1 - fx * fx)) 59 | 60 | def forward(self, x): 61 | return self.F.apply(x) 62 | 63 | 64 | # FReLU https://arxiv.org/abs/2007.11824 ------------------------------------------------------------------------------- 65 | class FReLU(nn.Module): 66 | def __init__(self, c1, k=3): # ch_in, kernel 67 | super().__init__() 68 | self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) 69 | self.bn = nn.BatchNorm2d(c1) 70 | 71 | def forward(self, x): 72 | return torch.max(x, self.bn(self.conv(x))) 73 | -------------------------------------------------------------------------------- /utils/autoanchor.py: -------------------------------------------------------------------------------- 1 | # Auto-anchor utils 2 | 3 | import numpy as np 4 | import torch 5 | import yaml 6 | from scipy.cluster.vq import kmeans 7 | from tqdm import tqdm 8 | 9 | from utils.general import colorstr 10 | 11 | 12 | def check_anchor_order(m): 13 | # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary 14 | a = m.anchor_grid.prod(-1).view(-1) # anchor area 15 | da = a[-1] - a[0] # delta a 16 | ds = m.stride[-1] - m.stride[0] # delta s 17 | if da.sign() != ds.sign(): # same order 18 | print('Reversing anchor order') 19 | m.anchors[:] = m.anchors.flip(0) 20 | m.anchor_grid[:] = m.anchor_grid.flip(0) 21 | 22 | 23 | def check_anchors(dataset, model, thr=4.0, imgsz=640): 24 | # Check anchor fit to data, recompute if necessary 25 | prefix = colorstr('autoanchor: ') 26 | print(f'\n{prefix}Analyzing anchors... ', end='') 27 | m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() 28 | shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) 29 | scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale 30 | if shapes.shape[1]!=2: 31 | wh = torch.tensor(np.concatenate([l[:, 3:5] * s[:-1] for s, l in zip(shapes * scale, dataset.labels)])).float() # wh 32 | else: 33 | wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() 34 | 35 | def metric(k): # compute metric 36 | r = wh[:, None] / k[None] 37 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 38 | best = x.max(1)[0] # best_x 39 | aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold 40 | bpr = (best > 1. / thr).float().mean() # best possible recall 41 | return bpr, aat 42 | 43 | anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors 44 | bpr, aat = metric(anchors) 45 | print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='') 46 | bpr=0.97 47 | if bpr < 0.98: # threshold to recompute 48 | print('. Attempting to improve anchors, please wait...') 49 | na = m.anchor_grid.numel() // 2 # number of anchors 50 | try: 51 | anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) 52 | except Exception as e: 53 | print(f'{prefix}ERROR: {e}') 54 | new_bpr = metric(anchors)[0] 55 | if new_bpr > bpr: # replace anchors 56 | anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) 57 | m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference 58 | m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss 59 | check_anchor_order(m) 60 | print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.') 61 | else: 62 | print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.') 63 | print('') # newline 64 | 65 | 66 | def kmean_anchors(path='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): 67 | """ Creates kmeans-evolved anchors from training dataset 68 | 69 | Arguments: 70 | path: path to dataset *.yaml, or a loaded dataset 71 | n: number of anchors 72 | img_size: image size used for training 73 | thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 74 | gen: generations to evolve anchors using genetic algorithm 75 | verbose: print all results 76 | 77 | Return: 78 | k: kmeans evolved anchors 79 | 80 | Usage: 81 | from utils.autoanchor import *; _ = kmean_anchors() 82 | """ 83 | thr = 1. / thr 84 | prefix = colorstr('autoanchor: ') 85 | 86 | def metric(k, wh): # compute metrics 87 | r = wh[:, None] / k[None] 88 | x = torch.min(r, 1. / r).min(2)[0] # ratio metric 89 | # x = wh_iou(wh, torch.tensor(k)) # iou metric 90 | return x, x.max(1)[0] # x, best_x 91 | 92 | def anchor_fitness(k): # mutation fitness 93 | _, best = metric(torch.tensor(k, dtype=torch.float32), wh) 94 | return (best * (best > thr).float()).mean() # fitness 95 | 96 | def print_results(k): 97 | k = k[np.argsort(k.prod(1))] # sort small to large 98 | x, best = metric(k, wh0) 99 | bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr 100 | print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr') 101 | print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' 102 | f'past_thr={x[x > thr].mean():.3f}-mean: ', end='') 103 | for i, x in enumerate(k): 104 | print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg 105 | return k 106 | 107 | if isinstance(path, str): # *.yaml file 108 | with open(path) as f: 109 | data_dict = yaml.load(f, Loader=yaml.SafeLoader) # model dict 110 | from utils.datasets import LoadImagesAndLabels 111 | dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) 112 | else: 113 | dataset = path # dataset 114 | 115 | # Get label wh 116 | shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) 117 | wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh 118 | 119 | # Filter 120 | i = (wh0 < 3.0).any(1).sum() 121 | if i: 122 | print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.') 123 | wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels 124 | # wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 125 | 126 | # Kmeans calculation 127 | print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...') 128 | s = wh.std(0) # sigmas for whitening 129 | k, dist = kmeans(wh / s, n, iter=30) # points, mean distance 130 | assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}') 131 | k *= s 132 | wh = torch.tensor(wh, dtype=torch.float32) # filtered 133 | wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered 134 | k = print_results(k) 135 | 136 | # Plot 137 | # k, d = [None] * 20, [None] * 20 138 | # for i in tqdm(range(1, 21)): 139 | # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance 140 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) 141 | # ax = ax.ravel() 142 | # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') 143 | # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh 144 | # ax[0].hist(wh[wh[:, 0]<100, 0],400) 145 | # ax[1].hist(wh[wh[:, 1]<100, 1],400) 146 | # fig.savefig('wh.png', dpi=200) 147 | 148 | # Evolve 149 | npr = np.random 150 | f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma 151 | pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar 152 | for _ in pbar: 153 | v = np.ones(sh) 154 | while (v == 1).all(): # mutate until a change occurs (prevent duplicates) 155 | v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) 156 | kg = (k.copy() * v).clip(min=2.0) 157 | fg = anchor_fitness(kg) 158 | if fg > f: 159 | f, k = fg, kg.copy() 160 | pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' 161 | if verbose: 162 | print_results(k) 163 | 164 | return print_results(k) 165 | -------------------------------------------------------------------------------- /utils/general.py: -------------------------------------------------------------------------------- 1 | # General utils 2 | 3 | import glob 4 | import logging 5 | import math 6 | import os 7 | import platform 8 | import random 9 | import re 10 | import subprocess 11 | import time 12 | from pathlib import Path 13 | 14 | import cv2 15 | import numpy as np 16 | import torch 17 | import torchvision 18 | import yaml 19 | 20 | from utils.google_utils import gsutil_getsize 21 | from utils.metrics import fitness 22 | from utils.torch_utils import init_torch_seeds 23 | 24 | # Settings 25 | torch.set_printoptions(linewidth=320, precision=5, profile='long') 26 | np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 27 | cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) 28 | os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8)) # NumExpr max threads 29 | 30 | 31 | def set_logging(rank=-1): 32 | logging.basicConfig( 33 | format="%(message)s", 34 | level=logging.INFO if rank in [-1, 0] else logging.WARN) 35 | 36 | 37 | def init_seeds(seed=0): 38 | # Initialize random number generator (RNG) seeds 39 | random.seed(seed) 40 | np.random.seed(seed) 41 | init_torch_seeds(seed) 42 | 43 | 44 | def get_latest_run(search_dir='.'): 45 | # Return path to most recent 'last.pt' in /runs (i.e. to --resume from) 46 | last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True) 47 | return max(last_list, key=os.path.getctime) if last_list else '' 48 | 49 | 50 | def isdocker(): 51 | # Is environment a Docker container 52 | return Path('/workspace').exists() # or Path('/.dockerenv').exists() 53 | 54 | 55 | def check_online(): 56 | # Check internet connectivity 57 | import socket 58 | try: 59 | socket.create_connection(("1.1.1.1", 443), 5) # check host accesability 60 | return True 61 | except OSError: 62 | return False 63 | 64 | 65 | def check_git_status(): 66 | # Recommend 'git pull' if code is out of date 67 | print(colorstr('github: '), end='') 68 | try: 69 | assert Path('.git').exists(), 'skipping check (not a git repository)' 70 | assert not isdocker(), 'skipping check (Docker image)' 71 | assert check_online(), 'skipping check (offline)' 72 | 73 | cmd = 'git fetch && git config --get remote.origin.url' 74 | url = subprocess.check_output(cmd, shell=True).decode().strip().rstrip('.git') # github repo url 75 | branch = subprocess.check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip() # checked out 76 | n = int(subprocess.check_output(f'git rev-list {branch}..origin/master --count', shell=True)) # commits behind 77 | if n > 0: 78 | s = f"⚠️ WARNING: code is out of date by {n} commit{'s' * (n > 1)}. " \ 79 | f"Use 'git pull' to update or 'git clone {url}' to download latest." 80 | else: 81 | s = f'up to date with {url} ✅' 82 | print(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) 83 | except Exception as e: 84 | print(e) 85 | 86 | 87 | def check_requirements(file='requirements.txt', exclude=()): 88 | # Check installed dependencies meet requirements 89 | import pkg_resources 90 | requirements = [f'{x.name}{x.specifier}' for x in pkg_resources.parse_requirements(Path(file).open()) 91 | if x.name not in exclude] 92 | pkg_resources.require(requirements) # DistributionNotFound or VersionConflict exception if requirements not met 93 | 94 | 95 | def check_img_size(img_size, s=32): 96 | # Verify img_size is a multiple of stride s 97 | new_size = make_divisible(img_size, int(s)) # ceil gs-multiple 98 | if new_size != img_size: 99 | print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size)) 100 | return new_size 101 | 102 | 103 | def check_imshow(): 104 | # Check if environment supports image displays 105 | try: 106 | assert not isdocker(), 'cv2.imshow() is disabled in Docker environments' 107 | cv2.imshow('test', np.zeros((1, 1, 3))) 108 | cv2.waitKey(1) 109 | cv2.destroyAllWindows() 110 | cv2.waitKey(1) 111 | return True 112 | except Exception as e: 113 | print(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}') 114 | return False 115 | 116 | 117 | def check_file(file): 118 | # Search for file if not found 119 | if os.path.isfile(file) or file == '': 120 | return file 121 | else: 122 | files = glob.glob('./**/' + file, recursive=True) # find file 123 | assert len(files), 'File Not Found: %s' % file # assert file was found 124 | assert len(files) == 1, "Multiple files match '%s', specify exact path: %s" % (file, files) # assert unique 125 | return files[0] # return file 126 | 127 | 128 | def check_dataset(dict): 129 | # Download dataset if not found locally 130 | val, s = dict.get('val'), dict.get('download') 131 | if val and len(val): 132 | val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path 133 | if not all(x.exists() for x in val): 134 | print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()]) 135 | if s and len(s): # download script 136 | print('Downloading %s ...' % s) 137 | if s.startswith('http') and s.endswith('.zip'): # URL 138 | f = Path(s).name # filename 139 | torch.hub.download_url_to_file(s, f) 140 | r = os.system('unzip -q %s -d ../ && rm %s' % (f, f)) # unzip 141 | else: # bash script 142 | r = os.system(s) 143 | print('Dataset autodownload %s\n' % ('success' if r == 0 else 'failure')) # analyze return value 144 | else: 145 | raise Exception('Dataset not found.') 146 | 147 | 148 | def make_divisible(x, divisor): 149 | # Returns x evenly divisible by divisor 150 | return math.ceil(x / divisor) * divisor 151 | 152 | 153 | def clean_str(s): 154 | # Cleans a string by replacing special characters with underscore _ 155 | return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s) 156 | 157 | 158 | def one_cycle(y1=0.0, y2=1.0, steps=100): 159 | # lambda function for sinusoidal ramp from y1 to y2 160 | return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1 161 | 162 | 163 | def colorstr(*input): 164 | # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world') 165 | *args, string = input if len(input) > 1 else ('blue', 'bold', input[0]) # color arguments, string 166 | colors = {'black': '\033[30m', # basic colors 167 | 'red': '\033[31m', 168 | 'green': '\033[32m', 169 | 'yellow': '\033[33m', 170 | 'blue': '\033[34m', 171 | 'magenta': '\033[35m', 172 | 'cyan': '\033[36m', 173 | 'white': '\033[37m', 174 | 'bright_black': '\033[90m', # bright colors 175 | 'bright_red': '\033[91m', 176 | 'bright_green': '\033[92m', 177 | 'bright_yellow': '\033[93m', 178 | 'bright_blue': '\033[94m', 179 | 'bright_magenta': '\033[95m', 180 | 'bright_cyan': '\033[96m', 181 | 'bright_white': '\033[97m', 182 | 'end': '\033[0m', # misc 183 | 'bold': '\033[1m', 184 | 'underline': '\033[4m'} 185 | return ''.join(colors[x] for x in args) + f'{string}' + colors['end'] 186 | 187 | 188 | def labels_to_class_weights(labels, nc=80): 189 | # Get class weights (inverse frequency) from training labels 190 | if labels[0] is None: # no labels loaded 191 | return torch.Tensor() 192 | 193 | labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO 194 | classes = labels[:, 0].astype(np.int) # labels = [class xywh] 195 | weights = np.bincount(classes, minlength=nc) # occurrences per class 196 | 197 | # Prepend gridpoint count (for uCE training) 198 | # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image 199 | # weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start 200 | 201 | weights[weights == 0] = 1 # replace empty bins with 1 202 | weights = 1 / weights # number of targets per class 203 | weights /= weights.sum() # normalize 204 | return torch.from_numpy(weights) 205 | 206 | 207 | def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)): 208 | # Produces image weights based on class_weights and image contents 209 | class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels]) 210 | image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1) 211 | # index = random.choices(range(n), weights=image_weights, k=1) # weight image sample 212 | return image_weights 213 | 214 | 215 | def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) 216 | # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ 217 | # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n') 218 | # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n') 219 | # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco 220 | # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet 221 | x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 222 | 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 223 | 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] 224 | return x 225 | 226 | 227 | def xyxy2xywh(x): 228 | # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right 229 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 230 | y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center 231 | y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center 232 | y[:, 2] = x[:, 2] - x[:, 0] # width 233 | y[:, 3] = x[:, 3] - x[:, 1] # height 234 | return y 235 | 236 | 237 | def xywh2xyxy(x): 238 | # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right 239 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 240 | y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x 241 | y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y 242 | y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x 243 | y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y 244 | return y 245 | 246 | 247 | def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0): 248 | # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right 249 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 250 | y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw # top left x 251 | y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh # top left y 252 | y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw # bottom right x 253 | y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh # bottom right y 254 | return y 255 | 256 | 257 | def xyn2xy(x, w=640, h=640, padw=0, padh=0): 258 | # Convert normalized segments into pixel segments, shape (n,2) 259 | y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) 260 | y[:, 0] = w * x[:, 0] + padw # top left x 261 | y[:, 1] = h * x[:, 1] + padh # top left y 262 | return y 263 | 264 | 265 | def segment2box(segment, width=640, height=640): 266 | # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy) 267 | x, y = segment.T # segment xy 268 | inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height) 269 | x, y, = x[inside], y[inside] 270 | return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4)) # cls, xyxy 271 | 272 | 273 | def segments2boxes(segments): 274 | # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh) 275 | boxes = [] 276 | for s in segments: 277 | x, y = s.T # segment xy 278 | boxes.append([x.min(), y.min(), x.max(), y.max()]) # cls, xyxy 279 | return xyxy2xywh(np.array(boxes)) # cls, xywh 280 | 281 | 282 | def resample_segments(segments, n=1000): 283 | # Up-sample an (n,2) segment 284 | for i, s in enumerate(segments): 285 | x = np.linspace(0, len(s) - 1, n) 286 | xp = np.arange(len(s)) 287 | segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T # segment xy 288 | return segments 289 | 290 | 291 | def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None): 292 | # Rescale coords (xyxy) from img1_shape to img0_shape 293 | if ratio_pad is None: # calculate from img0_shape 294 | gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new 295 | pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding 296 | else: 297 | gain = ratio_pad[0][0] 298 | pad = ratio_pad[1] 299 | 300 | coords[:, [0, 2]] -= pad[0] # x padding 301 | coords[:, [1, 3]] -= pad[1] # y padding 302 | coords[:, :4] /= gain 303 | clip_coords(coords, img0_shape) 304 | return coords 305 | 306 | 307 | def clip_coords(boxes, img_shape): 308 | # Clip bounding xyxy bounding boxes to image shape (height, width) 309 | boxes[:, 0].clamp_(0, img_shape[1]) # x1 310 | boxes[:, 1].clamp_(0, img_shape[0]) # y1 311 | boxes[:, 2].clamp_(0, img_shape[1]) # x2 312 | boxes[:, 3].clamp_(0, img_shape[0]) # y2 313 | 314 | 315 | def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-9): 316 | # Returns the IoU of box1 to box2. box1 is 4, box2 is nx4 317 | box2 = box2.T 318 | 319 | # Get the coordinates of bounding boxes 320 | if x1y1x2y2: # x1, y1, x2, y2 = box1 321 | b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3] 322 | b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3] 323 | else: # transform from xywh to xyxy 324 | b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2 325 | b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2 326 | b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2 327 | b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2 328 | 329 | # Intersection area 330 | inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ 331 | (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) 332 | 333 | # Union Area 334 | w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps 335 | w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps 336 | union = w1 * h1 + w2 * h2 - inter + eps 337 | 338 | iou = inter / union 339 | if GIoU or DIoU or CIoU: 340 | cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width 341 | ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height 342 | if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1 343 | c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared 344 | rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + 345 | (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared 346 | if DIoU: 347 | return iou - rho2 / c2 # DIoU 348 | elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47 349 | v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2) 350 | with torch.no_grad(): 351 | alpha = v / ((1 + eps) - iou + v) 352 | return iou - (rho2 / c2 + v * alpha) # CIoU 353 | else: # GIoU https://arxiv.org/pdf/1902.09630.pdf 354 | c_area = cw * ch + eps # convex area 355 | return iou - (c_area - union) / c_area # GIoU 356 | else: 357 | return iou # IoU 358 | 359 | 360 | def box_iou(box1, box2): 361 | # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py 362 | """ 363 | Return intersection-over-union (Jaccard index) of boxes. 364 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 365 | Arguments: 366 | box1 (Tensor[N, 4]) 367 | box2 (Tensor[M, 4]) 368 | Returns: 369 | iou (Tensor[N, M]): the NxM matrix containing the pairwise 370 | IoU values for every element in boxes1 and boxes2 371 | """ 372 | 373 | def box_area(box): 374 | # box = 4xn 375 | return (box[2] - box[0]) * (box[3] - box[1]) 376 | 377 | area1 = box_area(box1.T) 378 | area2 = box_area(box2.T) 379 | 380 | # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) 381 | inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2) 382 | return inter / (area1[:, None] + area2 - inter) # iou = inter / (area1 + area2 - inter) 383 | 384 | 385 | def wh_iou(wh1, wh2): 386 | # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2 387 | wh1 = wh1[:, None] # [N,1,2] 388 | wh2 = wh2[None] # [1,M,2] 389 | inter = torch.min(wh1, wh2).prod(2) # [N,M] 390 | return inter / (wh1.prod(2) + wh2.prod(2) - inter) # iou = inter / (area1 + area2 - inter) 391 | 392 | 393 | def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False, 394 | labels=()): 395 | """Runs Non-Maximum Suppression (NMS) on inference results 396 | 397 | Returns: 398 | list of detections, on (n,6) tensor per image [xyxy, conf, cls] 399 | """ 400 | 401 | nc = prediction.shape[2] - 5 # number of classes 402 | xc = prediction[..., 4] > conf_thres # candidates 403 | 404 | # Settings 405 | min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height 406 | max_det = 300 # maximum number of detections per image 407 | max_nms = 30000 # maximum number of boxes into torchvision.ops.nms() 408 | time_limit = 10.0 # seconds to quit after 409 | redundant = True # require redundant detections 410 | multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img) 411 | merge = False # use merge-NMS 412 | 413 | t = time.time() 414 | output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0] 415 | for xi, x in enumerate(prediction): # image index, image inference 416 | # Apply constraints 417 | # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height 418 | x = x[xc[xi]] # confidence 419 | 420 | # Cat apriori labels if autolabelling 421 | if labels and len(labels[xi]): 422 | l = labels[xi] 423 | v = torch.zeros((len(l), nc + 5), device=x.device) 424 | v[:, :4] = l[:, 1:5] # box 425 | v[:, 4] = 1.0 # conf 426 | v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls 427 | x = torch.cat((x, v), 0) 428 | 429 | # If none remain process next image 430 | if not x.shape[0]: 431 | continue 432 | 433 | # Compute conf 434 | x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf 435 | 436 | # Box (center x, center y, width, height) to (x1, y1, x2, y2) 437 | box = xywh2xyxy(x[:, :4]) 438 | 439 | # Detections matrix nx6 (xyxy, conf, cls) 440 | if multi_label: 441 | i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T 442 | x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) 443 | else: # best class only 444 | conf, j = x[:, 5:].max(1, keepdim=True) 445 | x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] 446 | 447 | # Filter by class 448 | if classes is not None: 449 | x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] 450 | 451 | # Apply finite constraint 452 | # if not torch.isfinite(x).all(): 453 | # x = x[torch.isfinite(x).all(1)] 454 | 455 | # Check shape 456 | n = x.shape[0] # number of boxes 457 | if not n: # no boxes 458 | continue 459 | elif n > max_nms: # excess boxes 460 | x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence 461 | 462 | # Batched NMS 463 | c = x[:, 5:6] * (0 if agnostic else max_wh) # classes 464 | boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores 465 | i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS 466 | if i.shape[0] > max_det: # limit detections 467 | i = i[:max_det] 468 | if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) 469 | # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) 470 | iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix 471 | weights = iou * scores[None] # box weights 472 | x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes 473 | if redundant: 474 | i = i[iou.sum(1) > 1] # require redundancy 475 | 476 | output[xi] = x[i] 477 | if (time.time() - t) > time_limit: 478 | print(f'WARNING: NMS time limit {time_limit}s exceeded') 479 | break # time limit exceeded 480 | 481 | return output 482 | 483 | 484 | def strip_optimizer(f='best.pt', s=''): # from utils.general import *; strip_optimizer() 485 | # Strip optimizer from 'f' to finalize training, optionally save as 's' 486 | x = torch.load(f, map_location=torch.device('cpu')) 487 | if x.get('ema'): 488 | x['model'] = x['ema'] # replace model with ema 489 | for k in 'optimizer', 'training_results', 'wandb_id', 'ema', 'updates': # keys 490 | x[k] = None 491 | x['epoch'] = -1 492 | x['model'].half() # to FP16 493 | for p in x['model'].parameters(): 494 | p.requires_grad = False 495 | torch.save(x, s or f) 496 | mb = os.path.getsize(s or f) / 1E6 # filesize 497 | print(f"Optimizer stripped from {f},{(' saved as %s,' % s) if s else ''} {mb:.1f}MB") 498 | 499 | 500 | def print_mutation(hyp, results, yaml_file='hyp_evolved.yaml', bucket=''): 501 | # Print mutation results to evolve.txt (for use with train.py --evolve) 502 | a = '%10s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys 503 | b = '%10.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values 504 | c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) 505 | print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c)) 506 | 507 | if bucket: 508 | url = 'gs://%s/evolve.txt' % bucket 509 | if gsutil_getsize(url) > (os.path.getsize('evolve.txt') if os.path.exists('evolve.txt') else 0): 510 | os.system('gsutil cp %s .' % url) # download evolve.txt if larger than local 511 | 512 | with open('evolve.txt', 'a') as f: # append result 513 | f.write(c + b + '\n') 514 | x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows 515 | x = x[np.argsort(-fitness(x))] # sort 516 | np.savetxt('evolve.txt', x, '%10.3g') # save sort by fitness 517 | 518 | # Save yaml 519 | for i, k in enumerate(hyp.keys()): 520 | hyp[k] = float(x[0, i + 7]) 521 | with open(yaml_file, 'w') as f: 522 | results = tuple(x[0, :7]) 523 | c = '%10.4g' * len(results) % results # results (P, R, mAP@0.5, mAP@0.5:0.95, val_losses x 3) 524 | f.write('# Hyperparameter Evolution Results\n# Generations: %g\n# Metrics: ' % len(x) + c + '\n\n') 525 | yaml.dump(hyp, f, sort_keys=False) 526 | 527 | if bucket: 528 | os.system('gsutil cp evolve.txt %s gs://%s' % (yaml_file, bucket)) # upload 529 | 530 | 531 | def apply_classifier(x, model, img, im0): 532 | # applies a second stage classifier to yolo outputs 533 | im0 = [im0] if isinstance(im0, np.ndarray) else im0 534 | for i, d in enumerate(x): # per image 535 | if d is not None and len(d): 536 | d = d.clone() 537 | 538 | # Reshape and pad cutouts 539 | b = xyxy2xywh(d[:, :4]) # boxes 540 | b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square 541 | b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad 542 | d[:, :4] = xywh2xyxy(b).long() 543 | 544 | # Rescale boxes from img_size to im0 size 545 | scale_coords(img.shape[2:], d[:, :4], im0[i].shape) 546 | 547 | # Classes 548 | pred_cls1 = d[:, 5].long() 549 | ims = [] 550 | for j, a in enumerate(d): # per item 551 | cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])] 552 | im = cv2.resize(cutout, (224, 224)) # BGR 553 | # cv2.imwrite('test%i.jpg' % j, cutout) 554 | 555 | im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 556 | im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32 557 | im /= 255.0 # 0 - 255 to 0.0 - 1.0 558 | ims.append(im) 559 | 560 | pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction 561 | x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections 562 | 563 | return x 564 | 565 | 566 | def increment_path(path, exist_ok=True, sep=''): 567 | # Increment path, i.e. runs/exp --> runs/exp{sep}0, runs/exp{sep}1 etc. 568 | path = Path(path) # os-agnostic 569 | if (path.exists() and exist_ok) or (not path.exists()): 570 | return str(path) 571 | else: 572 | dirs = glob.glob(f"{path}{sep}*") # similar paths 573 | matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs] 574 | i = [int(m.groups()[0]) for m in matches if m] # indices 575 | n = max(i) + 1 if i else 2 # increment number 576 | return f"{path}{sep}{n}" # update path 577 | -------------------------------------------------------------------------------- /utils/google_utils.py: -------------------------------------------------------------------------------- 1 | # Google utils: https://cloud.google.com/storage/docs/reference/libraries 2 | 3 | import os 4 | import platform 5 | import subprocess 6 | import time 7 | from pathlib import Path 8 | 9 | import requests 10 | import torch 11 | 12 | 13 | def gsutil_getsize(url=''): 14 | # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du 15 | s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') 16 | return eval(s.split(' ')[0]) if len(s) else 0 # bytes 17 | 18 | 19 | def attempt_download(file, repo='ultralytics/yolov5'): 20 | # Attempt file download if does not exist 21 | file = Path(str(file).strip().replace("'", '').lower()) 22 | if not file.exists(): 23 | try: 24 | proxy={"http":None,"https":None} 25 | #response = requests.get(r'D:\实验室\程序\YOLOV5\latest_access_token=86160778ebfa6f610de04e6f8609b4ffabae3d88',proxies=proxy).json() # github api 26 | response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest?access_token=86160778ebfa6f610de04e6f8609b4ffabae3d88',proxies=proxy).json() # github api 27 | assets = [x['name'] for x in response['assets']] # release assets, i.e. ['yolov5s.pt', 'yolov5m.pt', ...] 28 | tag = response['tag_name'] # i.e. 'v1.0' 29 | except: # fallback plan 30 | assets = ['yolov5.pt', 'yolov5.pt', 'yolov5l.pt', 'yolov5x.pt'] 31 | tag = subprocess.check_output('git tag', shell=True).decode('utf-8').split('\n')[-2] 32 | 33 | name = file.name 34 | if name in assets: 35 | msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/' 36 | redundant = False # second download option 37 | try: # GitHub 38 | url = f'https://github.com/{repo}/releases/download/{tag}/{name}' 39 | print(f'Downloading {url} to {file}...') 40 | torch.hub.download_url_to_file(url, file) 41 | assert file.exists() and file.stat().st_size > 1E6 # check 42 | except Exception as e: # GCP 43 | print(f'Download error: {e}') 44 | assert redundant, 'No secondary mirror' 45 | url = f'https://storage.googleapis.com/{repo}/ckpt/{name}' 46 | print(f'Downloading {url} to {file}...') 47 | os.system(f'curl -L {url} -o {file}') # torch.hub.download_url_to_file(url, weights) 48 | finally: 49 | if not file.exists() or file.stat().st_size < 1E6: # check 50 | file.unlink(missing_ok=True) # remove partial downloads 51 | print(f'ERROR: Download failure: {msg}') 52 | print('') 53 | return 54 | 55 | 56 | def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'): 57 | # Downloads a file from Google Drive. from yolov5.utils.google_utils import *; gdrive_download() 58 | t = time.time() 59 | file = Path(file) 60 | cookie = Path('cookie') # gdrive cookie 61 | print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') 62 | file.unlink(missing_ok=True) # remove existing file 63 | cookie.unlink(missing_ok=True) # remove existing cookie 64 | 65 | # Attempt file download 66 | out = "NUL" if platform.system() == "Windows" else "/dev/null" 67 | os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') 68 | if os.path.exists('cookie'): # large file 69 | s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' 70 | else: # small file 71 | s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' 72 | r = os.system(s) # execute, capture return 73 | cookie.unlink(missing_ok=True) # remove existing cookie 74 | 75 | # Error check 76 | if r != 0: 77 | file.unlink(missing_ok=True) # remove partial 78 | print('Download error ') # raise Exception('Download error') 79 | return r 80 | 81 | # Unzip if archive 82 | if file.suffix == '.zip': 83 | print('unzipping... ', end='') 84 | os.system(f'unzip -q {file}') # unzip 85 | file.unlink() # remove zip to free space 86 | 87 | print(f'Done ({time.time() - t:.1f}s)') 88 | return r 89 | 90 | 91 | def get_token(cookie="./cookie"): 92 | with open(cookie) as f: 93 | for line in f: 94 | if "download" in line: 95 | return line.split()[-1] 96 | return "" 97 | 98 | # def upload_blob(bucket_name, source_file_name, destination_blob_name): 99 | # # Uploads a file to a bucket 100 | # # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python 101 | # 102 | # storage_client = storage.Client() 103 | # bucket = storage_client.get_bucket(bucket_name) 104 | # blob = bucket.blob(destination_blob_name) 105 | # 106 | # blob.upload_from_filename(source_file_name) 107 | # 108 | # print('File {} uploaded to {}.'.format( 109 | # source_file_name, 110 | # destination_blob_name)) 111 | # 112 | # 113 | # def download_blob(bucket_name, source_blob_name, destination_file_name): 114 | # # Uploads a blob from a bucket 115 | # storage_client = storage.Client() 116 | # bucket = storage_client.get_bucket(bucket_name) 117 | # blob = bucket.blob(source_blob_name) 118 | # 119 | # blob.download_to_filename(destination_file_name) 120 | # 121 | # print('Blob {} downloaded to {}.'.format( 122 | # source_blob_name, 123 | # destination_file_name)) 124 | -------------------------------------------------------------------------------- /utils/loss.py: -------------------------------------------------------------------------------- 1 | # Loss functions 2 | 3 | import torch 4 | import torch.nn as nn 5 | 6 | from utils.general import bbox_iou 7 | from utils.torch_utils import is_parallel 8 | 9 | class AutomaticWeightedLoss(nn.Module): 10 | def __init__(self, num=2): 11 | super(AutomaticWeightedLoss, self).__init__() 12 | self.params = torch.nn.Parameter(torch.ones(num, requires_grad=True).cuda()) 13 | 14 | def forward(self, *x): 15 | loss_sum = 0 16 | for i, loss in enumerate(x): 17 | loss_sum += 0.5 / (self.params[i] ** 2) * loss + torch.log(1 + self.params[i] ** 2) 18 | return loss_sum 19 | 20 | def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 21 | # return positive, negative label smoothing BCE targets 22 | return 1.0 - 0.5 * eps, 0.5 * eps 23 | 24 | 25 | class BCEBlurWithLogitsLoss(nn.Module): 26 | # BCEwithLogitLoss() with reduced missing label effects. 27 | def __init__(self, alpha=0.05): 28 | super(BCEBlurWithLogitsLoss, self).__init__() 29 | self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() 30 | self.alpha = alpha 31 | 32 | def forward(self, pred, true): 33 | loss = self.loss_fcn(pred, true) 34 | pred = torch.sigmoid(pred) # prob from logits 35 | dx = pred - true # reduce only missing label effects 36 | # dx = (pred - true).abs() # reduce missing label and false label effects 37 | alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) 38 | loss *= alpha_factor 39 | return loss.mean() 40 | 41 | 42 | class FocalLoss(nn.Module): 43 | # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 44 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 45 | super(FocalLoss, self).__init__() 46 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 47 | self.gamma = gamma 48 | self.alpha = alpha 49 | self.reduction = loss_fcn.reduction 50 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 51 | 52 | def forward(self, pred, true): 53 | loss = self.loss_fcn(pred, true) 54 | # p_t = torch.exp(-loss) 55 | # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability 56 | 57 | # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py 58 | pred_prob = torch.sigmoid(pred) # prob from logits 59 | p_t = true * pred_prob + (1 - true) * (1 - pred_prob) 60 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 61 | modulating_factor = (1.0 - p_t) ** self.gamma 62 | loss *= alpha_factor * modulating_factor 63 | 64 | if self.reduction == 'mean': 65 | return loss.mean() 66 | elif self.reduction == 'sum': 67 | return loss.sum() 68 | else: # 'none' 69 | return loss 70 | 71 | 72 | class QFocalLoss(nn.Module): 73 | # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) 74 | def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): 75 | super(QFocalLoss, self).__init__() 76 | self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() 77 | self.gamma = gamma 78 | self.alpha = alpha 79 | self.reduction = loss_fcn.reduction 80 | self.loss_fcn.reduction = 'none' # required to apply FL to each element 81 | 82 | def forward(self, pred, true): 83 | loss = self.loss_fcn(pred, true) 84 | 85 | pred_prob = torch.sigmoid(pred) # prob from logits 86 | alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) 87 | modulating_factor = torch.abs(true - pred_prob) ** self.gamma 88 | loss *= alpha_factor * modulating_factor 89 | 90 | if self.reduction == 'mean': 91 | return loss.mean() 92 | elif self.reduction == 'sum': 93 | return loss.sum() 94 | else: # 'none' 95 | return loss 96 | 97 | 98 | class ComputeLoss: 99 | # Compute losses 100 | def __init__(self, model, autobalance=False): 101 | super(ComputeLoss, self).__init__() 102 | 103 | self.multiAdapt = AutomaticWeightedLoss(4) 104 | # self.multiAdapt = AutomaticWeightedLoss(2) # ForAuto 105 | 106 | device = next(model.parameters()).device # get model device 107 | h = model.hyp # hyperparameters 108 | 109 | # Define criteria 110 | BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device)) 111 | BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device)) 112 | 113 | # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 114 | self.cp, self.cn = smooth_BCE(eps=0.0) 115 | 116 | # Focal loss 117 | g = h['fl_gamma'] # focal loss gamma 118 | if g > 0: 119 | BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) 120 | 121 | det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module 122 | # self.balance = {2: [3.56, 1.0], 3: [3.67, 1.0, 0.43], 4: [3.78, 1.0, 0.39, 0.22], 5: [3.88, 1.0, 0.37, 0.17, 0.10]}[det.nl] 123 | # # self.balance = [1.0] * det.nl 124 | # self.ssi = (det.stride == 16).nonzero(as_tuple=False).item() # stride 16 index 125 | self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7 126 | self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index 127 | self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance 128 | self.dgMSE = nn.MSELoss(reduce=True, size_average=True) 129 | for k in 'na', 'nc', 'nl', 'anchors': 130 | setattr(self, k, getattr(det, k)) 131 | 132 | def __call__(self, p, dgimgs, targets): # predictions, targets, model 133 | device = targets.device 134 | lcls, lbox, lobj, ldg= torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device) 135 | tcls, tbox, indices, anchors = self.build_targets(p[0], targets) # targets 136 | 137 | # Losses 138 | for i, pi in enumerate(p[0]): # layer index, layer predictions 139 | b, a, gj, gi = indices[i] # image, anchor, gridy, gridx 140 | tobj = torch.zeros_like(pi[..., 0], device=device) # target obj 141 | 142 | n = b.shape[0] # number of targets 143 | if n: 144 | ps = pi[b, a, gj, gi] # prediction subset corresponding to targets 145 | 146 | # Regression 147 | pxy = ps[:, :2].sigmoid() * 2. - 0.5 148 | pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i] 149 | pbox = torch.cat((pxy, pwh), 1) # predicted box 150 | iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target) 151 | lbox += (1.0 - iou).mean() # iou loss 152 | 153 | # Objectness 154 | tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio 155 | 156 | # Classification 157 | if self.nc > 1: # cls loss (only if multiple classes) 158 | t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets 159 | t[range(n), tcls[i]] = self.cp 160 | lcls += self.BCEcls(ps[:, 5:], t) # BCE 161 | 162 | # Append targets to text file 163 | # with open('targets.txt', 'a') as file: 164 | # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] 165 | 166 | obji = self.BCEobj(pi[..., 4], tobj) 167 | lobj += obji * self.balance[i] # obj loss 168 | if self.autobalance: 169 | self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item() 170 | 171 | # "loss for patch" 172 | # if p[1]!=None: 173 | # ldgFactor = [0 if x not in targets[:, 0] else 1 for x in range(len(dgimgs))] 174 | # dgimgs = torch.stack([dgimgs[x] if ldgFactor[x] else p[1][x] for x in range(len(dgimgs))]).detach() 175 | # ldg += self.dgMSE(p[1].float(), dgimgs.float()) 176 | 177 | if p[1]!=None: 178 | ldg += self.dgMSE(p[1].float(), dgimgs.float()) 179 | 180 | if self.autobalance: 181 | self.balance = [x / self.balance[self.ssi] for x in self.balance] 182 | lbox *= self.hyp['box'] 183 | lobj *= self.hyp['obj'] 184 | lcls *= self.hyp['cls'] 185 | # ldg *= 1 # ForAuto 186 | ldg *= 0.1 187 | 188 | bs = tobj.shape[0] 189 | 190 | loss = self.multiAdapt(lbox, lobj, lcls, ldg) 191 | # loss = self.multiAdapt((lbox+ lobj+ lcls), ldg) # ForAuto 192 | # loss = lbox + lobj + lcls + ldg 193 | a=torch.cat((lbox, lobj, lcls, ldg, loss)) 194 | b=a.detach() 195 | return loss * bs, torch.cat((lbox, lobj, lcls, ldg, loss)).detach(), self.multiAdapt.params.detach() 196 | 197 | def build_targets(self, p, targets): 198 | # Build targets for compute_loss(), input targets(image,class,x,y,w,h) 199 | na, nt = self.na, targets.shape[0] # number of anchors, targets 200 | tcls, tbox, indices, anch = [], [], [], [] 201 | gain = torch.ones(7, device=targets.device) # normalized to gridspace gain 202 | ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) 203 | targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices 204 | 205 | g = 0.5 # bias 206 | off = torch.tensor([[0, 0], 207 | [1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m 208 | # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm 209 | ], device=targets.device).float() * g # offsets 210 | 211 | for i in range(self.nl): 212 | anchors = self.anchors[i] 213 | gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain 214 | 215 | # Match targets to anchors 216 | t = targets * gain 217 | if nt: 218 | # Matches 219 | r = t[:, :, 4:6] / anchors[:, None] # wh ratio 220 | j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare 221 | # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) 222 | t = t[j] # filter 223 | 224 | # Offsets 225 | gxy = t[:, 2:4] # grid xy 226 | gxi = gain[[2, 3]] - gxy # inverse 227 | j, k = ((gxy % 1. < g) & (gxy > 1.)).T 228 | l, m = ((gxi % 1. < g) & (gxi > 1.)).T 229 | j = torch.stack((torch.ones_like(j), j, k, l, m)) 230 | t = t.repeat((5, 1, 1))[j] 231 | offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] 232 | else: 233 | t = targets[0] 234 | offsets = 0 235 | 236 | # Define 237 | b, c = t[:, :2].long().T # image, class 238 | gxy = t[:, 2:4] # grid xy 239 | gwh = t[:, 4:6] # grid wh 240 | gij = (gxy - offsets).long() 241 | gi, gj = gij.T # grid xy indices 242 | 243 | # Append 244 | a = t[:, 6].long() # anchor indices 245 | indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices 246 | tbox.append(torch.cat((gxy - gij, gwh), 1)) # box 247 | anch.append(anchors[a]) # anchors 248 | tcls.append(c) # class 249 | 250 | return tcls, tbox, indices, anch 251 | -------------------------------------------------------------------------------- /utils/metrics.py: -------------------------------------------------------------------------------- 1 | # Model validation metrics 2 | 3 | from pathlib import Path 4 | 5 | import matplotlib.pyplot as plt 6 | import numpy as np 7 | import torch 8 | 9 | from . import general 10 | 11 | 12 | def fitness(x): 13 | # Model fitness as a weighted combination of metrics 14 | w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] 15 | return (x[:, :4] * w).sum(1) 16 | 17 | 18 | def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='precision-recall_curve.png', names=[]): 19 | """ Compute the average precision, given the recall and precision curves. 20 | Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. 21 | # Arguments 22 | tp: True positives (nparray, nx1 or nx10). 23 | conf: Objectness value from 0-1 (nparray). 24 | pred_cls: Predicted object classes (nparray). 25 | target_cls: True object classes (nparray). 26 | plot: Plot precision-recall curve at mAP@0.5 27 | save_dir: Plot save directory 28 | # Returns 29 | The average precision as computed in py-faster-rcnn. 30 | """ 31 | 32 | # Sort by objectness 33 | i = np.argsort(-conf) 34 | tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] 35 | 36 | # Find unique classes 37 | unique_classes = np.unique(target_cls) 38 | 39 | # Create Precision-Recall curve and compute AP for each class 40 | px, py = np.linspace(0, 1, 1000), [] # for plotting 41 | pr_score = 0.1 # score to evaluate P and R https://github.com/ultralytics/yolov3/issues/898 42 | s = [unique_classes.shape[0], tp.shape[1]] # number class, number iou thresholds (i.e. 10 for mAP0.5...0.95) 43 | ap, p, r = np.zeros(s), np.zeros(s), np.zeros(s) 44 | for ci, c in enumerate(unique_classes): 45 | i = pred_cls == c 46 | n_l = (target_cls == c).sum() # number of labels 47 | n_p = i.sum() # number of predictions 48 | 49 | if n_p == 0 or n_l == 0: 50 | continue 51 | else: 52 | # Accumulate FPs and TPs 53 | fpc = (1 - tp[i]).cumsum(0) 54 | tpc = tp[i].cumsum(0) 55 | 56 | # Recall 57 | recall = tpc / (n_l + 1e-16) # recall curve 58 | r[ci] = np.interp(-pr_score, -conf[i], recall[:, 0]) # r at pr_score, negative x, xp because xp decreases 59 | 60 | # Precision 61 | precision = tpc / (tpc + fpc) # precision curve 62 | p[ci] = np.interp(-pr_score, -conf[i], precision[:, 0]) # p at pr_score 63 | 64 | # AP from recall-precision curve 65 | for j in range(tp.shape[1]): 66 | ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) 67 | if plot and (j == 0): 68 | py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 69 | 70 | # Compute F1 score (harmonic mean of precision and recall) 71 | f1 = 2 * p * r / (p + r + 1e-16) 72 | 73 | if plot: 74 | plot_pr_curve(px, py, ap, save_dir, names) 75 | 76 | return p, r, ap, f1, unique_classes.astype('int32') 77 | 78 | 79 | def compute_ap(recall, precision): 80 | """ Compute the average precision, given the recall and precision curves 81 | # Arguments 82 | recall: The recall curve (list) 83 | precision: The precision curve (list) 84 | # Returns 85 | Average precision, precision curve, recall curve 86 | """ 87 | 88 | # Append sentinel values to beginning and end 89 | mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01])) 90 | mpre = np.concatenate(([1.], precision, [0.])) 91 | 92 | # Compute the precision envelope 93 | mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) 94 | 95 | # Integrate area under curve 96 | method = 'interp' # methods: 'continuous', 'interp' 97 | if method == 'interp': 98 | x = np.linspace(0, 1, 101) # 101-point interp (COCO) 99 | ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate 100 | else: # 'continuous' 101 | i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes 102 | ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve 103 | 104 | return ap, mpre, mrec 105 | 106 | 107 | class ConfusionMatrix: 108 | # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix 109 | def __init__(self, nc, conf=0.25, iou_thres=0.45): 110 | self.matrix = np.zeros((nc + 1, nc + 1)) 111 | self.nc = nc # number of classes 112 | self.conf = conf 113 | self.iou_thres = iou_thres 114 | 115 | def process_batch(self, detections, labels): 116 | """ 117 | Return intersection-over-union (Jaccard index) of boxes. 118 | Both sets of boxes are expected to be in (x1, y1, x2, y2) format. 119 | Arguments: 120 | detections (Array[N, 6]), x1, y1, x2, y2, conf, class 121 | labels (Array[M, 5]), class, x1, y1, x2, y2 122 | Returns: 123 | None, updates confusion matrix accordingly 124 | """ 125 | detections = detections[detections[:, 4] > self.conf] 126 | gt_classes = labels[:, 0].int() 127 | detection_classes = detections[:, 5].int() 128 | iou = general.box_iou(labels[:, 1:], detections[:, :4]) 129 | 130 | x = torch.where(iou > self.iou_thres) 131 | if x[0].shape[0]: 132 | matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() 133 | if x[0].shape[0] > 1: 134 | matches = matches[matches[:, 2].argsort()[::-1]] 135 | matches = matches[np.unique(matches[:, 1], return_index=True)[1]] 136 | matches = matches[matches[:, 2].argsort()[::-1]] 137 | matches = matches[np.unique(matches[:, 0], return_index=True)[1]] 138 | else: 139 | matches = np.zeros((0, 3)) 140 | 141 | n = matches.shape[0] > 0 142 | m0, m1, _ = matches.transpose().astype(np.int16) 143 | for i, gc in enumerate(gt_classes): 144 | j = m0 == i 145 | if n and sum(j) == 1: 146 | self.matrix[gc, detection_classes[m1[j]]] += 1 # correct 147 | else: 148 | self.matrix[self.nc, gc] += 1 # background FP 149 | 150 | if n: 151 | for i, dc in enumerate(detection_classes): 152 | if not any(m1 == i): 153 | self.matrix[dc, self.nc] += 1 # background FN 154 | 155 | def matrix(self): 156 | return self.matrix 157 | 158 | def plot(self, save_dir='', names=()): 159 | try: 160 | import seaborn as sn 161 | 162 | array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize 163 | array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) 164 | 165 | fig = plt.figure(figsize=(12, 9), tight_layout=True) 166 | sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size 167 | labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels 168 | sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True, 169 | xticklabels=names + ['background FP'] if labels else "auto", 170 | yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) 171 | fig.axes[0].set_xlabel('True') 172 | fig.axes[0].set_ylabel('Predicted') 173 | fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) 174 | except Exception as e: 175 | pass 176 | 177 | def print(self): 178 | for i in range(self.nc + 1): 179 | print(' '.join(map(str, self.matrix[i]))) 180 | 181 | 182 | # Plots ---------------------------------------------------------------------------------------------------------------- 183 | 184 | def plot_pr_curve(px, py, ap, save_dir='.', names=()): 185 | fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) 186 | py = np.stack(py, axis=1) 187 | 188 | if 0 < len(names) < 21: # show mAP in legend if < 10 classes 189 | for i, y in enumerate(py.T): 190 | ax.plot(px, y, linewidth=1, label=f'{names[i]} %.3f' % ap[i, 0]) # plot(recall, precision) 191 | else: 192 | ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) 193 | 194 | ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) 195 | ax.set_xlabel('Recall') 196 | ax.set_ylabel('Precision') 197 | ax.set_xlim(0, 1) 198 | ax.set_ylim(0, 1) 199 | plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") 200 | fig.savefig(Path(save_dir) / 'precision_recall_curve.png', dpi=250) 201 | -------------------------------------------------------------------------------- /utils/plots.py: -------------------------------------------------------------------------------- 1 | # Plotting utils 2 | 3 | import glob 4 | import math 5 | import os 6 | import random 7 | from copy import copy 8 | from pathlib import Path 9 | 10 | import cv2 11 | import matplotlib 12 | import matplotlib.pyplot as plt 13 | import numpy as np 14 | import pandas as pd 15 | import seaborn as sns 16 | import torch 17 | import yaml 18 | from PIL import Image, ImageDraw 19 | from scipy.signal import butter, filtfilt 20 | 21 | from utils.general import xywh2xyxy, xyxy2xywh 22 | from utils.metrics import fitness 23 | 24 | # Settings 25 | matplotlib.rc('font', **{'size': 11}) 26 | matplotlib.use('Agg') # for writing to files only 27 | 28 | 29 | def color_list(): 30 | # Return first 10 plt colors as (r,g,b) https://stackoverflow.com/questions/51350872/python-from-color-name-to-rgb 31 | def hex2rgb(h): 32 | return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) 33 | 34 | return [hex2rgb(h) for h in plt.rcParams['axes.prop_cycle'].by_key()['color']] 35 | 36 | 37 | def hist2d(x, y, n=100): 38 | # 2d histogram used in labels.png and evolve.png 39 | xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) 40 | hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) 41 | xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) 42 | yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) 43 | return np.log(hist[xidx, yidx]) 44 | 45 | 46 | def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): 47 | # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy 48 | def butter_lowpass(cutoff, fs, order): 49 | nyq = 0.5 * fs 50 | normal_cutoff = cutoff / nyq 51 | return butter(order, normal_cutoff, btype='low', analog=False) 52 | 53 | b, a = butter_lowpass(cutoff, fs, order=order) 54 | return filtfilt(b, a, data) # forward-backward filter 55 | 56 | 57 | def plot_one_box(x, img, color=None, label=None, line_thickness=3): 58 | # Plots one bounding box on image img 59 | tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness 60 | color = color or [random.randint(0, 255) for _ in range(3)] 61 | c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3])) 62 | cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA) 63 | if label: 64 | tf = max(tl - 1, 1) # font thickness 65 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 66 | c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3 67 | cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled 68 | cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA) 69 | 70 | 71 | def plot_wh_methods(): # from utils.plots import *; plot_wh_methods() 72 | # Compares the two methods for width-height anchor multiplication 73 | # https://github.com/ultralytics/yolov3/issues/168 74 | x = np.arange(-4.0, 4.0, .1) 75 | ya = np.exp(x) 76 | yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2 77 | 78 | fig = plt.figure(figsize=(6, 3), tight_layout=True) 79 | plt.plot(x, ya, '.-', label='YOLOv3') 80 | plt.plot(x, yb ** 2, '.-', label='YOLOv5 ^2') 81 | plt.plot(x, yb ** 1.6, '.-', label='YOLOv5 ^1.6') 82 | plt.xlim(left=-4, right=4) 83 | plt.ylim(bottom=0, top=6) 84 | plt.xlabel('input') 85 | plt.ylabel('output') 86 | plt.grid() 87 | plt.legend() 88 | fig.savefig('comparison.png', dpi=200) 89 | 90 | 91 | def output_to_target(output): 92 | # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] 93 | targets = [] 94 | for i, o in enumerate(output): 95 | for *box, conf, cls in o.cpu().numpy(): 96 | targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) 97 | return np.array(targets) 98 | 99 | def plot_images_together(images, targets, out, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): 100 | if isinstance(images, torch.Tensor): 101 | images = images.cpu().float().numpy() 102 | if isinstance(targets, torch.Tensor): 103 | targets = targets.cpu().numpy() 104 | if isinstance(out, torch.Tensor): 105 | out = out.cpu().numpy() 106 | if images.shape[1] != 3: 107 | images = images[:, :3, :, :] 108 | 109 | # un-normalise 110 | if np.max(images[0]) <= 1: 111 | images *= 255 112 | 113 | tl = 1 # line thickness 114 | tf = max(tl - 1, 1) # font thickness 115 | bs, _, h, w = images.shape # batch size, _, height, width 116 | bs = min(bs, max_subplots) # limit plot images 117 | 118 | # Check if we should resize 119 | scale_factor = max_size / max(h, w) 120 | if scale_factor < 1: 121 | h = math.ceil(scale_factor * h) 122 | w = math.ceil(scale_factor * w) 123 | 124 | colors = color_list() # list of colors 125 | for i, img in enumerate(images): 126 | if i == max_subplots: # if last batch has fewer images than we expect 127 | break 128 | img = img.transpose(1, 2, 0) 129 | if scale_factor < 1: 130 | img = cv2.resize(img, (w, h)) 131 | 132 | if len(targets) > 0: 133 | image_targets = targets[targets[:, 0] == i] 134 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 135 | classes = image_targets[:, 1].astype('int') 136 | labels = image_targets.shape[1] == 6 # labels if no conf column 137 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 138 | 139 | if boxes.shape[1]: 140 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 141 | boxes[[0, 2]] *= w # scale to pixels 142 | boxes[[1, 3]] *= h 143 | elif scale_factor < 1: # absolute coords need scale if image scales 144 | boxes *= scale_factor 145 | boxes[[0, 2]] += int(0) 146 | boxes[[1, 3]] += int(0) 147 | for j, box in enumerate(boxes.T): 148 | cls = int(classes[j]) 149 | color = colors[cls % len(colors)+3] 150 | cls = names[cls] if names else cls 151 | if labels or conf[j] > 0.25: # 0.25 conf thresh 152 | label = '' 153 | img = cv2.resize(img, (w, h)) 154 | plot_one_box(box, img, label=label, color=color, line_thickness=tl) 155 | 156 | if len(out) > 0: 157 | image_targets = out[out[:, 0] == i] 158 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 159 | classes = image_targets[:, 1].astype('int') 160 | labels = image_targets.shape[1] == 6 # labels if no conf column 161 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 162 | 163 | if boxes.shape[1]: 164 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 165 | boxes[[0, 2]] *= w # scale to pixels 166 | boxes[[1, 3]] *= h 167 | elif scale_factor < 1: # absolute coords need scale if image scales 168 | boxes *= scale_factor 169 | boxes[[0, 2]] += int(0) 170 | boxes[[1, 3]] += int(0) 171 | for j, box in enumerate(boxes.T): 172 | cls = int(classes[j]) 173 | color = colors[cls % len(colors)] 174 | cls = names[cls] if names else cls 175 | if labels or conf[j] > 0.25: # 0.25 conf thresh 176 | label = '%s %.1f' % (cls, conf[j]) 177 | img = cv2.resize(img, (w, h)) 178 | plot_one_box(box, img, label=label, color=color, line_thickness=tl) 179 | cv2.imwrite( 180 | os.path.dirname(str(fname)) + os.path.sep + os.path.basename(paths[i])[:-4] + os.path.basename(str(fname)), 181 | cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) 182 | 183 | 184 | def plot_images_each(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): 185 | # Plot image grid with labels 186 | 187 | if isinstance(images, torch.Tensor): 188 | images = images.cpu().float().numpy() 189 | if isinstance(targets, torch.Tensor): 190 | targets = targets.cpu().numpy() 191 | if images.shape[1]!=3: 192 | images = images[:,:3,:,:] 193 | 194 | # un-normalise 195 | if np.max(images[0]) <= 1: 196 | images *= 255 197 | 198 | tl = 1 # line thickness 199 | tf = max(tl - 1, 1) # font thickness 200 | bs, _, h, w = images.shape # batch size, _, height, width 201 | bs = min(bs, max_subplots) # limit plot images 202 | 203 | # Check if we should resize 204 | scale_factor = max_size / max(h, w) 205 | if scale_factor < 1: 206 | h = math.ceil(scale_factor * h) 207 | w = math.ceil(scale_factor * w) 208 | 209 | colors = color_list() # list of colors 210 | for i, img in enumerate(images): 211 | if i == max_subplots: # if last batch has fewer images than we expect 212 | break 213 | img = img.transpose(1, 2, 0) 214 | if scale_factor < 1: 215 | img = cv2.resize(img, (w, h)) 216 | 217 | if len(targets) > 0: 218 | image_targets = targets[targets[:, 0] == i] 219 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 220 | classes = image_targets[:, 1].astype('int') 221 | labels = image_targets.shape[1] == 6 # labels if no conf column 222 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 223 | 224 | if boxes.shape[1]: 225 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 226 | boxes[[0, 2]] *= w # scale to pixels 227 | boxes[[1, 3]] *= h 228 | elif scale_factor < 1: # absolute coords need scale if image scales 229 | boxes *= scale_factor 230 | boxes[[0, 2]] += int(0) 231 | boxes[[1, 3]] += int(0) 232 | for j, box in enumerate(boxes.T): 233 | cls = int(classes[j]) 234 | color = colors[cls % len(colors)] 235 | cls = names[cls] if names else cls 236 | if labels or conf[j] > 0.25: # 0.25 conf thresh 237 | label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j]) 238 | img = cv2.resize(img, (w, h)) 239 | plot_one_box(box, img, label=label, color=color, line_thickness=tl) 240 | cv2.imwrite(os.path.dirname(str(fname))+os.path.sep+ os.path.basename(paths[i])[:-4]+os.path.basename(str(fname)),cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) 241 | 242 | def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16): 243 | # Plot image grid with labels 244 | 245 | if isinstance(images, torch.Tensor): 246 | images = images.cpu().float().numpy() 247 | if isinstance(targets, torch.Tensor): 248 | targets = targets.cpu().numpy() 249 | if images.shape[1]!=3: 250 | images = images[:,:3,:,:] 251 | 252 | # un-normalise 253 | if np.max(images[0]) <= 1: 254 | images *= 255 255 | 256 | tl = 3 # line thickness 257 | tf = max(tl - 1, 1) # font thickness 258 | bs, _, h, w = images.shape # batch size, _, height, width 259 | bs = min(bs, max_subplots) # limit plot images 260 | ns = np.ceil(bs ** 0.5) # number of subplots (square) 261 | 262 | # Check if we should resize 263 | scale_factor = max_size / max(h, w) 264 | if scale_factor < 1: 265 | h = math.ceil(scale_factor * h) 266 | w = math.ceil(scale_factor * w) 267 | 268 | colors = color_list() # list of colors 269 | mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init 270 | for i, img in enumerate(images): 271 | if i == max_subplots: # if last batch has fewer images than we expect 272 | break 273 | 274 | block_x = int(w * (i // ns)) 275 | block_y = int(h * (i % ns)) 276 | 277 | img = img.transpose(1, 2, 0) 278 | if scale_factor < 1: 279 | img = cv2.resize(img, (w, h)) 280 | 281 | mosaic[block_y:block_y + h, block_x:block_x + w, :] = img 282 | if len(targets) > 0: 283 | image_targets = targets[targets[:, 0] == i] 284 | boxes = xywh2xyxy(image_targets[:, 2:6]).T 285 | classes = image_targets[:, 1].astype('int') 286 | labels = image_targets.shape[1] == 6 # labels if no conf column 287 | conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred) 288 | 289 | if boxes.shape[1]: 290 | if boxes.max() <= 1.01: # if normalized with tolerance 0.01 291 | boxes[[0, 2]] *= w # scale to pixels 292 | boxes[[1, 3]] *= h 293 | elif scale_factor < 1: # absolute coords need scale if image scales 294 | boxes *= scale_factor 295 | boxes[[0, 2]] += block_x 296 | boxes[[1, 3]] += block_y 297 | for j, box in enumerate(boxes.T): 298 | cls = int(classes[j]) 299 | color = colors[cls % len(colors)] 300 | cls = names[cls] if names else cls 301 | if labels or conf[j] > 0.25: # 0.25 conf thresh 302 | label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j]) 303 | plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl) 304 | 305 | # Draw image filename labels 306 | if paths: 307 | label = Path(paths[i]).name[:40] # trim to 40 char 308 | t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0] 309 | cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf, 310 | lineType=cv2.LINE_AA) 311 | 312 | # Image border 313 | cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3) 314 | 315 | if fname: 316 | r = min(1280. / max(h, w) / ns, 1.0) # ratio to limit image size 317 | mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA) 318 | # cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB)) # cv2 save 319 | Image.fromarray(mosaic).save(fname) # PIL save 320 | return mosaic 321 | 322 | 323 | def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): 324 | # Plot LR simulating training for full epochs 325 | optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals 326 | y = [] 327 | for _ in range(epochs): 328 | scheduler.step() 329 | y.append(optimizer.param_groups[0]['lr']) 330 | plt.plot(y, '.-', label='LR') 331 | plt.xlabel('epoch') 332 | plt.ylabel('LR') 333 | plt.grid() 334 | plt.xlim(0, epochs) 335 | plt.ylim(0) 336 | plt.savefig(Path(save_dir) / 'LR.png', dpi=200) 337 | plt.close() 338 | 339 | 340 | def plot_test_txt(): # from utils.plots import *; plot_test() 341 | # Plot test.txt histograms 342 | x = np.loadtxt('test.txt', dtype=np.float32) 343 | box = xyxy2xywh(x[:, :4]) 344 | cx, cy = box[:, 0], box[:, 1] 345 | 346 | fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) 347 | ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) 348 | ax.set_aspect('equal') 349 | plt.savefig('hist2d.png', dpi=300) 350 | 351 | fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) 352 | ax[0].hist(cx, bins=600) 353 | ax[1].hist(cy, bins=600) 354 | plt.savefig('hist1d.png', dpi=200) 355 | 356 | 357 | def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() 358 | # Plot targets.txt histograms 359 | x = np.loadtxt('targets.txt', dtype=np.float32).T 360 | s = ['x targets', 'y targets', 'width targets', 'height targets'] 361 | fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) 362 | ax = ax.ravel() 363 | for i in range(4): 364 | ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std())) 365 | ax[i].legend() 366 | ax[i].set_title(s[i]) 367 | plt.savefig('targets.jpg', dpi=200) 368 | 369 | 370 | def plot_study_txt(path='study/', x=None): # from utils.plots import *; plot_study_txt() 371 | # Plot study.txt generated by test.py 372 | fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True) 373 | ax = ax.ravel() 374 | 375 | fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) 376 | for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolov5s', 'yolov5m', 'yolov5l', 'yolov5x']]: 377 | y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T 378 | x = np.arange(y.shape[1]) if x is None else np.array(x) 379 | s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)'] 380 | for i in range(7): 381 | ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) 382 | ax[i].set_title(s[i]) 383 | 384 | j = y[3].argmax() + 1 385 | ax2.plot(y[6, :j], y[3, :j] * 1E2, '.-', linewidth=2, markersize=8, 386 | label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) 387 | 388 | ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], 389 | 'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet') 390 | 391 | ax2.grid() 392 | ax2.set_yticks(np.arange(30, 60, 5)) 393 | ax2.set_xlim(0, 30) 394 | ax2.set_ylim(29, 51) 395 | ax2.set_xlabel('GPU Speed (ms/img)') 396 | ax2.set_ylabel('COCO AP val') 397 | ax2.legend(loc='lower right') 398 | plt.savefig('test_study.png', dpi=300) 399 | 400 | 401 | def plot_labels(labels, names=(), save_dir=Path(''), loggers=None): 402 | # plot dataset labels 403 | print('Plotting labels... ') 404 | c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes 405 | nc = int(c.max() + 1) # number of classes 406 | colors = color_list() 407 | x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) 408 | 409 | # seaborn correlogram 410 | sns.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9)) 411 | plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) 412 | plt.close() 413 | 414 | # matplotlib labels 415 | matplotlib.use('svg') # faster 416 | ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() 417 | ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) 418 | ax[0].set_ylabel('instances') 419 | if 0 < len(names) < 30: 420 | ax[0].set_xticks(range(len(names))) 421 | ax[0].set_xticklabels(names, rotation=90, fontsize=10) 422 | else: 423 | ax[0].set_xlabel('classes') 424 | sns.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9) 425 | sns.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9) 426 | 427 | # rectangles 428 | labels[:, 1:3] = 0.5 # center 429 | labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 430 | img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) 431 | for cls, *box in labels[:1000]: 432 | ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10]) # plot 433 | ax[1].imshow(img) 434 | ax[1].axis('off') 435 | 436 | for a in [0, 1, 2, 3]: 437 | for s in ['top', 'right', 'left', 'bottom']: 438 | ax[a].spines[s].set_visible(False) 439 | 440 | plt.savefig(save_dir / 'labels.jpg', dpi=200) 441 | matplotlib.use('Agg') 442 | plt.close() 443 | 444 | # loggers 445 | for k, v in loggers.items() or {}: 446 | if k == 'wandb' and v: 447 | v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}) 448 | 449 | 450 | def plot_evolution(yaml_file='data/hyp.finetune.yaml'): # from utils.plots import *; plot_evolution() 451 | # Plot hyperparameter evolution results in evolve.txt 452 | with open(yaml_file) as f: 453 | hyp = yaml.load(f, Loader=yaml.SafeLoader) 454 | x = np.loadtxt('evolve.txt', ndmin=2) 455 | f = fitness(x) 456 | # weights = (f - f.min()) ** 2 # for weighted results 457 | plt.figure(figsize=(10, 12), tight_layout=True) 458 | matplotlib.rc('font', **{'size': 8}) 459 | for i, (k, v) in enumerate(hyp.items()): 460 | y = x[:, i + 7] 461 | # mu = (y * weights).sum() / weights.sum() # best weighted result 462 | mu = y[f.argmax()] # best single result 463 | plt.subplot(6, 5, i + 1) 464 | plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none') 465 | plt.plot(mu, f.max(), 'k+', markersize=15) 466 | plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters 467 | if i % 5 != 0: 468 | plt.yticks([]) 469 | print('%15s: %.3g' % (k, mu)) 470 | plt.savefig('evolve.png', dpi=200) 471 | print('\nPlot saved as evolve.png') 472 | 473 | 474 | def profile_idetection(start=0, stop=0, labels=(), save_dir=''): 475 | # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection() 476 | ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel() 477 | s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS'] 478 | files = list(Path(save_dir).glob('frames*.txt')) 479 | for fi, f in enumerate(files): 480 | try: 481 | results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows 482 | n = results.shape[1] # number of rows 483 | x = np.arange(start, min(stop, n) if stop else n) 484 | results = results[:, x] 485 | t = (results[0] - results[0].min()) # set t0=0s 486 | results[0] = x 487 | for i, a in enumerate(ax): 488 | if i < len(results): 489 | label = labels[fi] if len(labels) else f.stem.replace('frames_', '') 490 | a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5) 491 | a.set_title(s[i]) 492 | a.set_xlabel('time (s)') 493 | # if fi == len(files) - 1: 494 | # a.set_ylim(bottom=0) 495 | for side in ['top', 'right']: 496 | a.spines[side].set_visible(False) 497 | else: 498 | a.remove() 499 | except Exception as e: 500 | print('Warning: Plotting error for %s; %s' % (f, e)) 501 | 502 | ax[1].legend() 503 | plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200) 504 | 505 | 506 | def plot_results_overlay(start=0, stop=0): # from utils.plots import *; plot_results_overlay() 507 | # Plot training 'results*.txt', overlaying train and val losses 508 | s = ['train', 'train', 'train', 'Precision', 'mAP@0.5', 'val', 'val', 'val', 'Recall', 'mAP@0.5:0.95'] # legends 509 | t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles 510 | for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')): 511 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 512 | n = results.shape[1] # number of rows 513 | x = range(start, min(stop, n) if stop else n) 514 | fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True) 515 | ax = ax.ravel() 516 | for i in range(5): 517 | for j in [i, i + 5]: 518 | y = results[j, x] 519 | ax[i].plot(x, y, marker='.', label=s[j]) 520 | # y_smooth = butter_lowpass_filtfilt(y) 521 | # ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j]) 522 | 523 | ax[i].set_title(t[i]) 524 | ax[i].legend() 525 | ax[i].set_ylabel(f) if i == 0 else None # add filename 526 | fig.savefig(f.replace('.txt', '.png'), dpi=200) 527 | 528 | 529 | def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''): 530 | # Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp') 531 | fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) 532 | ax = ax.ravel() 533 | s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall', 534 | 'val Box', 'val Objectness', 'val Classification', 'mAP@0.5', 'mAP@0.5:0.95'] 535 | if bucket: 536 | # files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id] 537 | files = ['results%g.txt' % x for x in id] 538 | c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id) 539 | os.system(c) 540 | else: 541 | files = list(Path(save_dir).glob('results*.txt')) 542 | assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir) 543 | for fi, f in enumerate(files): 544 | try: 545 | results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T 546 | n = results.shape[1] # number of rows 547 | x = range(start, min(stop, n) if stop else n) 548 | for i in range(10): 549 | y = results[i, x] 550 | if i in [0, 1, 2, 5, 6, 7]: 551 | y[y == 0] = np.nan # don't show zero loss values 552 | # y /= y[0] # normalize 553 | label = labels[fi] if len(labels) else f.stem 554 | ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8) 555 | ax[i].set_title(s[i]) 556 | # if i in [5, 6, 7]: # share train and val loss y axes 557 | # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) 558 | except Exception as e: 559 | print('Warning: Plotting error for %s; %s' % (f, e)) 560 | 561 | ax[1].legend() 562 | fig.savefig(Path(save_dir) / 'results.png', dpi=200) 563 | 564 | colors = color_list() 565 | pass -------------------------------------------------------------------------------- /utils/torch_utils.py: -------------------------------------------------------------------------------- 1 | # PyTorch utils 2 | 3 | import logging 4 | import math 5 | import os 6 | import subprocess 7 | import time 8 | from contextlib import contextmanager 9 | from copy import deepcopy 10 | from pathlib import Path 11 | 12 | import torch 13 | import torch.backends.cudnn as cudnn 14 | import torch.nn as nn 15 | import torch.nn.functional as F 16 | import torchvision 17 | 18 | try: 19 | import thop # for FLOPS computation 20 | except ImportError: 21 | thop = None 22 | logger = logging.getLogger(__name__) 23 | 24 | 25 | @contextmanager 26 | def torch_distributed_zero_first(local_rank: int): 27 | """ 28 | Decorator to make all processes in distributed training wait for each local_master to do something. 29 | """ 30 | if local_rank not in [-1, 0]: 31 | torch.distributed.barrier() 32 | yield 33 | if local_rank == 0: 34 | torch.distributed.barrier() 35 | 36 | 37 | def init_torch_seeds(seed=0): 38 | # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html 39 | torch.manual_seed(seed) 40 | if seed == 0: # slower, more reproducible 41 | cudnn.benchmark, cudnn.deterministic = False, True 42 | else: # faster, less reproducible 43 | cudnn.benchmark, cudnn.deterministic = True, False 44 | 45 | 46 | def git_describe(): 47 | # return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe 48 | if Path('.git').exists(): 49 | return subprocess.check_output('git describe --tags --long --always', shell=True).decode('utf-8')[:-1] 50 | else: 51 | return '' 52 | 53 | 54 | def select_device(device='', batch_size=None): 55 | # device = 'cpu' or '0' or '0,1,2,3' 56 | s = f'YOLOv5 {git_describe()} torch {torch.__version__} ' # string 57 | cpu = device.lower() == 'cpu' 58 | if cpu: 59 | os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False 60 | elif device: # non-cpu device requested 61 | os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable 62 | assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested' # check availability 63 | 64 | cuda = not cpu and torch.cuda.is_available() 65 | if cuda: 66 | n = torch.cuda.device_count() 67 | if n > 1 and batch_size: # check that batch_size is compatible with device_count 68 | assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}' 69 | space = ' ' * len(s) 70 | for i, d in enumerate(device.split(',') if device else range(n)): 71 | p = torch.cuda.get_device_properties(i) 72 | s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n" # bytes to MB 73 | else: 74 | s += 'CPU\n' 75 | 76 | logger.info(s) # skip a line 77 | return torch.device('cuda:0' if cuda else 'cpu') 78 | 79 | 80 | def time_synchronized(): 81 | # pytorch-accurate time 82 | if torch.cuda.is_available(): 83 | torch.cuda.synchronize() 84 | return time.time() 85 | 86 | 87 | def profile(x, ops, n=100, device=None): 88 | # profile a pytorch module or list of modules. Example usage: 89 | # x = torch.randn(16, 3, 640, 640) # input 90 | # m1 = lambda x: x * torch.sigmoid(x) 91 | # m2 = nn.SiLU() 92 | # profile(x, [m1, m2], n=100) # profile speed over 100 iterations 93 | 94 | device = device or torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') 95 | x = x.to(device) 96 | x.requires_grad = True 97 | print(torch.__version__, device.type, torch.cuda.get_device_properties(0) if device.type == 'cuda' else '') 98 | print(f"\n{'Params':>12s}{'GFLOPS':>12s}{'forward (ms)':>16s}{'backward (ms)':>16s}{'input':>24s}{'output':>24s}") 99 | for m in ops if isinstance(ops, list) else [ops]: 100 | m = m.to(device) if hasattr(m, 'to') else m # device 101 | m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m # type 102 | dtf, dtb, t = 0., 0., [0., 0., 0.] # dt forward, backward 103 | try: 104 | flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPS 105 | except: 106 | flops = 0 107 | 108 | for _ in range(n): 109 | t[0] = time_synchronized() 110 | y = m(x) 111 | t[1] = time_synchronized() 112 | try: 113 | _ = y.sum().backward() 114 | t[2] = time_synchronized() 115 | except: # no backward method 116 | t[2] = float('nan') 117 | dtf += (t[1] - t[0]) * 1000 / n # ms per op forward 118 | dtb += (t[2] - t[1]) * 1000 / n # ms per op backward 119 | 120 | s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' 121 | s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list' 122 | p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0 # parameters 123 | print(f'{p:12.4g}{flops:12.4g}{dtf:16.4g}{dtb:16.4g}{str(s_in):>24s}{str(s_out):>24s}') 124 | 125 | 126 | def is_parallel(model): 127 | return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) 128 | 129 | 130 | def intersect_dicts(da, db, exclude=()): 131 | # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values 132 | return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape} 133 | 134 | 135 | def initialize_weights(model): 136 | for m in model.modules(): 137 | t = type(m) 138 | if t is nn.Conv2d: 139 | pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 140 | elif t is nn.BatchNorm2d: 141 | m.eps = 1e-3 142 | m.momentum = 0.03 143 | elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]: 144 | m.inplace = True 145 | 146 | 147 | def find_modules(model, mclass=nn.Conv2d): 148 | # Finds layer indices matching module class 'mclass' 149 | return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] 150 | 151 | 152 | def sparsity(model): 153 | # Return global model sparsity 154 | a, b = 0., 0. 155 | for p in model.parameters(): 156 | a += p.numel() 157 | b += (p == 0).sum() 158 | return b / a 159 | 160 | 161 | def prune(model, amount=0.3): 162 | # Prune model to requested global sparsity 163 | import torch.nn.utils.prune as prune 164 | print('Pruning model... ', end='') 165 | for name, m in model.named_modules(): 166 | if isinstance(m, nn.Conv2d): 167 | prune.l1_unstructured(m, name='weight', amount=amount) # prune 168 | prune.remove(m, 'weight') # make permanent 169 | print(' %.3g global sparsity' % sparsity(model)) 170 | 171 | 172 | def fuse_conv_and_bn(conv, bn): 173 | # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ 174 | fusedconv = nn.Conv2d(conv.in_channels, 175 | conv.out_channels, 176 | kernel_size=conv.kernel_size, 177 | stride=conv.stride, 178 | padding=conv.padding, 179 | groups=conv.groups, 180 | bias=True).requires_grad_(False).to(conv.weight.device) 181 | 182 | # prepare filters 183 | w_conv = conv.weight.clone().view(conv.out_channels, -1) 184 | w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) 185 | fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.size())) 186 | 187 | # prepare spatial bias 188 | b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias 189 | b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) 190 | fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) 191 | 192 | return fusedconv 193 | 194 | 195 | def model_info(model, verbose=False, img_size=512): 196 | # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] 197 | n_p = sum(x.numel() for x in model.parameters()) # number parameters 198 | n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients 199 | if verbose: 200 | print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma')) 201 | for i, (name, p) in enumerate(model.named_parameters()): 202 | name = name.replace('module_list.', '') 203 | print('%5g %40s %9s %12g %20s %10.3g %10.3g' % 204 | (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) 205 | 206 | try: # FLOPS 207 | from thop import profile 208 | stride = int(model.stride.max()) if hasattr(model, 'stride') else 32 209 | stride=512 210 | img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input 211 | flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPS 212 | img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float 213 | fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPS 214 | except (ImportError, Exception): 215 | fs = '' 216 | 217 | logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") 218 | 219 | 220 | def load_classifier(name='resnet101', n=2): 221 | # Loads a pretrained model reshaped to n-class output 222 | model = torchvision.models.__dict__[name](pretrained=True) 223 | 224 | # ResNet model properties 225 | # input_size = [3, 224, 224] 226 | # input_space = 'RGB' 227 | # input_range = [0, 1] 228 | # mean = [0.485, 0.456, 0.406] 229 | # std = [0.229, 0.224, 0.225] 230 | 231 | # Reshape output to n classes 232 | filters = model.fc.weight.shape[1] 233 | model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True) 234 | model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True) 235 | model.fc.out_features = n 236 | return model 237 | 238 | 239 | def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416) 240 | # scales img(bs,3,y,x) by ratio constrained to gs-multiple 241 | if ratio == 1.0: 242 | return img 243 | else: 244 | h, w = img.shape[2:] 245 | s = (int(h * ratio), int(w * ratio)) # new size 246 | img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize 247 | if not same_shape: # pad/crop img 248 | h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)] 249 | return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean 250 | 251 | 252 | def copy_attr(a, b, include=(), exclude=()): 253 | # Copy attributes from b to a, options to only include [...] and to exclude [...] 254 | for k, v in b.__dict__.items(): 255 | if (len(include) and k not in include) or k.startswith('_') or k in exclude: 256 | continue 257 | else: 258 | setattr(a, k, v) 259 | 260 | 261 | class ModelEMA: 262 | """ Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models 263 | Keep a moving average of everything in the model state_dict (parameters and buffers). 264 | This is intended to allow functionality like 265 | https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage 266 | A smoothed version of the weights is necessary for some training schemes to perform well. 267 | This class is sensitive where it is initialized in the sequence of model init, 268 | GPU assignment and distributed training wrappers. 269 | """ 270 | 271 | def __init__(self, model, decay=0.9999, updates=0): 272 | # Create EMA 273 | self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA 274 | # if next(model.parameters()).device.type != 'cpu': 275 | # self.ema.half() # FP16 EMA 276 | self.updates = updates # number of EMA updates 277 | self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs) 278 | for p in self.ema.parameters(): 279 | p.requires_grad_(False) 280 | 281 | def update(self, model): 282 | # Update EMA parameters 283 | with torch.no_grad(): 284 | self.updates += 1 285 | d = self.decay(self.updates) 286 | 287 | msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict 288 | for k, v in self.ema.state_dict().items(): 289 | if v.dtype.is_floating_point: 290 | v *= d 291 | v += (1. - d) * msd[k].detach() 292 | 293 | def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): 294 | # Update EMA attributes 295 | copy_attr(self.ema, model, include, exclude) 296 | --------------------------------------------------------------------------------