yjh0410 1 年之前
父节点
当前提交
b29a4ad336
共有 1 个文件被更改,包括 4 次插入4 次删除
  1. 4 4
      README.md

+ 4 - 4
README.md

@@ -34,13 +34,13 @@
 |-------------|-------|-------|--------------------------|-----------------------|-------------------|-------------------|--------------------|--------|------|
 | YOLOv1-R18  | 1xb16 |  640  |           124            |         27.6          |       46.8        |   37.8            |   21.3             | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov1_r18_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv1-R18-COCO.txt) |
 | YOLOv2-R18  | 1xb16 |  640  |           128            |         28.4          |       47.4        |   38.0            |   21.5             | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov2_r18_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv2-R18-COCO.txt) |
-| YOLOv3-S    | 1xb16 |  640  |           107            |         31.3          |       49.2        |   25.2            |   7.              | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov3_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv3-S-COCO.txt) |
+| YOLOv3-S    | 1xb16 |  640  |           107            |         31.3          |       49.2        |   25.2            |   7.3              | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov3_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv3-S-COCO.txt) |
 | YOLOv5-S    | 1xb16 |  640  |            80            |         38.8          |       56.9        |   27.3            |   9.0              | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov5_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv5-S-COCO.txt) |
 | YOLOv5-AF-S | 1xb16 |  640  |            83            |         39.6          |       58.7        |   26.9            |   8.9              | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov5_af_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv5-AF-S-COCO.txt) |
-| YOLOv8-S    | 1xb16 |  640  |            70            |         42.5          |       59.3        |   28.4            |   11.3            | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov8_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv8-S-COCO.txt) |
-| GELAN-S     | 1xb16 |  640  |            32            |         42.6          |       58.8        |   27.1 (26.4)     |   7.1 (7.2)            | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/gelan_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/GELAN-S-COCO.txt) |
+| YOLOv8-S    | 1xb16 |  640  |            79            |         42.5          |       59.3        |   28.4            |   11.3            | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/yolov8_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/YOLOv8-S-COCO.txt) |
+| GELAN-S     | 1xb16 |  640  |            34(38)       |         42.6          |       58.8        |   27.1 (26.4)     |   7.1 (7.2)            | [ckpt](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/gelan_s_coco.pth) | [log](https://github.com/yjh0410/YOLO-Tutorial-v2/releases/download/yolo_tutorial_ckpt/GELAN-S-COCO.txt) |
 
-需要说明的是,对于GELAN-S,未进行重参数化时,模型参数量为7.1 M,理论计算量为27.1 GFLOPs;经过重参数化处理后,模型参数量为7.2 M,理论计算量为26.4 GFLOPs。然而,GELAN-S的FPS很低,起初,以为是因为它的regression head部分用到了group=4的分组卷积,由于PyTorch框架本身没有对这个操作做优化,因此,虽然分组卷积的理论计算量会更低,但在不做加快优化的情况下,推理速度会慢于group=1的普通卷积,类似的现象在depthwise卷积中也能看到。但是,即便将group=4修改为group=1,依旧不超过35 FPS,远远低于YOLOv8-S的速度。
+需要说明的是,对于GELAN-S,未进行重参数化时,模型参数量为7.1 M,理论计算量为27.1 GFLOPs;经过重参数化处理后,模型参数量为7.2 M,理论计算量为26.4 GFLOPs。然而,GELAN-S的FPS很低,起初,以为是因为它的regression head部分用到了group=4的分组卷积,由于PyTorch框架本身没有对这个操作做优化,因此,虽然分组卷积的理论计算量会更低,但在不做加快优化的情况下,推理速度会慢于group=1的普通卷积,类似的现象在depthwise卷积中也能看到。但是,即便将group=4修改为group=1,依旧不超过40 FPS,显著低于YOLOv8-S的速度。
 
 ### RT-DETR系列
 下表汇报了本项目的RT-DETR系列在COCO数据集上的性能指标。所有模型都采用4张3090显卡训练的,在训练中,每张3090显卡上的batch size被设置为4,并使用多卡同步BN来计算BN层的统计量。需要说明的是,官方的RT-DETR所汇报的FPS指标,是经过各种加速处理后所测得的,因而会很高,而这里我们没有做加速处理,也没有编译CUDA版本的Deformable Attention算子,纯纯的PyTorch框架实现的,且使用的是4060显卡,而非诸如3090和V100等高算力显卡,因此,FPS指标会显著低于论文中所汇报的指标。