对比实验--罗宇鑫

FLOPs-SSD

from nets.ssd import SSD300
import torchvision.models as models
from ptflops import get_model_complexity_info
def get_classes(classes_path): # 去取name和数量
    with open(classes_path, encoding='utf-8') as f:
        class_names = f.readlines()
    class_names = [c.strip() for c in class_names]
    return class_names, len(class_names)
if __name__ == "__main__":
    classes_path = 'model_data/voc_classes.txt'
    class_names, num_classes  = get_classes(classes_path)
    # 对应的模型创建
    myNet = SSD300(num_classes + 1, 'vgg')  
    # 根据层数 以及跑的图片大小进行设置
    flops, params = get_model_complexity_info(myNet, (3,416,416), as_strings=True, print_per_layer_stat=True)
    print("Flops: {}".format(flops))
    print("Params: " + params)

FLOPs-yolov4

from nets.yolo import YoloBody
from ptflops import get_model_complexity_info
def get_classes(classes_path): # 去取name和数量
    with open(classes_path, encoding='utf-8') as f:
        class_names = f.readlines()
    class_names = [c.strip() for c in class_names]
    return class_names, len(class_names)
if __name__ == "__main__":
    classes_path = 'model_data/voc_clothes_classes.txt'
    class_names, num_classes  = get_classes(classes_path)
    # 对应的模型创建
    anchors_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
    myNet = YoloBody(anchors_mask, num_classes)  
    # 根据层数 以及跑的图片大小进行设置
    flops, params = get_model_complexity_info(myNet, (3,416,416), as_strings=True, print_per_layer_stat=True)
    print("Flops: {}".format(flops))
    print("Params: " + params)

YoloV4（失败）

• ep008-loss7.070-val_loss6.145

• ep017-loss5.949-val_loss5.581

• ep030-loss5.037-val_loss4.922

• ep001-loss73.319-val_loss11.971

• ep002-loss10.649-val_loss8.547

• ep003-loss9.075-val_loss7.746

• ep026-loss5.229-val_loss5.078

• ep043-loss4.501-val_loss5.139

• ep052-loss4.376-val_loss5.250

SSD(成功)

• ep001-loss9.099-val_loss6.616

• ep006-loss6.029-val_loss5.649

• ep011-loss5.529-val_loss5.349

• ep018-loss5.089-val_loss5.075

• ep038-loss4.575-val_loss4.881

• ep061-loss4.020-val_loss4.222

• ep071-loss3.524-val_loss3.855

• ep057-loss4.357-val_loss4.379

• ep030-loss4.735-val_loss4.917

• ep065-loss3.776-val_loss3.973

• ep026-loss4.829-val_loss4.921

• FLOPs

YoloV4（不满意）

• ep006-loss6.976-val_loss6.224

• ep013-loss5.956-val_loss5.406

• ep027-loss4.986-val_loss4.570

• ep060-loss4.038-val_loss4.029

• ep100-loss3.251-val_loss3.683

• ep085-loss3.434-val_loss3.769

• ep070-loss3.645-val_loss3.861.pth

• ep093-loss3.337-val_loss3.662

• ep063-loss3.844-val_loss3.929

• ep066-loss3.726-val_loss3.892

• ep088-loss3.379-val_loss3.738

Faster（resnet50 不满意）

• ep002-loss1.465-val_loss1.353

• ep012-loss1.203-val_loss1.302

• ep024-loss1.124-val_loss1.162

• ep035-loss1.079-val_loss1.071

• ep045-loss1.062-val_loss1.171

• ep055-loss1.060-val_loss1.047

• ep065-loss0.998-val_loss1.175

• ep075-loss0.996-val_loss1.183

• ep085-loss0.973-val_loss1.119

• ep095-loss0.941-val_loss1.170

Faster(vgg 断开了)

• ep002-loss1.838-val_loss1.865

• ep011-loss1.564-val_loss1.710

• ep020-loss1.457-val_loss1.627

• ep031-loss1.373-val_loss1.585

Faster（vgg 最高逼近50）

• ep001-loss2.065-val_loss1.938

• ep002-loss1.866-val_loss1.888

• ep008-loss1.608-val_loss1.758

• ep014-loss1.534-val_loss1.660

• ep021-loss1.438-val_loss1.632

• ep031-loss1.394-val_loss1.580

• ep041-loss1.339-val_loss1.568

• ep051-loss1.357-val_loss1.554

• ep061-loss1.223-val_loss1.490

• ep071-loss1.146-val_loss1.439

• ep081-loss1.094-val_loss1.413

• ep091-loss1.076-val_loss1.413

• ep100-loss1.054-val_loss1.431

• ep075-loss1.144-val_loss1.413

• ep085-loss1.097-val_loss1.447

• ep095-loss1.062-val_loss1.390

• ep098-loss1.073-val_loss1.383

• ep087-loss1.084-val_loss1.418

• 
  | 权重 | mAP | | —- | —- | |
  - ep001-loss2.065-val_loss1.938
  | 8.5 | |
  - ep002-loss1.866-val_loss1.888
  | 12.55 | |
  - ep008-loss1.608-val_loss1.758
  | 22.41 | |
  - ep014-loss1.534-val_loss1.660
  | 25.99 | |
  - ep021-loss1.438-val_loss1.632
  | 28.93 | |
  - ep031-loss1.394-val_loss1.580
  | 32.33 | |
  - ep041-loss1.339-val_loss1.568
  | 33.94 | |
  - ep051-loss1.357-val_loss1.554
  | 33.79 | |
  - ep061-loss1.223-val_loss1.490
  | 41.51 | |
  - ep071-loss1.146-val_loss1.439
  | 44.79 | |
  - ep081-loss1.094-val_loss1.413
  | 47.28 | |
  - ep091-loss1.076-val_loss1.413
  | 45.23 | |
  - ep100-loss1.054-val_loss1.431
  | 46.78 | |
  - ep095-loss1.062-val_loss1.390
  | 47.01 | |
  - ep075-loss1.144-val_loss1.413
  | 45.02 | |
  - ep085-loss1.097-val_loss1.447
  | 45.18 | |
  - ep098-loss1.073-val_loss1.383
  | 47.47 | |
  - ep087-loss1.084-val_loss1.418
  | 46.05 |

YoloV4（不满意）

• ep002-loss8.676-val_loss7.802

• ep023-loss5.622-val_loss5.088

• ep045-loss4.705-val_loss4.343

• ep061-loss4.051-val_loss4.030

• ep086-loss3.541-val_loss3.919

• ep100-loss3.434-val_loss3.835

• ep070-loss3.370-val_loss3.732

• ep092-loss3.003-val_loss3.709

• ep100-loss2.901-val_loss3.688

• ep110-loss2.799-val_loss3.615

• ep119-loss2.881-val_loss3.661

• ep128-loss2.804-val_loss3.560

• ep141-loss2.784-val_loss3.626

• ep151-loss2.871-val_loss3.645

• ep160-loss2.802-val_loss3.651

YoloV4（成功）

• ep144-loss2.422-val_loss3.563.pth

• FLOPs