YOLO5Landmark 实验记录
目录
- 前言
- yolo5s6
-
-
- 训练：
-
- yolo5s6 + NoPretrain
-
- 训练：
- 测试
- yolo7stem
-
- 训练
- 测试
- 4gpu+32bs训练
- yolo7 + RepVGGBlockStem
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem
-
- 训练：
- 测试 best.pt：
- yolo7M + RepStem
-
- 训练：
- 测试 best.pt：
- yolo7L + 4GPU + RepStem
-
- 训练：
- 测试 best.pt：
- yolo7L + 1GPU + RepStem
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepBot
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepBot + IKeypoint
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + ShuffleBot
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + Shuffle2Bot
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot (RepVGG Block试下，看能否解决维度大于2)
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot + Ikeypoint (RepVGG Block能解决维度大于2)
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot + Ikeypoint +DW-CNN KPoint
-
- 训练：
- 测试 best.pt：
- yolo7m + RepStem + RepShuffle2Bot + Ikeypoint
-
- 训练：2GPU
- 测试 best.pt：
- 训练（1GPU）
-
- 测试 best.pt：
- yolo7L + RepStem + RepShuffle2Bot + Ikeypoint
-
- 训练：4GPU
- 测试 best.pt：
- 训练（1GPU）
-
- 测试 best.pt：
- yolo5L + RepStem + RepShuffle2Bot + Ikeypoint (RepVGG Block能解决维度大于2)
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot (RepConv因为维度为1,导致fuse出错)
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot (RepConv后去掉BN层)
-
- 训练：
- 测试 best.pt：
- yolo7 + RepStem + RepShuffle2Bot (DW换成group/2)
-
- 训练：
- 测试 best.pt：
- 二、使用步骤
-
- 1.引入库
前言
这里记录每一次运行的实验记录
yolo5s6
本地运行：
训练：
Model Summary: 377 layers, 13149336 parameters, 13149336 gradients, 18.6 GFLOPS
## 测试 last.pt：
0.925 0.982 0.958 0.913
nme:0.04828772978002897
Speed: 4.0/0.6/4.7 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp4
### 测试 best.pt：
0.93 0.955 0.978 0.933
nme:0.04658510462525046
Speed: 4.1/0.7/4.8 ms inference/NMS/total per 640x640 image at batch-size 4
yolo5s6 + NoPretrain
本地运行：
训练：
377 layers, 13149336 parameters, 13149336 gradients, 18.6 GFLOPS
Transferred 155/480 items from weights/yolov5s.pt
测试
Fusing layers…
Model Summary: 300 layers, 13135096 parameters, 0 gradients, 18.4 GFLOPS
0.952 0.96 0.982 0.94
nme:0.04449336785480576
Speed: 4.0/0.6/4.6 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp6
yolo7stem
训练
Model Summary: 385 layers, 13142776 parameters, 13142776 gradients, 18.0 GFLOPS
Transferred 12/492 items from weights/yolov5s.pt
测试
Fusing layers…
Model Summary: 306 layers, 13128472 parameters, 0 gradients, 17.8 GFLOPS
0.946 0.951 0.979 0.931
nme:0.04381672315071253
Speed: 4.2/0.6/4.7 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp7
4gpu+32bs训练
all 554 554 0.923 0.968 0.979 0.931
nme:0.04370922955195993
Speed: 4.8/5.5/10.3 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp2
IOU=0.5 conf =0.1:
all 554 554 0.923 0.968 0.98 0.931
nme:0.04878550926798806
Speed: 4.7/3.8/8.5 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp3
yolo7 + RepVGGBlockStem
本地运行：
训练：
Model Summary: 395 layers, 13144024 parameters, 13144024 gradients, 18.2 GFLOPS
Transferred 30/504 items from weights/yolov5s6.pt
测试 best.pt：
Fusing layers…
Model Summary: 308 layers, 13128472 parameters, 10144 gradients, 17.8 GFLOPS
0.943 0.958 0.978 0.929
nme:0.04634316943127357
Speed: 4.2/0.6/4.8 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp5
yolo7 + RepStem
本地运行：
训练：
Model Summary: 393 layers, 13144024 parameters, 13144024 gradients, 18.2 GFLOPS
Transferred 12/504 items from weights/yolov5s.pt
测试 best.pt：
Fusing layers…
RepConv.fuse_repvgg_block
RepConv.fuse_repvgg_block
Model Summary: 306 layers, 13128472 parameters, 10144 gradients, 17.8 GFLOPS
0.952 0.939 0.981 0.935
nme:0.043321637678518515
Speed: 4.3/0.6/4.8 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp8
yolo7M + RepStem
python train.py --data w300_kpts.yaml --cfg yolov5m6_w300kpts_RepStem.yaml --weights weights/yolov5m6.pt --batch-size 32 --workers 4 --device 3 --img 640 --kpt-label
本地运行：
训练：
Model Summary: 528 layers, 36628248 parameters, 36628248 gradients, 53.6 GFLOPS
Transferred 42/684 items from weights/yolov5m6.pt
Image sizes 640 train, 640 test
Using 4 dataloader workers batch size 20
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.1 --iou 0.5 --batch-size 4 --device 2 --weights runs/train/exp6/weights/best.pt --kpt-label
测试 best.pt：
YOLOv5 � 2022-12-30 torch 1.11.0 CUDA:2 (NVIDIA GeForce RTX 3090, 24268.3125MB)
Fusing layers…
RepConv.fuse_repvgg_block
RepConv.fuse_repvgg_block
Model Summary: 411 layers, 36598104 parameters, 22128 gradients, 53.0 GFLOPS
all 554 554 0.931 0.977 0.983 0.936
nme:0.043028827631862665
Speed: 6.7/4.9/11.6 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp5
yolo7L + 4GPU + RepStem
python train.py --data w300_kpts.yaml --cfg yolov5l6_w300kpts_RepStem.yaml --weights weights/yolov5l6.pt --batch-size 16 --workers 4 --device 0,1,2,3 --img 640 --kpt-label
本地运行：
训练：
Model Summary: 663 layers, 78207960 parameters, 78207960 gradients, 119.4 GFLOPS
Transferred 54/864 items from weights/yolov5l6.pt
Scaled weight_decay = 0.00046875
Optimizer groups: 149 .bias, 149 conv.weight, 141 other
Image sizes 640 train, 640 test
Using 4 dataloader workers batch size 16
all 554 554 0.896 0.962 0.964 0.916
300 epochs completed in 15.514 hours.
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.1 --iou 0.5 --batch-size 4 --device 2 --weights runs/train/exp11/weights/best.pt --kpt-label
测试 best.pt：
YOLOv5 � 2022-12-30 torch 1.11.0 CUDA:2 (NVIDIA GeForce RTX 3090, 24268.3125MB)
Model Summary: 516 layers, 78158680 parameters, 38720 gradients, 118.4 GFLOPS
all 554 554 0.897 0.958 0.965 0.919
nme:0.04442429334286103
Speed: 6.2/0.7/6.9 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp6
yolo7L + 1GPU + RepStem
python train.py --data w300_kpts.yaml --cfg yolov5l6_w300kpts_RepStem.yaml --weights weights/yolov5l6.pt --batch-size 8 --workers 2 --device 0 --img 640 --kpt-label
本地运行：
训练：
Model Summary: 663 layers, 78207960 parameters, 78207960 gradients, 119.4 GFLOPS
Transferred 54/864 items from weights/yolov5l6.pt
Scaled weight_decay = 0.00046875
Optimizer groups: 149 .bias, 149 conv.weight, 141 other
Image sizes 640 train, 640 test
Using 2 dataloader workers batch size 8
0.956 0.946 0.981 0.933
300 epochs completed in 14.915 hours.
Optimizer stripped from runs/train/exp13/weights/last.pt, 157.0MB
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.001 --iou 0.65 --batch-size 4 --device 2 --weights runs/train/exp13/weights/best.pt --kpt-label
测试 best.pt：
YOLOv5 � 2022-12-30 torch 1.11.0 CUDA:2 (NVIDIA GeForce RTX 3090, 24268.3125MB)
Model Summary: 516 layers, 78158680 parameters, 38720 gradients, 118.4 GFLOPS
0.948 0.947 0.981 0.935
nme:0.04148737189304469
Speed: 7.7/2.9/10.5 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp7
yolo7 + RepStem + RepBot
本地运行：
训练：
Model Summary: 648 layers, 13497112 parameters, 13497112 gradients, 18.9 GFLOPS
Transferred 337/864 items from runs/train/exp5/weights/best.pt
测试 best.pt：
Model Summary: 411 layers, 13130488 parameters, 3083104 gradients, 17.8 GFLOPS
0.941 0.958 0.98 0.934
nme:0.04262262558182343
Speed: 4.7/0.6/5.3 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp9
yolo7 + RepStem + RepBot + IKeypoint
本地运行：
训练：
Model Summary: 658 layers, 13498464 parameters, 13498464 gradients, 18.9 GFLOPS
Transferred 862/872 items from runs/train/exp6/weights/best.pt
测试 best.pt：
Model Summary: 421 layers, 13131840 parameters, 3083104 gradients, 17.8 GFLOPS
0.934 0.975 0.98 0.938
nme:0.04131272826509372
Speed: 4.7/0.6/5.3 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp15
yolo7 + RepStem + ShuffleBot
本地运行：
训练：
Model Summary: 423 layers, 9911656 parameters, 9911656 gradients, 12.1 GFLOPS
Transferred 322/594 items from runs/train/exp5/weights/best.pt
测试 best.pt：
Fusing layers…
RepConv.fuse_repvgg_block
RepConv.fuse_repvgg_block
Model Summary: 366 layers, 9900136 parameters, 10144 gradients, 11.8 GFLOPS
0.951 0.942 0.977 0.929
nme:0.04308880185174582
Speed: 4.0/0.6/4.6 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp10
yolo7 + RepStem + Shuffle2Bot
本地运行：
训练：
Model Summary: 423 layers, 13254648 parameters, 13254648 gradients, 17.2 GFLOPS
Transferred 246/594 items from runs/train/exp8/weights/best.pt
测试 best.pt：
Fusing layers…
RepConv.fuse_repvgg_block
RepConv.fuse_repvgg_block
Model Summary: 366 layers, 13239864 parameters, 10144 gradients, 16.9 GFLOPS
0.936 0.955 0.978 0.931
nme:0.042702530044724414
Speed: 4.7/0.6/5.4 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp11
yolo7 + RepStem + RepShuffle2Bot (RepVGG Block试下，看能否解决维度大于2)
本地运行：
训练：
Model Summary: 558 layers, 13268760 parameters, 13268760 gradients, 17.3 GFLOPS
Transferred 577/834 items from runs/train/exp9/weights/best.pt
测试 best.pt：
switch_to_deploy
Model Summary: 411 layers, 13241880 parameters, 30304 gradients, 16.9 GFLOPS
0.922 0.968 0.979 0.932
nme:0.04306018083638277
Speed: 4.8/0.6/5.4 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp16
yolo7 + RepStem + RepShuffle2Bot + Ikeypoint (RepVGG Block能解决维度大于2)
本地运行：
训练：
Model Summary: 568 layers, 13270112 parameters, 13270112 gradients, 17.3 GFLOPS
Transferred 832/842 items from runs/train/exp13/weights/best.pt
测试 best.pt：
switch_to_deploy
Model Summary: 421 layers, 13243232 parameters, 30304 gradients, 16.9 GFLOPS
0.937 0.962 0.976 0.933
nme:0.039646686695125256
Speed: 4.8/0.7/5.5 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp17
yolo7 + RepStem + RepShuffle2Bot + Ikeypoint +DW-CNN KPoint
本地运行：
训练：
Model Summary: 588 layers, 13367392 parameters, 13367392 gradients, 17.5 GFLOPS
Transferred 832/858 items from runs/train/exp15/weights/best.pt
Scaled weight_decay = 0.0005
Optimizer groups: 157 .bias, 127 conv.weight, 141 other
Image sizes 640 train, 640 test
Using 2 dataloader workers
Logging results to runs/train/exp19
测试 best.pt：
LKA换成Conv2d-Det+ImplictA-Kpt
all 554 554 0.937 0.969 0.977 0.935
Model Summary: 421 layers, 13245608 parameters, 30304 gradients, 16.9 GFLOPS
nme:0.039575465426819015
Speed: 10.6/0.7/11.3 ms inference/NMS/total per 640x640 image at batch-size 1
Results saved to runs/test/exp21
yolo7m + RepStem + RepShuffle2Bot + Ikeypoint
训练：2GPU
python train.py --data w300_kpts.yaml --cfg yolov5m6_w300kpts_RepStem_RepShuffle2Bot_IKeypoint.yaml --weights runs/train/exp6/weights/best.pt --batch-size 16 --workers 4 --device 3,4 --img 640 --kpt-label
Logging results to runs/train/exp17
0.948 0.964 0.983 0.94
300 epochs completed in 9.100 hours.
Optimizer stripped from runs/train/exp17/weights/last.pt, 62.4MB
测试 best.pt：
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.001 --iou 0.65 --batch-size 4 --device 2 --weights runs/train/exp17/weights/best.pt --kpt-label
0.952 0.962 0.984 0.941
nme:0.041504019045823154
Speed: 7.6/2.2/9.9 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp9
训练（1GPU）
python train.py --data w300_kpts.yaml --cfg yolov5m6_w300kpts_RepStem_RepShuffle2Bot_IKeypoint.yaml --weights runs/train/exp6/weights/best.pt --batch-size 8 --workers 2 --device 7 --img 640 --kpt-label
Model Summary: 868 layers, 30795360 parameters, 30795360 gradients, 40.3 GFLOPS
Transferred 201/1352 items from runs/train/exp6/weights/best.pt
Image sizes 640 train, 640 test
Using 2 dataloader workers
Logging results to runs/train/exp18
测试 best.pt：
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.001 --iou 0.65 --batch-size 4 --device 2 --weights runs/train/exp18/weights/best.pt --kpt-label
0.938 0.962 0.975 0.93
nme:0.04204403050927501
Speed: 8.0/3.0/11.0 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp10
yolo7L + RepStem + RepShuffle2Bot + Ikeypoint
训练：4GPU
python train.py --data w300_kpts.yaml --cfg yolov5l6_w300kpts_RepStem_RepShuffle2Bot_IKeypoint.yaml --weights runs/train/exp13/weights/best.pt --batch-size 16 --workers 4 --device 3,4,5,6 --img 640 --kpt-label
Model Summary: 1168 layers, 56945568 parameters, 56945568 gradients, 75.5 GFLOPS
Transferred 201/1862 items from runs/train/exp13/weights/best.pt
Image sizes 640 train, 640 test
Using 4 dataloader workers
Logging results to runs/train/exp19
测试 best.pt：
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.001 --iou 0.65 --batch-size 4 --device 2 --weights runs/train/exp17/weights/best.pt --kpt-label
0.952 0.962 0.984 0.941
nme:0.041504019045823154
Speed: 7.6/2.2/9.9 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp9
训练（1GPU）
python train.py --data w300_kpts.yaml --cfg yolov5l6_w300kpts_RepStem_RepShuffle2Bot_IKeypoint.yaml --weights runs/train/exp13/weights/best.pt --batch-size 8 --workers 2 --device 7 --img 640 --kpt-label
Model Summary: 1168 layers, 56945568 parameters, 56945568 gradients, 75.5 GFLOPS
Transferred 201/1862 items from runs/train/exp13/weights/best.pt
Image sizes 640 train, 640 test
Using 2 dataloader workers
Logging results to runs/train/exp20
Starting training for 300 epochs…
测试 best.pt：
(yolo) root@f5e52d307978:~/Projects/yolo5landmark# python test300WV1.py --data w300_kpts.yaml --img 640 --conf 0.001 --iou 0.65 --batch-size 4 --device 2 --weights runs/train/exp18/weights/best.pt --kpt-label
0.938 0.962 0.975 0.93
nme:0.04204403050927501
Speed: 8.0/3.0/11.0 ms inference/NMS/total per 640x640 image at batch-size 4
Results saved to runs/test/exp10
yolo5L + RepStem + RepShuffle2Bot + Ikeypoint (RepVGG Block能解决维度大于2)
本地运行：
训练：
Model Summary: 1168 layers, 56945568 parameters, 56945568 gradients, 75.5 GFLOPS
Transferred 1862/1862 items from runs/train/exp17/weights/last.pt
测试 best.pt：
switch_to_deploy
Model Summary: 841 layers, 56841376 parameters, 159680 gradients, 74.2 GFLOPS
nme:0.041191850604322004
Speed: 19.4/0.7/20.2 ms inference/NMS/total per 640x640 image at batch-size 1
Results saved to runs/test/exp19
yolo7 + RepStem + RepShuffle2Bot (RepConv因为维度为1,导致fuse出错)
本地运行：
训练：
Model Summary: 543 layers, 13268760 parameters, 13268760 gradients, 17.3 GFLOPS
Transferred 577/834 items from runs/train/exp9/weights/best.pt
测试 best.pt：
Model Summary: 396 layers, 13241880 parameters, 30304 gradients, 16.9 GFLOPS
0.842 0.791 0.794 0.598
nme:0.1406149227627627
Speed: 4.8/0.6/5.4 ms inference/NMS/total per 640x640 image at batch-size 4
yolo7 + RepStem + RepShuffle2Bot (RepConv后去掉BN层)
本地运行：
训练：
Model Summary: 528 layers, 13264728 parameters, 13264728 gradients, 17.3 GFLOPS
Transferred 412/759 items from runs/train/exp9/weights/best.pt
测试 best.pt：
出错
yolo7 + RepStem + RepShuffle2Bot (DW换成group/2)
本地运行：
训练：
Model Summary: 543 layers, 9928792 parameters, 9928792 gradients, 12.2 GFLOPS
Transferred 246/834 items from runs/train/exp9/weights/best.pt
测试 best.pt：
出错
二、使用步骤
1.引入库
代码如下（示例）：
{
// Use IntelliSense to learn about possible attributes.
// Hover to view descriptions of existing attributes.
// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
"version": "0.2.0",
"configurations": [
{
"name": "Python: Current File",
"type": "python",
"request": "launch",
"program": "${file}",
"console": "integratedTerminal",
"args": [
// // preprocess data_300W or generate300W challenge and common dataset
// "data_300W"
// "COFW"
// "data_300W_COFW_WFLW"
// // train yolo_pose
// "--data","coco_kpts.yaml",
// "--cfg", "yolov5s6_kpts.yaml",
// "--weights", "weights/yolov5s6.pt",
// "--batch-size", "4",
// "--img", "640",
// "--kpt-label"
// // // train yolo_facepose
// "--data","w300_kpts.yaml",
// // "--cfg", "yolov5s6_w300kpts_SimpleRepStem_ShuffleCSPAIAuxDetect.yaml",
// "--cfg", "yolov5s6_w300kpts_SimpleRepStem_ShuffleCSPA.yaml", // exp90 117 118 122
// // "--cfg", "yolov5s6_w300kpts_SimpleRepStem_ResX.yaml", // exp77 + 113
// // "--cfg", "yolov5s6_w300kpts_SimpleRepStemRepBot.yaml", //119
// // "--cfg", "yolov5s6_w300kpts_SimpleRepStemGhostBot.yaml",
// // "--cfg", "yolov5s6_w300kpts_SimpleRepStem.yaml", //110
// // "--resume", "runs/train/exp72/weights/last.pt",
// // "--cfg", "yolov5s6_w300kpts_stem_rep.yaml",
// // "--resume", "runs/train/exp57/weights/last.pt",
// // "--cfg", "yolov5s6_w300kpts.yaml", //109
// // "--resume", "runs/train/exp109/weights/last.pt",
// // "--cfg", "yolov5s6_w300kpts_stem_yolo7.yaml",
// // "--cfg", "yolov5s6_w300kpts_repmore.yaml",
// // "--resume", "runs/train/exp49/weights/last.pt",
// // "--cfg", "yolov5s6_w300kpts_ti_lite.yaml",
// // "--resume", "runs/train/exp28/weights/last.pt",
// // "--resume", "runs/train/exp/weights/last.pt",
// // "--cfg", "yolov5s6_w300kpts_ti_lite_deconv.yaml",
// // "--resume", "runs/train/exp27/weights/last.pt",
// // "--cfg", "yolov5l6_w300kpts_ti_lite.yaml",
// // "--resume", "runs/train/exp22/weights/last.pt",
// "--weights", "weights/yolov5s.pt",
// // "--weights", "weights/yolov5s6_640_ti_lite.pt",
// "--batch-size", "8",
// "--img", "640",
// // "--img", "960",
// // "--workers","8",
// "--epochs","300",
// // "--bbox_interval","30",
// "--kpt-label",
// "--adam"
// // "--facekpt-label"
// // // python train.py --data coco_kpts.yaml --cfg yolov5s6_kpts.yaml --weights 'path to the pre-trained ckpts' --batch-size 64 --img 640 --kpt-label
// // //test yolo_facepose
// "--data","w300_kpts.yaml",
// // "--img", "960",
// "--img", "640",
// "--iou"," 0.5", // use default value 0.65
// "--conf", "0.1", //original
// // "--conf", "0.02",
// // "--conf", "0.3",
// "--batch-size","16",
// // "--weights", "weights/yolov5s6.pt",
// "--weights", "runs/train/exp118/weights/best.pt",
// "--kpt-label"
// //"--facekpt-label"
// // python test.py --data coco_kpts.yaml --img 960 --conf 0.001 --iou 0.65 --weights "path to the pre-trained ckpt" --kpt-label
// to run detect_face
"--img-size", "800",
"--iou"," 0.65",
"--conf", "0.003", //original
// "--conf", "0.1",
"--weights", "runs/train/exp19/weights/best.pt",
"--source", "data/images/zidane.jpg"
// "--image", "data/images/jordan.jpg"
// "--image", "data/images/she2.jpg"
// "--image", "data/images/lfpw_testset_image_0178.png"
// "--image", "data/yolo_300W/images/test/ibug_image_047_1.jpg"
// "--image", "data/yolo_300W/images/test/ibug_image_048.jpg" //bad
// "--image", "data/yolo_300W/images/test/ibug_image_076_1.jpg"
// "--image", "data/yolo_300W/images/test/ibug_image_065_1.jpg"
// "--image", "data/yolo_300W/images/test/helen_testset_2988557119_1.jpg"
// "--image", "data/yolo_300W/images/train/helen_trainset_2203538277_1.jpg"
// "--image", "data/yolo_300W/images/test/helen_testset_315336719_1.jpg"
// "--image", "data/yolo_300W/images/test/helen_testset_2973812613_1.jpg"
// "--kpt-label"
// // to export model
// "--weights", "runs/train/exp19/weights/best.pt",
// "--batch", "1",
// "--img", "640",
// "--simplify",
// "--export-nms"
// // python export.py --weights "path to the pre-trained ckpt" --img 640 --batch 1 --simplify --export-nms # export at 640x640 with batch size 1
]
}
]
}
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import warnings
warnings.filterwarnings('ignore')
import ssl
ssl._create_default_https_context = ssl._create_unverified_context