YOLOv7 0.00 mAP

[emersondelemmus]

I am training on a one-class dataset. I made sure to change 'nc: 1' (number of classes = 1) in the yolov7.cfg. I am running it on anaconda and my data is being read correctly + training is executing. (AVG: Epoch Length 5 min -> 10 min @ 32 batch size).

This is my training command:
E:/yolov7/python train.py --batch 32 --img-size 640 --cfg E:/yolov7/cfg/training/yolov7.yaml --epochs 100 -- data E:/Trainig_Data/data.yaml --weights E:/yolov7/yolov7.pt --device 0

I set epochs to 100 just for a preliminary attempt. Overnight, I checked wandb.ai for the metrics at 67/100 epochs and there is no training done. mAP = 0.0, precision = 0.0, recall =0.0.

The training dataset should be completely fine because I used it to train YOLOv5 with no problems. Can anyone shed light on why mAP is stuck 0.0?

[Abandon-ht]

Can you provide more information about hyperparameters?

[emersondelemmus]

Can you provide more information about hyperparameters?

Yes! Of course.

E:/yolov7/python train.py --batch 32 --img-size 640
--cfg E:/yolov7/cfg/training/yolov7.yaml --epochs 100
--data E:/Trainig_Data/data.yaml
--hyp E:/yolov7/data/hyp.scratch.p5.yaml
--weights E:/yolov7/yolov7.pt --device 0

[emersondelemmus]

This is the end result incase it helps visualize my problem. For reference, training on the same dataset using YOLOv5 yielded 80% mAP.

[emersondelemmus]

Can you provide more information about hyperparameters?

Also I should mention the hyp.scratch.p5.yaml I am using is the default one provided by this github with no changes.

[Abandon-ht]

Can you provide more information about hyperparameters?

Also I should mention the hyp.scratch.p5.yaml I am using is the default one provided by this github with no changes.

In README.md:
If you have previously used a different version of YOLO, we strongly recommend that you delete train2017.cache and val2017.cache files, and redownload labels.
Did you finished it?

[emersondelemmus]

Can you provide more information about hyperparameters?

Also I should mention the hyp.scratch.p5.yaml I am using is the default one provided by this github with no changes.

In README.md: If you have previously used a different version of YOLO, we strongly recommend that you delete train2017.cache and val2017.cache files, and redownload labels. Did you finished it?

Correct, the .cache files were deleted and regenerated. I trained again last night, but I still received mAP 0.0 results over 100 epochs.

I am working with a tiled image dataset so I know my images contained a large number of Empty images, however, this should not be a problem.

[AlexeyAB]

Attach all charts from runs/train/
And check if it detect anything, to check if the issue is in training or validation.

[emersondelemmus]

Results:

Confusion Matrix

results.txt

Test Batch 0 Label

Test Batch 0 Prediction

Test Batch 2 Label

Test Batch 2 Predict

Attach all charts from runs/train/ And check if it detect anything, to check if the issue is in training or validation.

[parto-mahdi]

Hello,
I have also the same issue, I trained the network without changing anything in codes with my own dataset for both yolo-tiny7 and yolo7-p5 and only for one class and I have the same problem for both. I changed the number of classes in every possible places;
Yolov7-tiny:
--data custom.yaml , --cfg yolov7-tiny.yaml
Yolov7-p5:
--data custom.yaml , --cfg yolov7.yaml
my custom.yaml file:
train: yolo-tiny/tiny7/train
test: yolo-tiny/tiny7/test
val: yolo-tiny/tiny7/val
#Classes
nc: 1 # replace according to your number of classes
#classes names
#replace all class names list with your classes names
names: ['myobject']
Training was done successfully, 100 epochs and it is finding my object perfectly fine with the average of 0.96 for IoUs and MAP 1.00 based on IoUs
but when it comes to confidence scores, the average is 0.50 for tiny and 0.52 for p5 and most of them are between 0.49 and 0.50
because of this the Map for threshold 0.7 according to confidence scores is 0
I think it is a small problem in settings but I have no idea where, I would appreciate if you help me for this....

[parto-mahdi]

@AlexeyAB
I would appreciate if consider my problem or if I need to start a new topic just please let me know.
Thanks

[AlexeyAB]

@parto-mahdi Could you try to train using the latest code

• with loss_ota: 0 https://github.com/WongKinYiu/yolov7/blob/main/data/hyp.scratch.p5.yaml#L31
• or with all these parameters 0 https://github.com/WongKinYiu/yolov7/blob/main/data/hyp.scratch.p5.yaml#L27-L31

[Guiltytrees]

我遇到了同样的问题

[Guiltytrees]

是因为YOLOv7-tiny相比YOLOv5s、YOLOv4等更难收敛吗？同样的数据与参数设置，后两者可以正常训练