search

meituan / YOLOv6

YOLOv6: a single-stage object detection framework dedicated to industrial applications.
GNU General Public License v3.0
5.71k stars 1.03k forks source link

map=-1，所有结果都是-1 #527

Open wxh1402714457 opened 2 years ago

wxh1402714457 commented 2 years ago

Before Asking

Search before asking

Question

无论batch怎么设置，结果都是 -1

Additional

No response

mtjhl commented 2 years ago

您好，请问您测试一下下载我们训练的模型，在 coco 上评估，是否 mAP 也是 －1 呢？ 另外，您是在自己的数据上训练，结果是 －1 吗？如果不是 coco 数据，是否在 [这里] (https://github.com/meituan/YOLOv6/blob/a76f0e6b3ebb5960d4b826b3a9be7621c7c6010f/data/coco.yaml#L9 )设置 is_coco 为 false 呢？

wxh1402714457 commented 2 years ago

您好，请问您测试一下下载我们训练的模型，在 coco 上评估，是否 mAP 也是 －1 呢？ 另外，您是在自己的数据上训练，结果是 －1 吗？如果不是 coco 数据，是否在 [这里] (

https://github.com/meituan/YOLOv6/blob/a76f0e6b3ebb5960d4b826b3a9be7621c7c6010f/data/coco.yaml#L9

)设置 is_coco 为 false 呢？

是的 设置了False，用的自己的数据集，我去改一下True继续测试

mtjhl commented 2 years ago

您可以评估一下我们发布的模型，在 coco 上评估是否正常。

riskeverything commented 2 years ago

How can you fix this bug? I have the same question like that all map is -1. Here is my command for train net: python tools/train.py --batch 64 --conf ./configs/yolov6s.py --data ./data/my.yaml --device 0 --output-dir $MNT_DIR/YOLOv6-main/result Here is my data:

train: /workspace/images/train
val: /workspace/images/test
# whether it is coco dataset, only coco dataset should be set to True.
is_coco: False
# Classes
nc: 7 # number of classes
names: ['face', 'person', 'pet', 'vehicle', 'plate', 'nonmotor', "head"]

Here is my dataset dir: /workspace/images/train /workspace/images/test /workspace/labels/train /workspace/labels/test

Chilicyy commented 2 years ago

@wxh1402714457 @riskeverything Hi, it may be caused by the wrong annotation cache-files created before, please add the args --check-images --check-labels in your training command and try again. Besides, if you want to train your custom dataset, we recommend you to use the config files with suffix of _finetune.py.

anhTuan0712 commented 2 years ago

@wxh1402714457 @riskeverything Hi, it may be caused by the wrong annotation cache-files created before, please add the args --check-images --check-labels in your training command and try again. Besides, if you want to train your custom dataset, we recommend you to use the config files with suffix of _finetune.py.

I tried all of these suggestions but mAP still = -1