Skip to content

Low MAP when training out custom dataset #194

Open
@genghuan2005

Description

@genghuan2005

Describe the bug

I'm trying to train a yolo v9 model with COCO128 dataset. There are images and labels inside the dataset folder. inside each folder I have train and val folders.
--images
----train
------1.jpg
------2.jpg
----val
------3.jpg
------4.jpg
--labels
----train
------1.txt
------2.txt
----val
------3.txt
------4.txt

data in each txt file, labels are format like [x_min,y_min,x_max,y_max]

Then I Create a coco128.yml file and located in yolo/config/dataset/coco128.yml with this simple data:

path: Dataset
train: train
validation: val
class_num: 80
class_list: ['..', '..']#省略80个coco类别
auto_download:

I run this command to train a model with my custom dataset:
python yolo/lazy.py task=train task.data.batch_size=8 model=v9-c dataset=coco128 device=cuda

The training result like this. Only the MAP of the first epoch looks normal. After the first epoch, the MAP is nearly zero. Are there any question in my training steps?

Epoch ┃ Avg. Precision ┃ % ┃ Avg. Recall ┃ % ┃
┡━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━┩
│ 0 │ AP @ .5:.95 │ 40.53 │ AP @ .5 │ 51.54 │
│ 1 │ AP @ .5:.95 │ 0.01 │ AP @ .5 │ 0.05 │
│ │ │ │ │ │
│ 2 │ AP @ .5:.95 │ 0.00 │ AR maxDets 1 │ 0.00 │
│ 2 │ AP @ .5 │ 0.00 │ AR maxDets 10 │ 0.00 │
│ 2 │ AP @ .75 │ 0.00 │ AR maxDets 100 │ 0.00 │
│ 2 │ AP (small) │ 0.00 │ AR (small) │ 0.00 │
│ 2 │ AP (medium) │ 0.00 │ AR (medium) │ 0.00 │
│ 2 │ AP (large) │ 0.00 │ AR (large) │ 0.00 │

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions