YOLOV8训练时loss为0

[kaixin-bai]

问题确认 Search before asking

• 我已经查询历史issue，没有发现相似的bug。I have searched the issues and found no similar bug report.

Bug组件 Bug Component

Training

Bug描述 Describe the Bug

在paddleyolo中使用yolov8训练自定义数据集时，显示loss全为0，对保存的checkpoint进行推理无法推理出结果。同样的数据集在paddledetection的yolov3上可以顺利训练并推理。

训练命令：

python3 tools/train.py -c ./configs/yolov8/yolov8_l_500e_domainR_RGB_coco.yml --use_vdl=true --vdl_log_dir=vdl_dir/yolov8_l_500e_domainR_RGB_coco/scalar

推理命令：

python3 tools/infer.py -c ./configs/yolov8/yolov8_l_500e_domainR_RGB_coco.yml --output_dir=./ --draw_threshold=0.0 --infer_img=./ceshi/input/rgb_0900.png -o weights=./output/yolov8_l_500e_domainR_RGB_coco/9.pdparams

配置文件：

_BASE_: [
  '../datasets/coco_detection_domainR_RGB.yml',
  '../runtime.yml',
  '_base_/optimizer_500e_high.yml',
  '_base_/yolov8_cspdarknet.yml',
  '_base_/yolov8_reader_high_aug.yml',
]
depth_mult: 1.0
width_mult: 1.0

log_iter: 50
snapshot_epoch: 10
weights: output/yolov8_l_500e_domainR_RGB_coco/model_final


YOLOv8CSPDarkNet:
  last_stage_ch: 512 # The actual channel is int(512 * width_mult), not int(1024 * width_mult) as in YOLOv5


TrainReader:
  batch_size: 8  # 16 # default 8 gpus, total bs = 128

(paddle) kb@ar-gpu01:/data-r10/kb/Projects/SynDataGen/PaddleYOLO$ python3 tools/train.py -c ./configs/yolov8/yolov8_l_500e_domainR_RGB_coco.yml --use_vdl=true --vdl_log_dir=vdl_dir/yolov8_l_500e_domainR_RGB_coco/scalar
/data-r10/kb/anaconda3/envs/paddle/lib/python3.8/site-packages/pkg_resources/__init__.py:121: DeprecationWarning: pkg_resources is deprecated as an API
  warnings.warn("pkg_resources is deprecated as an API", DeprecationWarning)
/data-r10/kb/anaconda3/envs/paddle/lib/python3.8/site-packages/pkg_resources/__init__.py:2870: DeprecationWarning: Deprecated call to `pkg_resources.declare_namespace('mpl_toolkits')`.
Implementing implicit namespace packages (as specified in PEP 420) is preferred to `pkg_resources.declare_namespace`. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packages
  declare_namespace(pkg)
/data-r10/kb/anaconda3/envs/paddle/lib/python3.8/site-packages/pkg_resources/__init__.py:2870: DeprecationWarning: Deprecated call to `pkg_resources.declare_namespace('google')`.
Implementing implicit namespace packages (as specified in PEP 420) is preferred to `pkg_resources.declare_namespace`. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packages
  declare_namespace(pkg)
Warning: import ppdet from source directory without installing, run 'python setup.py install' to install ppdet firstly
loading annotations into memory...
Done (t=3.93s)
creating index...
index created!
[07/07 16:50:19] ppdet.data.source.coco INFO: Load [900 samples valid, 0 samples invalid] in file /data-r10/kb/Projects/SynDataGen/datasets/domain_randomization_rgb_mps1/train.json.
W0707 16:50:20.097818 25033 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 6.1, Driver API Version: 11.7, Runtime API Version: 11.2
W0707 16:50:20.101763 25033 gpu_resources.cc:91] device: 0, cuDNN Version: 8.2.
[07/07 16:50:26] ppdet.engine INFO: Epoch: [0] [  0/113] eta: 2 days, 16:35:47 lr: 0.000000 loss: 478.704224 loss_cls: 478.704224 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 4.1159 data_cost: 1.6733 ips: 1.9437 images/s
[07/07 16:51:08] ppdet.engine INFO: Epoch: [0] [ 50/113] eta: 13:29:23 lr: 0.000218 loss: 247.558243 loss_cls: 247.558243 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.7952 data_cost: 0.2914 ips: 10.0607 images/s
[07/07 16:51:53] ppdet.engine INFO: Epoch: [0] [100/113] eta: 13:30:54 lr: 0.000870 loss: 1.999204 loss_cls: 1.999204 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8651 data_cost: 0.3560 ips: 9.2475 images/s
[07/07 16:52:07] ppdet.engine INFO: Epoch: [1] [  0/113] eta: 13:46:36 lr: 0.001111 loss: 1.268529 loss_cls: 1.268529 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9100 data_cost: 0.3909 ips: 8.7912 images/s
[07/07 16:52:51] ppdet.engine INFO: Epoch: [1] [ 50/113] eta: 13:34:34 lr: 0.002312 loss: 0.301968 loss_cls: 0.301968 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8401 data_cost: 0.3263 ips: 9.5229 images/s
[07/07 16:53:35] ppdet.engine INFO: Epoch: [1] [100/113] eta: 13:31:01 lr: 0.003948 loss: 0.102459 loss_cls: 0.102459 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8547 data_cost: 0.3380 ips: 9.3601 images/s
[07/07 16:53:49] ppdet.engine INFO: Epoch: [2] [  0/113] eta: 13:39:27 lr: 0.004444 loss: 0.079679 loss_cls: 0.079679 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9041 data_cost: 0.3965 ips: 8.8488 images/s
[07/07 16:54:32] ppdet.engine INFO: Epoch: [2] [ 50/113] eta: 13:30:48 lr: 0.006629 loss: 0.042159 loss_cls: 0.042159 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8269 data_cost: 0.3053 ips: 9.6741 images/s
[07/07 16:55:17] ppdet.engine INFO: Epoch: [2] [100/113] eta: 13:29:32 lr: 0.009248 loss: 0.012407 loss_cls: 0.012407 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8614 data_cost: 0.3499 ips: 9.2871 images/s
[07/07 16:55:31] ppdet.engine INFO: Epoch: [3] [  0/113] eta: 13:34:04 lr: 0.010000 loss: 0.011372 loss_cls: 0.011372 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9075 data_cost: 0.3917 ips: 8.8153 images/s
[07/07 16:56:14] ppdet.engine INFO: Epoch: [3] [ 50/113] eta: 13:26:51 lr: 0.009946 loss: 0.004152 loss_cls: 0.004152 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8156 data_cost: 0.3029 ips: 9.8085 images/s
[07/07 16:56:58] ppdet.engine INFO: Epoch: [3] [100/113] eta: 13:25:00 lr: 0.009946 loss: 0.001564 loss_cls: 0.001564 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8520 data_cost: 0.3256 ips: 9.3893 images/s
[07/07 16:57:12] ppdet.engine INFO: Epoch: [4] [  0/113] eta: 13:27:49 lr: 0.009946 loss: 0.001540 loss_cls: 0.001540 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8985 data_cost: 0.3776 ips: 8.9034 images/s
[07/07 16:57:55] ppdet.engine INFO: Epoch: [4] [ 50/113] eta: 13:24:32 lr: 0.009946 loss: 0.001526 loss_cls: 0.001526 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8372 data_cost: 0.3240 ips: 9.5555 images/s
[07/07 16:58:40] ppdet.engine INFO: Epoch: [4] [100/113] eta: 13:22:30 lr: 0.009946 loss: 0.001526 loss_cls: 0.001526 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8464 data_cost: 0.3216 ips: 9.4520 images/s
[07/07 16:58:55] ppdet.engine INFO: Epoch: [5] [  0/113] eta: 13:27:25 lr: 0.009928 loss: 0.001525 loss_cls: 0.001525 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9087 data_cost: 0.3855 ips: 8.8035 images/s
[07/07 16:59:38] ppdet.engine INFO: Epoch: [5] [ 50/113] eta: 13:24:35 lr: 0.009928 loss: 0.001523 loss_cls: 0.001523 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8382 data_cost: 0.3152 ips: 9.5445 images/s
[07/07 17:00:23] ppdet.engine INFO: Epoch: [5] [100/113] eta: 13:23:55 lr: 0.009928 loss: 0.001500 loss_cls: 0.001500 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8646 data_cost: 0.3505 ips: 9.2525 images/s
[07/07 17:00:36] ppdet.engine INFO: Epoch: [6] [  0/113] eta: 13:25:56 lr: 0.009910 loss: 0.001483 loss_cls: 0.001483 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9161 data_cost: 0.4078 ips: 8.7323 images/s
[07/07 17:01:20] ppdet.engine INFO: Epoch: [6] [ 50/113] eta: 13:23:23 lr: 0.009910 loss: 0.000978 loss_cls: 0.000978 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8377 data_cost: 0.3267 ips: 9.5498 images/s
[07/07 17:02:03] ppdet.engine INFO: Epoch: [6] [100/113] eta: 13:20:28 lr: 0.009910 loss: 0.000144 loss_cls: 0.000144 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8274 data_cost: 0.3139 ips: 9.6688 images/s
[07/07 17:02:17] ppdet.engine INFO: Epoch: [7] [  0/113] eta: 13:22:47 lr: 0.009892 loss: 0.000080 loss_cls: 0.000080 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8882 data_cost: 0.3785 ips: 9.0067 images/s
[07/07 17:03:01] ppdet.engine INFO: Epoch: [7] [ 50/113] eta: 13:20:49 lr: 0.009892 loss: 0.000017 loss_cls: 0.000017 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8421 data_cost: 0.3204 ips: 9.5006 images/s
[07/07 17:03:46] ppdet.engine INFO: Epoch: [7] [100/113] eta: 13:20:00 lr: 0.009892 loss: 0.000007 loss_cls: 0.000007 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8613 data_cost: 0.3310 ips: 9.2879 images/s
[07/07 17:04:00] ppdet.engine INFO: Epoch: [8] [  0/113] eta: 13:21:21 lr: 0.009874 loss: 0.000005 loss_cls: 0.000005 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.9011 data_cost: 0.3771 ips: 8.8778 images/s
[07/07 17:04:42] ppdet.engine INFO: Epoch: [8] [ 50/113] eta: 13:18:15 lr: 0.009874 loss: 0.000000 loss_cls: 0.000000 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8158 data_cost: 0.3034 ips: 9.8058 images/s
[07/07 17:05:27] ppdet.engine INFO: Epoch: [8] [100/113] eta: 13:16:58 lr: 0.009874 loss: 0.000000 loss_cls: 0.000000 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8500 data_cost: 0.3314 ips: 9.4116 images/s
[07/07 17:05:41] ppdet.engine INFO: Epoch: [9] [  0/113] eta: 13:18:17 lr: 0.009856 loss: 0.000000 loss_cls: 0.000000 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8904 data_cost: 0.3790 ips: 8.9844 images/s
[07/07 17:06:24] ppdet.engine INFO: Epoch: [9] [ 50/113] eta: 13:16:09 lr: 0.009856 loss: 0.000000 loss_cls: 0.000000 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8305 data_cost: 0.3138 ips: 9.6330 images/s
[07/07 17:07:09] ppdet.engine INFO: Epoch: [9] [100/113] eta: 13:14:56 lr: 0.009856 loss: 0.000000 loss_cls: 0.000000 loss_iou: 0.000000 loss_dfl: 0.000000 loss_l1: 0.000000 batch_cost: 0.8496 data_cost: 0.3301 ips: 9.4162 images/s

复现环境 Environment

cudatoolkit 11.2.2 hbe64b41_10 conda-forge
cudnn 8.2.1.32 h86fa8c9_0 conda-forge
paddlepaddle-gpu 2.4.2.post112 pypi_0 pypi

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!

[kaixin-bai]

yolov7可以顺利训练，loss也显示比较正常，请问config文件中我有哪些需要修改？yolov8训练时我将reader文件的share_memory给设置为False，因为会报错。其他的没有什么修改了

[nemonameless]

数据集建议再检查下，loss_dfl loss_iou 在第0iter理论上不会为0
此外训练配置的lr bs不合理，训练自定义数据集最好加上coco权重预训练。
配置文件里加一行 pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov8_l_500e_coco.pdparams
或运行命令后加上 -o pretrain_weights=https://paddledet.bj.bcebos.com/models/yolov8_l_500e_coco.pdparams
同时lr改小1/10
#43

[54wb]

hi,请问你这个问题解决了嘛，我遇到跟你一摸一样的问题，将reader文件的share_memory给设置为False,我加载了预训练模型，我学习率设置是正常的

[54wb]

数据集建议再检查下，loss_dfl loss_iou 在第0iter理论上不会为0 此外训练配置的lr bs不合理，训练自定义数据集最好加上coco权重预训练。 配置文件里加一行 pretrain_weights: https://paddledet.bj.bcebos.com/models/yolov8_l_500e_coco.pdparams 或运行命令后加上 -o pretrain_weights=https://paddledet.bj.bcebos.com/models/yolov8_l_500e_coco.pdparams 同时lr改小1/10 #43

我的情况跟上面的一样，只是中途loss变为了0，加载了预训练模型，bs为16，lr设置的0.00125