loss=nan

[meixiangzhang]

During training, the initial loss was effective, but later the loss became nan. I am using the URPC2019 dataset in VOC format, and the environment is configured according to your instructions. Apart from the dataset format, I have not modified any other configuration files. How can I solve this problem?

-------------------- 2023-08-16 19:47:44,175 - mmdet - INFO - workflow: [('train', 1)], max: 12 epochs 2023-08-16 19:47:44,175 - mmdet - INFO - Checkpoints will be saved to D:\ZMX\Boosting-R-CNN-master\work_dirs\boosting_rcnn_r50_pafpn_1x_voc by HardDiskB ackend. 2023-08-16 19:48:46,645 - mmdet - INFO - Epoch [1][50/951] lr: 4.945e-04, eta: 3:56:35, time: 1.249, data_time: 0.801, memory: 6129, loss_rpn_cls: 0.5699, loss_rpn_bbox: 0.4639, loss_rpn_iou: 0.6679, loss_bbox: 0.3773, loss_cls: 1.0322, acc: 82.9473, loss: 3.1113, grad_norm: 36.3708 2023-08-16 19:49:05,882 - mmdet - INFO - Epoch [1][100/951] lr: 9.940e-04, eta: 2:34:02, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.5379, loss_rpn_bbox: 0.4408, loss_rpn_iou: 0.6699, loss_bbox: 0.4450, loss_cls: 0.5393, acc: 94.2852, loss: 2.6328, grad_norm: 25.0444 2023-08-16 19:49:25,151 - mmdet - INFO - Epoch [1][150/951] lr: 1.494e-03, eta: 2:06:21, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7055, loss_rpn_bbox: 0.4066, loss_rpn_iou: 0.6846, loss_bbox: 0.6769, loss_cls: 0.5754, acc: 93.0977, loss: 3.0490, grad_norm: 18.9466 2023-08-16 19:49:44,496 - mmdet - INFO - Epoch [1][200/951] lr: 1.993e-03, eta: 1:52:25, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6782, loss_rpn_bbox: 0.3794, loss_rpn_iou: 0.6740, loss_bbox: 0.6967, loss_cls: 0.4901, acc: 93.7217, loss: 2.9185, grad_norm: 15.8338 2023-08-16 19:50:03,819 - mmdet - INFO - Epoch [1][250/951] lr: 2.493e-03, eta: 1:43:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6850, loss_rpn_bbox: 0.3607, loss_rpn_iou: 0.6641, loss_bbox: 0.6738, loss_cls: 0.4112, acc: 94.6426, loss: 2.7948, grad_norm: 15.0501 2023-08-16 19:50:23,153 - mmdet - INFO - Epoch [1][300/951] lr: 2.992e-03, eta: 1:38:08, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7968, loss_rpn_bbox: 0.3522, loss_rpn_iou: 0.6599, loss_bbox: 0.6060, loss_cls: 0.3809, acc: 94.4990, loss: 2.7958, grad_norm: 12.3667 2023-08-16 19:50:42,446 - mmdet - INFO - Epoch [1][350/951] lr: 3.492e-03, eta: 1:33:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6889, loss_rpn_bbox: 0.3408, loss_rpn_iou: 0.6569, loss_bbox: 0.6401, loss_cls: 0.3920, acc: 94.2471, loss: 2.7187, grad_norm: 11.3766 2023-08-16 19:51:01,799 - mmdet - INFO - Epoch [1][400/951] lr: 3.991e-03, eta: 1:30:40, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6298, loss_rpn_bbox: 0.3567, loss_rpn_iou: 0.6685, loss_bbox: 0.6711, loss_cls: 0.3629, acc: 94.6396, loss: 2.6889, grad_norm: 10.2118 2023-08-16 19:51:21,118 - mmdet - INFO - Epoch [1][450/951] lr: 4.491e-03, eta: 1:28:04, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7109, loss_rpn_bbox: 0.3316, loss_rpn_iou: 0.6526, loss_bbox: 0.6230, loss_cls: 0.3773, acc: 94.5322, loss: 2.6955, grad_norm: 9.5380 2023-08-16 19:51:39,808 - mmdet - INFO - Epoch [1][500/951] lr: 4.990e-03, eta: 1:25:42, time: 0.374, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 85.8550, loss: nan, grad_norm: nan 2023-08-16 19:51:53,656 - mmdet - INFO - Epoch [1][550/951] lr: 5.000e-03, eta: 1:22:06, time: 0.277, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 12.9807, loss: nan, grad_norm: nan

[mousecpn]

During training, the initial loss was effective, but later the loss became nan. I am using the URPC2019 dataset in VOC format, and the environment is configured according to your instructions. Apart from the dataset format, I have not modified any other configuration files. How can I solve this problem?

-------------------- 2023-08-16 19:47:44,175 - mmdet - INFO - workflow: [('train', 1)], max: 12 epochs 2023-08-16 19:47:44,175 - mmdet - INFO - Checkpoints will be saved to D:\ZMX\Boosting-R-CNN-master\work_dirs\boosting_rcnn_r50_pafpn_1x_voc by HardDiskB ackend. 2023-08-16 19:48:46,645 - mmdet - INFO - Epoch [1][50/951] lr: 4.945e-04, eta: 3:56:35, time: 1.249, data_time: 0.801, memory: 6129, loss_rpn_cls: 0.5699, loss_rpn_bbox: 0.4639, loss_rpn_iou: 0.6679, loss_bbox: 0.3773, loss_cls: 1.0322, acc: 82.9473, loss: 3.1113, grad_norm: 36.3708 2023-08-16 19:49:05,882 - mmdet - INFO - Epoch [1][100/951] lr: 9.940e-04, eta: 2:34:02, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.5379, loss_rpn_bbox: 0.4408, loss_rpn_iou: 0.6699, loss_bbox: 0.4450, loss_cls: 0.5393, acc: 94.2852, loss: 2.6328, grad_norm: 25.0444 2023-08-16 19:49:25,151 - mmdet - INFO - Epoch [1][150/951] lr: 1.494e-03, eta: 2:06:21, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7055, loss_rpn_bbox: 0.4066, loss_rpn_iou: 0.6846, loss_bbox: 0.6769, loss_cls: 0.5754, acc: 93.0977, loss: 3.0490, grad_norm: 18.9466 2023-08-16 19:49:44,496 - mmdet - INFO - Epoch [1][200/951] lr: 1.993e-03, eta: 1:52:25, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6782, loss_rpn_bbox: 0.3794, loss_rpn_iou: 0.6740, loss_bbox: 0.6967, loss_cls: 0.4901, acc: 93.7217, loss: 2.9185, grad_norm: 15.8338 2023-08-16 19:50:03,819 - mmdet - INFO - Epoch [1][250/951] lr: 2.493e-03, eta: 1:43:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6850, loss_rpn_bbox: 0.3607, loss_rpn_iou: 0.6641, loss_bbox: 0.6738, loss_cls: 0.4112, acc: 94.6426, loss: 2.7948, grad_norm: 15.0501 2023-08-16 19:50:23,153 - mmdet - INFO - Epoch [1][300/951] lr: 2.992e-03, eta: 1:38:08, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7968, loss_rpn_bbox: 0.3522, loss_rpn_iou: 0.6599, loss_bbox: 0.6060, loss_cls: 0.3809, acc: 94.4990, loss: 2.7958, grad_norm: 12.3667 2023-08-16 19:50:42,446 - mmdet - INFO - Epoch [1][350/951] lr: 3.492e-03, eta: 1:33:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6889, loss_rpn_bbox: 0.3408, loss_rpn_iou: 0.6569, loss_bbox: 0.6401, loss_cls: 0.3920, acc: 94.2471, loss: 2.7187, grad_norm: 11.3766 2023-08-16 19:51:01,799 - mmdet - INFO - Epoch [1][400/951] lr: 3.991e-03, eta: 1:30:40, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6298, loss_rpn_bbox: 0.3567, loss_rpn_iou: 0.6685, loss_bbox: 0.6711, loss_cls: 0.3629, acc: 94.6396, loss: 2.6889, grad_norm: 10.2118 2023-08-16 19:51:21,118 - mmdet - INFO - Epoch [1][450/951] lr: 4.491e-03, eta: 1:28:04, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7109, loss_rpn_bbox: 0.3316, loss_rpn_iou: 0.6526, loss_bbox: 0.6230, loss_cls: 0.3773, acc: 94.5322, loss: 2.6955, grad_norm: 9.5380 2023-08-16 19:51:39,808 - mmdet - INFO - Epoch [1][500/951] lr: 4.990e-03, eta: 1:25:42, time: 0.374, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 85.8550, loss: nan, grad_norm: nan 2023-08-16 19:51:53,656 - mmdet - INFO - Epoch [1][550/951] lr: 5.000e-03, eta: 1:22:06, time: 0.277, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 12.9807, loss: nan, grad_norm: nan

That's super weird. But I did realize that the VOC format URPC often caused a problem. My advices are:
(1) Check whether the problem comes from the dataset. Locate the image which causes nan. Checking whether the images that cause the problem are the same image.
(2) Try a grad_clip.
(3) turn this VOC format to COCO format.

[meixiangzhang]

During training, the initial loss was effective, but later the loss became nan. I am using the URPC2019 dataset in VOC format, and the environment is configured according to your instructions. Apart from the dataset format, I have not modified any other configuration files. How can I solve this problem?
-------------------- 2023-08-16 19:47:44,175 - mmdet - INFO - workflow: [('train', 1)], max: 12 epochs 2023-08-16 19:47:44,175 - mmdet - INFO - Checkpoints will be saved to D:\ZMX\Boosting-R-CNN-master\work_dirs\boosting_rcnn_r50_pafpn_1x_voc by HardDiskB ackend. 2023-08-16 19:48:46,645 - mmdet - INFO - Epoch [1][50/951] lr: 4.945e-04, eta: 3:56:35, time: 1.249, data_time: 0.801, memory: 6129, loss_rpn_cls: 0.5699, loss_rpn_bbox: 0.4639, loss_rpn_iou: 0.6679, loss_bbox: 0.3773, loss_cls: 1.0322, acc: 82.9473, loss: 3.1113, grad_norm: 36.3708 2023-08-16 19:49:05,882 - mmdet - INFO - Epoch [1][100/951] lr: 9.940e-04, eta: 2:34:02, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.5379, loss_rpn_bbox: 0.4408, loss_rpn_iou: 0.6699, loss_bbox: 0.4450, loss_cls: 0.5393, acc: 94.2852, loss: 2.6328, grad_norm: 25.0444 2023-08-16 19:49:25,151 - mmdet - INFO - Epoch [1][150/951] lr: 1.494e-03, eta: 2:06:21, time: 0.385, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7055, loss_rpn_bbox: 0.4066, loss_rpn_iou: 0.6846, loss_bbox: 0.6769, loss_cls: 0.5754, acc: 93.0977, loss: 3.0490, grad_norm: 18.9466 2023-08-16 19:49:44,496 - mmdet - INFO - Epoch [1][200/951] lr: 1.993e-03, eta: 1:52:25, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6782, loss_rpn_bbox: 0.3794, loss_rpn_iou: 0.6740, loss_bbox: 0.6967, loss_cls: 0.4901, acc: 93.7217, loss: 2.9185, grad_norm: 15.8338 2023-08-16 19:50:03,819 - mmdet - INFO - Epoch [1][250/951] lr: 2.493e-03, eta: 1:43:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6850, loss_rpn_bbox: 0.3607, loss_rpn_iou: 0.6641, loss_bbox: 0.6738, loss_cls: 0.4112, acc: 94.6426, loss: 2.7948, grad_norm: 15.0501 2023-08-16 19:50:23,153 - mmdet - INFO - Epoch [1][300/951] lr: 2.992e-03, eta: 1:38:08, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7968, loss_rpn_bbox: 0.3522, loss_rpn_iou: 0.6599, loss_bbox: 0.6060, loss_cls: 0.3809, acc: 94.4990, loss: 2.7958, grad_norm: 12.3667 2023-08-16 19:50:42,446 - mmdet - INFO - Epoch [1][350/951] lr: 3.492e-03, eta: 1:33:54, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6889, loss_rpn_bbox: 0.3408, loss_rpn_iou: 0.6569, loss_bbox: 0.6401, loss_cls: 0.3920, acc: 94.2471, loss: 2.7187, grad_norm: 11.3766 2023-08-16 19:51:01,799 - mmdet - INFO - Epoch [1][400/951] lr: 3.991e-03, eta: 1:30:40, time: 0.387, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.6298, loss_rpn_bbox: 0.3567, loss_rpn_iou: 0.6685, loss_bbox: 0.6711, loss_cls: 0.3629, acc: 94.6396, loss: 2.6889, grad_norm: 10.2118 2023-08-16 19:51:21,118 - mmdet - INFO - Epoch [1][450/951] lr: 4.491e-03, eta: 1:28:04, time: 0.386, data_time: 0.003, memory: 6129, loss_rpn_cls: 0.7109, loss_rpn_bbox: 0.3316, loss_rpn_iou: 0.6526, loss_bbox: 0.6230, loss_cls: 0.3773, acc: 94.5322, loss: 2.6955, grad_norm: 9.5380 2023-08-16 19:51:39,808 - mmdet - INFO - Epoch [1][500/951] lr: 4.990e-03, eta: 1:25:42, time: 0.374, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 85.8550, loss: nan, grad_norm: nan 2023-08-16 19:51:53,656 - mmdet - INFO - Epoch [1][550/951] lr: 5.000e-03, eta: 1:22:06, time: 0.277, data_time: 0.003, memory: 6129, loss_rpn_cls: nan, loss_rpn_bbox: nan, loss_rpn_iou: nan, loss_bbox: nan, loss_cls: nan, acc: 12.9807, loss: nan, grad_norm: nan

That's super weird. But I did realize that the VOC format URPC often caused a problem. My advices are: (1) Check whether the problem comes from the dataset. Locate the image which causes nan. Checking whether the images that cause the problem are the same image. (2) Try a grad_clip. (3) turn this VOC format to COCO format.

Thank you for your suggestions. I replaced the dataset in COCO format and she is now working properly.