我在coco数据集上用yoloxm蒸馏yoloxs，损失曲线也正常，但是第一个epoch，精度为0，我已经按照reademe一步一步改了，之前是在自己数据集上精度为0，我想排查下到底哪里有问题，就在coco上进行了实验，发现在coco上也是同样问题。请问我这可能是哪里的问题那，已经困扰我很多天了。。。。

Hi, I have runnd faster_rcnn_res50 for detection and mask_rcnn_res50 for segmentation, all configs are same with paper, I got bbox_mAP 40.8 and segm_mAP 36.6, witch is a gap with paper bbox_mAP 42.0 and segm_mAP 37.8. I want to know your benchmark running once or several times pick out best results? Thank your~

thanks for sharing your excellent job! can you update your training logs? when I have runned configs/distillers/fgd/fgd_retina_rx101_64x4d_distill_retina_r50_fpn_2x_coco.py , I only got 30.4 mAP Epoch(val) [24][625] bbox_mAP: 0.3040, bbox_mAP_50: 0.4980, bbox_mAP_75: 0.3140, bbox_mAP_s: 0.1640, bbox_mAP_m: 0.3300, bbox_mAP_l: 0.4160, bbox_mAP_copypaste: 0.304 0.498 0.314 0.164 0.330 0.416

I don't konw why thank your~

Hi,Thanks your great work I use FDG in solov2, the teacher model is solov2-resNet101 and the student model is solov2-mobileNetV2 the train log is below: 20220520_164942.log and the config file is below: fgd_solov2_r101_c8_ms_distill_solov2_mobileNetV2_slim_c8_txt.txt I compared the mAP between solov2-mobileNetV2 distiller model and no distiller model, but the result is not ideal. the mAP of solov2-mobileNetV2 model (0.465) ,which is higher than the distiller model (0.426) why the accuary of the distiller model reduced ?

Hello, Thank you for your insight work! I have a question when I add a DCN structure for the teacher or student, e.g, faster_rcnn_r50_fpn_dconv_c3-c5, cascade_mask_rcnn_x101_32x4d_fpn_dconv_c3-c5, loss_fgd_fpn and total loss will become too large to train? I am wondering the reason behind it.

Here is the log output:

2022-10-10 15:07:23,967 - mmdet - INFO - Epoch [1][50/16178]	lr: 9.890e-04, eta: 3 days, 7:50:51, time: 1.481, data_time: 0.072, memory: 11330, loss_cls: 0.6365, loss_bbox: 0.6870, loss_dfl: 0.3843, loss_fgd_fpn_4: 0.0002, loss_fgd_fpn_3: 0.0181, loss_fgd_fpn_2: 0.1217, loss_fgd_fpn_1: 1.3931, loss_fgd_fpn_0: 6.3135, loss: 9.5544, grad_norm: 68.6224
2022-10-10 15:08:34,256 - mmdet - INFO - Epoch [1][100/16178]	lr: 1.988e-03, eta: 3 days, 5:47:53, time: 1.406, data_time: 0.010, memory: 11331, loss_cls: 0.5288, loss_bbox: 0.5811, loss_dfl: 0.3424, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0081, loss_fgd_fpn_2: 0.0561, loss_fgd_fpn_1: 0.9112, loss_fgd_fpn_0: 4.2769, loss: 6.7047, grad_norm: 23.8445
2022-10-10 15:09:45,544 - mmdet - INFO - Epoch [1][150/16178]	lr: 2.987e-03, eta: 3 days, 5:27:40, time: 1.426, data_time: 0.010, memory: 11332, loss_cls: 0.5084, loss_bbox: 0.5301, loss_dfl: 0.3167, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0086, loss_fgd_fpn_2: 0.0718, loss_fgd_fpn_1: 1.0272, loss_fgd_fpn_0: 4.8596, loss: 7.3225, grad_norm: 37.7240
2022-10-10 15:10:57,655 - mmdet - INFO - Epoch [1][200/16178]	lr: 3.986e-03, eta: 3 days, 5:30:15, time: 1.442, data_time: 0.010, memory: 11332, loss_cls: 0.5128, loss_bbox: 0.5244, loss_dfl: 0.3146, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0079, loss_fgd_fpn_2: 0.0572, loss_fgd_fpn_1: 1.0461, loss_fgd_fpn_0: 4.8641, loss: 7.3272, grad_norm: 30.0396
2022-10-10 15:12:08,429 - mmdet - INFO - Epoch [1][250/16178]	lr: 4.985e-03, eta: 3 days, 5:14:02, time: 1.415, data_time: 0.010, memory: 11332, loss_cls: 0.5229, loss_bbox: 0.5141, loss_dfl: 0.3086, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0068, loss_fgd_fpn_2: 0.0506, loss_fgd_fpn_1: 0.8142, loss_fgd_fpn_0: 3.9057, loss: 6.1230, grad_norm: 12.6570
2022-10-10 15:13:19,710 - mmdet - INFO - Epoch [1][300/16178]	lr: 5.984e-03, eta: 3 days, 5:08:18, time: 1.426, data_time: 0.010, memory: 11332, loss_cls: 0.4991, loss_bbox: 0.5199, loss_dfl: 0.3110, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0072, loss_fgd_fpn_2: 0.0523, loss_fgd_fpn_1: 0.8164, loss_fgd_fpn_0: 3.7634, loss: 5.9694, grad_norm: 11.5621
2022-10-10 15:14:30,729 - mmdet - INFO - Epoch [1][350/16178]	lr: 6.983e-03, eta: 3 days, 5:01:27, time: 1.420, data_time: 0.010, memory: 11332, loss_cls: 0.5207, loss_bbox: 0.5134, loss_dfl: 0.3113, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0073, loss_fgd_fpn_2: 0.0541, loss_fgd_fpn_1: 0.8442, loss_fgd_fpn_0: 4.1035, loss: 6.3546, grad_norm: 13.2545
2022-10-10 15:15:41,433 - mmdet - INFO - Epoch [1][400/16178]	lr: 7.982e-03, eta: 3 days, 4:53:28, time: 1.414, data_time: 0.010, memory: 11332, loss_cls: 0.4979, loss_bbox: 0.5297, loss_dfl: 0.3136, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0070, loss_fgd_fpn_2: 0.0515, loss_fgd_fpn_1: 0.7945, loss_fgd_fpn_0: 3.7389, loss: 5.9332, grad_norm: 8.9079
2022-10-10 15:16:52,334 - mmdet - INFO - Epoch [1][450/16178]	lr: 8.981e-03, eta: 3 days, 4:48:25, time: 1.418, data_time: 0.010, memory: 11332, loss_cls: 0.5142, loss_bbox: 0.5102, loss_dfl: 0.3119, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0068, loss_fgd_fpn_2: 0.0489, loss_fgd_fpn_1: 0.7655, loss_fgd_fpn_0: 3.7806, loss: 5.9382, grad_norm: 9.4913
2022-10-10 15:18:03,443 - mmdet - INFO - Epoch [1][500/16178]	lr: 9.980e-03, eta: 3 days, 4:45:28, time: 1.422, data_time: 0.010, memory: 11332, loss_cls: 0.5051, loss_bbox: 0.5061, loss_dfl: 0.3060, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0071, loss_fgd_fpn_2: 0.0511, loss_fgd_fpn_1: 0.7981, loss_fgd_fpn_0: 3.7618, loss: 5.9355, grad_norm: 9.1807
2022-10-10 15:19:14,859 - mmdet - INFO - Epoch [1][550/16178]	lr: 1.000e-02, eta: 3 days, 4:44:39, time: 1.428, data_time: 0.010, memory: 11332, loss_cls: 0.5116, loss_bbox: 0.4758, loss_dfl: 0.2929, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0071, loss_fgd_fpn_2: 0.0524, loss_fgd_fpn_1: 0.7696, loss_fgd_fpn_0: 3.6697, loss: 5.7791, grad_norm: 10.2593
2022-10-10 15:20:26,284 - mmdet - INFO - Epoch [1][600/16178]	lr: 1.000e-02, eta: 3 days, 4:43:49, time: 1.428, data_time: 0.010, memory: 11332, loss_cls: 0.5084, loss_bbox: 0.5035, loss_dfl: 0.3048, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0074, loss_fgd_fpn_2: 0.0538, loss_fgd_fpn_1: 0.8274, loss_fgd_fpn_0: 3.9499, loss: 6.1554, grad_norm: 11.5289
2022-10-10 15:21:38,146 - mmdet - INFO - Epoch [1][650/16178]	lr: 1.000e-02, eta: 3 days, 4:45:06, time: 1.437, data_time: 0.010, memory: 11332, loss_cls: 0.5201, loss_bbox: 0.5220, loss_dfl: 0.3139, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0105, loss_fgd_fpn_2: 0.0730, loss_fgd_fpn_1: 0.8779, loss_fgd_fpn_0: 4.0663, loss: 6.3838, grad_norm: 12.7775
2022-10-10 15:22:49,456 - mmdet - INFO - Epoch [1][700/16178]	lr: 1.000e-02, eta: 3 days, 4:43:29, time: 1.426, data_time: 0.010, memory: 11332, loss_cls: 0.5132, loss_bbox: 0.5461, loss_dfl: 0.3226, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0079, loss_fgd_fpn_2: 0.0564, loss_fgd_fpn_1: 0.8152, loss_fgd_fpn_0: 3.8834, loss: 6.1450, grad_norm: 11.4293
2022-10-10 15:24:00,110 - mmdet - INFO - Epoch [1][750/16178]	lr: 1.000e-02, eta: 3 days, 4:39:06, time: 1.413, data_time: 0.010, memory: 11332, loss_cls: 0.5316, loss_bbox: 0.5172, loss_dfl: 0.3141, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0076, loss_fgd_fpn_2: 0.0548, loss_fgd_fpn_1: 0.8379, loss_fgd_fpn_0: 3.9216, loss: 6.1849, grad_norm: 8.1157
2022-10-10 15:25:11,329 - mmdet - INFO - Epoch [1][800/16178]	lr: 1.000e-02, eta: 3 days, 4:37:24, time: 1.424, data_time: 0.010, memory: 11332, loss_cls: 0.4892, loss_bbox: 0.4858, loss_dfl: 0.2968, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0069, loss_fgd_fpn_2: 0.0519, loss_fgd_fpn_1: 0.7800, loss_fgd_fpn_0: 3.5984, loss: 5.7091, grad_norm: 7.2720
2022-10-10 15:26:21,786 - mmdet - INFO - Epoch [1][850/16178]	lr: 1.000e-02, eta: 3 days, 4:32:53, time: 1.409, data_time: 0.010, memory: 11332, loss_cls: 0.5169, loss_bbox: 0.5149, loss_dfl: 0.3076, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0076, loss_fgd_fpn_2: 0.0551, loss_fgd_fpn_1: 0.8749, loss_fgd_fpn_0: 3.9996, loss: 6.2768, grad_norm: 11.8264
2022-10-10 15:27:32,728 - mmdet - INFO - Epoch [1][900/16178]	lr: 1.000e-02, eta: 3 days, 4:30:27, time: 1.419, data_time: 0.010, memory: 11332, loss_cls: 0.5167, loss_bbox: 0.4732, loss_dfl: 0.2936, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0070, loss_fgd_fpn_2: 0.0519, loss_fgd_fpn_1: 0.7775, loss_fgd_fpn_0: 3.6971, loss: 5.8172, grad_norm: 8.8075
2022-10-10 15:28:44,486 - mmdet - INFO - Epoch [1][950/16178]	lr: 1.000e-02, eta: 3 days, 4:30:56, time: 1.435, data_time: 0.010, memory: 11332, loss_cls: 0.4985, loss_bbox: 0.4969, loss_dfl: 0.3020, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0066, loss_fgd_fpn_2: 0.0498, loss_fgd_fpn_1: 0.7534, loss_fgd_fpn_0: 3.6059, loss: 5.7131, grad_norm: 7.6654
2022-10-10 15:29:55,254 - mmdet - INFO - Epoch [1][1000/16178]	lr: 1.000e-02, eta: 3 days, 4:28:03, time: 1.415, data_time: 0.010, memory: 11332, loss_cls: 0.4810, loss_bbox: 0.5045, loss_dfl: 0.3053, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0063, loss_fgd_fpn_2: 0.0447, loss_fgd_fpn_1: 0.6467, loss_fgd_fpn_0: 3.1656, loss: 5.1542, grad_norm: 6.4960
2022-10-10 15:31:05,415 - mmdet - INFO - Epoch [1][1050/16178]	lr: 1.000e-02, eta: 3 days, 4:23:28, time: 1.403, data_time: 0.010, memory: 11332, loss_cls: 0.5539, loss_bbox: 0.5281, loss_dfl: 0.3202, loss_fgd_fpn_4: 0.0002, loss_fgd_fpn_3: 0.0451, loss_fgd_fpn_2: 0.2862, loss_fgd_fpn_1: 10.4438, loss_fgd_fpn_0: 30.4854, loss: 42.6630, grad_norm: 3034.1699
2022-10-10 15:32:16,896 - mmdet - INFO - Epoch [1][1100/16178]	lr: 1.000e-02, eta: 3 days, 4:23:04, time: 1.430, data_time: 0.010, memory: 11332, loss_cls: 0.5293, loss_bbox: 0.5076, loss_dfl: 0.3075, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0076, loss_fgd_fpn_2: 0.0534, loss_fgd_fpn_1: 0.8656, loss_fgd_fpn_0: 3.8698, loss: 6.1410, grad_norm: 15.3256
2022-10-10 15:33:27,140 - mmdet - INFO - Epoch [1][1150/16178]	lr: 1.000e-02, eta: 3 days, 4:19:08, time: 1.405, data_time: 0.010, memory: 11332, loss_cls: 0.5029, loss_bbox: 0.4510, loss_dfl: 0.2841, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0064, loss_fgd_fpn_2: 0.0465, loss_fgd_fpn_1: 0.6350, loss_fgd_fpn_0: 3.1598, loss: 5.0856, grad_norm: 6.5924
2022-10-10 15:34:37,744 - mmdet - INFO - Epoch [1][1200/16178]	lr: 1.000e-02, eta: 3 days, 4:16:23, time: 1.412, data_time: 0.010, memory: 11332, loss_cls: 0.4919, loss_bbox: 0.4949, loss_dfl: 0.3020, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0075, loss_fgd_fpn_2: 0.0529, loss_fgd_fpn_1: 0.7247, loss_fgd_fpn_0: 3.5468, loss: 5.6208, grad_norm: 6.6782
2022-10-10 15:35:48,395 - mmdet - INFO - Epoch [1][1250/16178]	lr: 1.000e-02, eta: 3 days, 4:13:54, time: 1.413, data_time: 0.010, memory: 11332, loss_cls: 0.5052, loss_bbox: 0.4780, loss_dfl: 0.2934, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0071, loss_fgd_fpn_2: 0.0507, loss_fgd_fpn_1: 0.6617, loss_fgd_fpn_0: 3.1991, loss: 5.1952, grad_norm: 7.1381
2022-10-10 15:36:59,383 - mmdet - INFO - Epoch [1][1300/16178]	lr: 1.000e-02, eta: 3 days, 4:12:20, time: 1.420, data_time: 0.010, memory: 11332, loss_cls: 0.4702, loss_bbox: 0.4689, loss_dfl: 0.2900, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0067, loss_fgd_fpn_2: 0.0483, loss_fgd_fpn_1: 0.6287, loss_fgd_fpn_0: 3.1048, loss: 5.0178, grad_norm: 6.6730
2022-10-10 15:38:09,195 - mmdet - INFO - Epoch [1][1350/16178]	lr: 1.000e-02, eta: 3 days, 4:08:00, time: 1.396, data_time: 0.010, memory: 11332, loss_cls: 0.5176, loss_bbox: 0.4569, loss_dfl: 0.2849, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0062, loss_fgd_fpn_2: 0.0443, loss_fgd_fpn_1: 0.5653, loss_fgd_fpn_0: 2.8550, loss: 4.7303, grad_norm: 6.2432
2022-10-10 15:39:19,317 - mmdet - INFO - Epoch [1][1400/16178]	lr: 1.000e-02, eta: 3 days, 4:04:37, time: 1.402, data_time: 0.010, memory: 11332, loss_cls: 0.4808, loss_bbox: 0.4700, loss_dfl: 0.2906, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0062, loss_fgd_fpn_2: 0.0451, loss_fgd_fpn_1: 0.5628, loss_fgd_fpn_0: 2.8688, loss: 4.7243, grad_norm: 5.8212
2022-10-10 15:40:29,784 - mmdet - INFO - Epoch [1][1450/16178]	lr: 1.000e-02, eta: 3 days, 4:02:08, time: 1.409, data_time: 0.010, memory: 11332, loss_cls: 0.4922, loss_bbox: 0.4429, loss_dfl: 0.2802, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0061, loss_fgd_fpn_2: 0.0443, loss_fgd_fpn_1: 0.5911, loss_fgd_fpn_0: 2.9405, loss: 4.7974, grad_norm: 7.1132
2022-10-10 15:41:40,482 - mmdet - INFO - Epoch [1][1500/16178]	lr: 1.000e-02, eta: 3 days, 4:00:15, time: 1.414, data_time: 0.010, memory: 11332, loss_cls: 0.4567, loss_bbox: 0.4931, loss_dfl: 0.3016, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0066, loss_fgd_fpn_2: 0.0472, loss_fgd_fpn_1: 0.6008, loss_fgd_fpn_0: 2.9786, loss: 4.8846, grad_norm: 6.0787
2022-10-10 15:42:51,679 - mmdet - INFO - Epoch [1][1550/16178]	lr: 1.000e-02, eta: 3 days, 3:59:26, time: 1.424, data_time: 0.010, memory: 11332, loss_cls: 0.4667, loss_bbox: 0.4836, loss_dfl: 0.2966, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0064, loss_fgd_fpn_2: 0.0461, loss_fgd_fpn_1: 0.6019, loss_fgd_fpn_0: 2.9062, loss: 4.8075, grad_norm: 7.3193
2022-10-10 15:44:02,559 - mmdet - INFO - Epoch [1][1600/16178]	lr: 1.000e-02, eta: 3 days, 3:57:57, time: 1.418, data_time: 0.010, memory: 11332, loss_cls: 0.5035, loss_bbox: 0.4874, loss_dfl: 0.2984, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0208, loss_fgd_fpn_2: 0.1415, loss_fgd_fpn_1: 12.0289, loss_fgd_fpn_0: 22.1169, loss: 35.5976, grad_norm: 9077.5504
2022-10-10 15:45:13,235 - mmdet - INFO - Epoch [1][1650/16178]	lr: 1.000e-02, eta: 3 days, 3:56:06, time: 1.414, data_time: 0.010, memory: 11332, loss_cls: 0.4893, loss_bbox: 0.4638, loss_dfl: 0.2870, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0065, loss_fgd_fpn_2: 0.0473, loss_fgd_fpn_1: 0.6394, loss_fgd_fpn_0: 3.2284, loss: 5.1618, grad_norm: 46.0868
2022-10-10 15:46:23,125 - mmdet - INFO - Epoch [1][1700/16178]	lr: 1.000e-02, eta: 3 days, 3:52:48, time: 1.398, data_time: 0.010, memory: 11332, loss_cls: 0.4880, loss_bbox: 0.4438, loss_dfl: 0.2836, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0061, loss_fgd_fpn_2: 0.0430, loss_fgd_fpn_1: 0.5120, loss_fgd_fpn_0: 2.5668, loss: 4.3435, grad_norm: 5.6961
2022-10-10 15:47:34,943 - mmdet - INFO - Epoch [1][1750/16178]	lr: 1.000e-02, eta: 3 days, 3:53:10, time: 1.436, data_time: 0.010, memory: 11332, loss_cls: 0.4759, loss_bbox: 0.4977, loss_dfl: 0.3028, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0070, loss_fgd_fpn_2: 0.0493, loss_fgd_fpn_1: 0.6108, loss_fgd_fpn_0: 3.0078, loss: 4.9514, grad_norm: 6.7499
2022-10-10 15:48:45,031 - mmdet - INFO - Epoch [1][1800/16178]	lr: 1.000e-02, eta: 3 days, 3:50:21, time: 1.402, data_time: 0.010, memory: 11332, loss_cls: 0.5020, loss_bbox: 0.4571, loss_dfl: 0.2838, loss_fgd_fpn_4: 0.0001, loss_fgd_fpn_3: 0.0067, loss_fgd_fpn_2: 0.0485, loss_fgd_fpn_1: 0.5821, loss_fgd_fpn_0: 2.9682, loss: 4.8486, grad_norm: 8.6155
2022-10-10 15:49:55,865 - mmdet - INFO - Epoch [1][1850/16178]	lr: 1.000e-02, eta: 3 days, 3:48:56, time: 1.417, data_time: 0.010, memory: 11332, loss_cls: 0.5524, loss_bbox: 0.5681, loss_dfl: 0.3350, loss_fgd_fpn_4: 0.0234, loss_fgd_fpn_3: 20.7488, loss_fgd_fpn_2: 135.5364, loss_fgd_fpn_1: 4514.5350, loss_fgd_fpn_0: 5980.8744, loss: 10653.1733, grad_norm: 693299.5836

Dear author, I configured my training model on MMDetection, and the bbox_mAP_50 of the validation set was preferably 0.883, but the test result was 0.My detailed steps are as follows: Add and Replace the codes 1.Add folders a_my_configs/ to the configs/ in mmdetectin's codes. 2.Add my_faster_rcnn_r50_fpn_2x_coco.py and my_faster_rcnn_r101_fpn_2x_coco.py to the configs/a_my_configs/. 3.In my_faster_rcnn_r50_fpn_2x_coco.py, base as same as configs/faster_rcnn/faster_rcnn_r50_fpn_2x_coco.py, add data=dict( samples_per_gpu=1, # batch size workers_per_gpu=1, # num_workers train=dict(path1),val=dict(path2),test=dict(path3)) 4.Modify the
student_cfg = 'configs/a_my_configs/my_faster_rcnn_r50_fpn_2x_coco.py' teacher_cfg = 'configs/a_my_configs/my_faster_rcnn_r101_fpn_2x_coco.py' in configs/distillers/fgd/fgd_faster_rcnn_r101_fpn_2x_distill_faster_rcnn_r50_fpn_2x_coco.py Train python tools/train.py configs/distillers/fgd/fgd_faster_rcnn_r101_fpn_2x_distill_faster_rcnn_r50_fpn_2x_coco.py The training process went smoothly and the best bbox_mAP_50=0.883 Test python tools/test.py configs/a_my_configs/my_faster_rcnn_r50_fpn_2x_coco.py work_dirs/fgd_faster_rcnn_r101_fpn_2x_distill_faster_rcnn_r50_fpn_2x_coco/latest.pth --eval bbox

But the test results(bbox_mAP,bbox_mAP_50,bbox_mAP_l……) were 0, I guess the problem appears on the configuration file, but do not know how to solve, looking forward to your reply, I can provide my complete code

作者你好，我使用FCOS框架进行蒸馏，发现训练过程中，随着epoch增加，测试结果在不断降低，不知道是不是配置的问题，能否帮帮忙看下？谢谢！

环境配置： pytorch=1.7.1，mmdetection=2.11.0，mmcv=1.2.4，cuda=10.2

训练命令行： bash tools/dist_train.sh configs/distillers/fgd/fgd_fcos_r101_distill_fcos_r50_2x_coco.py 2

训练log（只包含前5个epoch）： fgd_fcos.log

@yzd-v Hi, thanks for your simple and powerfull KD method. I have one question, why use GCNet rather commonly used Non-local module,
Have you test these two OPs since GCNet is little better than Non-local.(Maybe some experiments about these two modules can be added into your paper, since this is one question reviewers intuitively will ask)

Thank you for open this repo. $A^S, A^C$ denote the spatial and channel attention mask of the teacher detector In formulate (9) , However $A^S_t , A^C_t $ is contrain in formulate (10). Why not fuse $A^S_t , A^C_t $ in formulate (9)?

是框架的问题吗？ missing keys in source state_dict: backbone.conv1.weight, backbone.bn1.weight, backbone.bn1.bias, backbone.bn1.running_mean, backbone.bn1.running_var, backbone.layer1.0.conv1.weight, backbone.layer1.0.bn1.weight, backbone.layer1.0.bn1.bias, backbone.layer1.0.bn1.running_mean,............

load checkpoint from local path: ../tools/Retina_ResX101-50_40.7_final.pth The model and loaded state dict do not match exactly

size mismatch for bbox_head.retina_cls.weight: copying a param with shape torch.Size([720, 256, 3, 3]) from checkpoint, the shape in current model is torch.Size([36, 256, 3, 3]). size mismatch for bbox_head.retina_cls.bias: copying a param with shape torch.Size([720]) from checkpoint, the shape in current model is torch.Size([36]).

Hey,bro. I wonder using two stage method like Faster-RCNN to distill one stage method like Yolo, only using neck feature to distill, will it work？what's your opinion? Thanks.

当前有两个YOLOX-l模型（分别检测抽烟（YOLOX-l-cigar）和玩手机（YOLOX-l-cellphone）），利用FGD将YOLOX-l-cigar和YOLOX-l-cellphone模型对YOLOX-m学生进行训练，训练后的模型既能检测抽烟 也能检测玩手机，有相关的示例代码或者指导吗？多谢

Line 86 in the forward function in mmdet/distillation/losses/fgd.py There are two for loops here, in my test, these two for loops slow down the training seriously, is there any solution?

Traceback (most recent call last): File "tools/train.py", line 15, in from mmdet.apis import set_random_seed, train_detector File "/home/renyu/anzhuang/anaconda3/envs/FGD/lib/python3.7/site-packages/mmdet/apis/init.py", line 2, in from .inference import (async_inference_detector, inference_detector, File "/home/renyu/anzhuang/anaconda3/envs/FGD/lib/python3.7/site-packages/mmdet/apis/inference.py", line 7, in from mmcv.ops import RoIPool File "/home/renyu/anzhuang/anaconda3/envs/FGD/lib/python3.7/site-packages/mmcv/ops/init.py", line 10, in from .corner_pool import CornerPool File "/home/renyu/anzhuang/anaconda3/envs/FGD/lib/python3.7/site-packages/mmcv/ops/corner_pool.py", line 11, in 'right_pool_forward', 'right_pool_backward' File "/home/renyu/anzhuang/anaconda3/envs/FGD/lib/python3.7/site-packages/mmcv/utils/ext_loader.py", line 15, in load_ext assert hasattr(ext, fun), f'{fun} miss in module {name}' AssertionError: top_pool_forward miss in module _ext

how about the results of some lightweight backones such as mobbilenetv3。I used resnet50 GFL as teacher and mobilenetv3 GFL as student but got very small improvement in my own datasets. Any guideline to these model?