Oneflow-Inc / oneflow

OneFlow is a deep learning framework designed to be user-friendly, scalable and efficient.
http://www.oneflow.org
Apache License 2.0
5.86k stars 661 forks source link

复现 TDAN 时,可形变卷积相关的 loss 不收敛 #9872

Open wwhio opened 1 year ago

wwhio commented 1 year ago

Description

论文:TDAN: Temporally-Deformable Alignment Network for Video Super-Resolution, CVPR 2020 问题:TDAN 的 loss 由两部分组成,和可形变卷积相关的部分不能正常收敛,但在 PyTorch 实现中可以。

wwhio commented 1 year ago

训练日志:

22-12-25 02:16:57.041 - INFO: Start training.
22-12-25 02:17:35.141 - INFO: Iter: 100, LR: 1.000e-04, pixel: 36512.547, align: 56599.852, 
22-12-25 02:18:12.655 - INFO: Iter: 200, LR: 1.000e-04, pixel: 224289.969, align: 14450721.000, 
22-12-25 02:18:49.696 - INFO: Iter: 300, LR: 1.000e-04, pixel: 64504.664, align: 20301796.000, 
22-12-25 02:19:26.969 - INFO: Iter: 400, LR: 1.000e-04, pixel: 11645.910, align: 3653067.000, 
22-12-25 02:20:04.346 - INFO: Iter: 500, LR: 1.000e-04, pixel: 33667.680, align: 7805409.000, 
…………………………………………………………………………
22-12-25 12:40:56.857 - INFO: Iter: 99500, LR: 5.000e-05, pixel: 3847.587, align: 15787.362, 
22-12-25 12:41:33.367 - INFO: Iter: 99600, LR: 5.000e-05, pixel: 4950.891, align: 12813.590, 
22-12-25 12:42:09.883 - INFO: Iter: 99700, LR: 5.000e-05, pixel: 8242.460, align: 10576.354, 
22-12-25 12:42:46.160 - INFO: Iter: 99800, LR: 5.000e-05, pixel: 5750.407, align: 13752.346, 
22-12-25 12:43:22.516 - INFO: Iter: 99900, LR: 5.000e-05, pixel: 1096.244, align: 12073.321, 
22-12-25 12:43:59.005 - INFO: Iter: 100000, LR: 5.000e-05, pixel: 2564.331, align: 17589.385
wwhio commented 1 year ago

Pytorch 实现的日志:

22-12-24 22:20:15.461 - INFO: Start training.
22-12-24 22:21:03.231 - INFO: Iter: 100, LR: 1.000e-04, pixel: 23383.805, align: 1757.192, 
22-12-24 22:21:47.388 - INFO: Iter: 200, LR: 1.000e-04, pixel: 16763.250, align: 1979.344, 
22-12-24 22:22:31.734 - INFO: Iter: 300, LR: 1.000e-04, pixel: 6542.937, align: 1074.003, 
22-12-24 22:23:16.155 - INFO: Iter: 400, LR: 1.000e-04, pixel: 1002.172, align: 279.012, 
22-12-24 22:24:00.620 - INFO: Iter: 500, LR: 1.000e-04, pixel: 4013.701, align: 558.434, 
...................................
22-12-25 10:50:32.517 - INFO: Iter: 99500, LR: 5.000e-05, pixel: 2584.350, align: 58.607, 
22-12-25 10:51:17.293 - INFO: Iter: 99600, LR: 5.000e-05, pixel: 3799.376, align: 39.234, 
22-12-25 10:52:02.064 - INFO: Iter: 99700, LR: 5.000e-05, pixel: 6548.685, align: 64.043, 
22-12-25 10:52:46.820 - INFO: Iter: 99800, LR: 5.000e-05, pixel: 4459.707, align: 53.848, 
22-12-25 10:53:31.587 - INFO: Iter: 99900, LR: 5.000e-05, pixel: 1096.945, align: 76.437, 
22-12-25 10:54:16.362 - INFO: Iter: 100000, LR: 5.000e-05, pixel: 2012.993, align: 47.685,