Poor performance when do testing(i3d_nonlocal_net_400k_8gpu)

[LossNAN]

When training, it seems that there are no problems,
| Train ETA: 0:00:00 LR: 0.00010000 Iters [400000/400000] [109.10ep] Time 0.991 Loss 1.0673 top1 25.000 top5 12.500
* Finished #iters [400000|500500]: top1: 40.155 top5: 19.018
But after test, i got a very poor result.
[INFO: test_net_video.py: 140]: 9258/9263 iter (19754/19761 videos): Time: 0.679 (ETA: 0:00:04). ID: [3323, 3325, 3323, 3323, 3323, 3324, 3325, 3323]
[INFO: test_net_video.py: 140]: 9259/9263 iter (19756/19761 videos): Time: 1.408 (ETA: 0:00:03). ID: [3325, 3327, 3326, 3327, 3326, 3325, 33, 3326]
[INFO: test_net_video.py: 140]: 9260/9263 iter (19758/19761 videos): Time: 1.354 (ETA: 0:00:02). ID: [3327, 3328, 3329, 332, 3327, 3329, 3328, 3328]
[INFO: test_net_video.py: 140]: 9261/9263 iter (19760/19761 videos): Time: 0.759 (ETA: 0:00:01). ID: [3329, 3330, 3330, 333, 3329, 3330, 3330, 3331]
[WARNING: test_net_video.py: 126]: Video id 0 have been seen. Skip.
[WARNING: test_net_video.py: 126]: Video id 0 have been seen. Skip.
[INFO: test_net_video.py: 140]: 9262/9263 iter (19761/19761 videos): Time: 1.370 (ETA: 0:00:00). ID: [3332, 3333, 3331, 3331, 333, 3331, 3332, 3332]
[INFO: test_net_video.py: 247]: Num of empty videos: 0
[INFO: test_net_video.py: 248]: Num of corrupted videos: 0
[INFO: test_net_video.py: 249]: Max num of clips in a video: 30
[INFO: test_net_video.py: 250]: Min num of clips in a video: 30
[INFO: test_net_video.py: 254]: Clip1 accuracy: 0.22 percent (44/19761)
[INFO: test_net_video.py: 258]: Clip accuracy: 0.22 percent (1293/592830)
[INFO: test_net_video.py: 281]: --------------------------------------------------------------------------------
[INFO: test_net_video.py: 282]: top-1 accuracy: 0.27 percent
[INFO: test_net_video.py: 283]: top-5 accuracy: 1.29 percent
[INFO: test_net_video.py: 284]: --------------------------------------------------------------------------------
[INFO: test_net_video.py: 181]: Temporary file saved to: ../data/checkpoints/run_i3d_nlnet_400k/results_probs.pkl
[INFO: train_net_video.py: 234]: 10-clip spatial fcn testing finished
Can you give me some suggestions?