search

MCG-NJU / TDN

[CVPR 2021] TDN: Temporal Difference Networks for Efficient Action Recognition
https://arxiv.org/abs/2012.10071
Apache License 2.0
368 stars 55 forks source link

Problems related to model performance #51

Closed jwfanDL closed 2 years ago

jwfanDL commented 2 years ago

Hi~ I have trained a TDN(ResNet50+8frame) on something-v1 as well as something-v2. However, there exists some gaps between the results we do and the reported one. On something-v2, best Prec@1 is only 61.567. We follow the command and process that you mentioned in README.md. I wonder that maybe some hyper-parameters settings in our experiments is not the same as your experiments.

截屏2022-01-19 下午3 12 31
yztongzhan commented 2 years ago

Hi @JwFbupt! Thanks for your question! Please provide your command here, so I could help you.

jwfanDL commented 2 years ago

rlaunch --cpu=48 --gpu=8 --memory=$((1024*140)) --max-wait-time=12h --charged-group=v_event -- python -m torch.distributed.launch --master_port 12347 --nproc_per_node=8 \ main.py somethingv2 RGB --arch resnet50 --num_segments 8 --gd 20 --lr 0.01 \ --lr_scheduler step --lr_steps 30 45 55 --epochs 60 --batch-size 8 \ --wd 5e-4 --dropout 0.5 --consensus_type=avg --eval-freq=1 -j 4 --npb

yztongzhan commented 2 years ago

Hi @JwFbupt! The command is right. Could you share the training log? For your convenience, we provide the training log and tensorboard file for TDN-R50-Segment-8 on SSV2. We hope these could be helpful for you.

yztongzhan commented 2 years ago

Hi @JwFbupt! Could you share your training log? Or, have you addressed this issue now? I will close this issue if you have no reply in two weeks.

jwfanDL commented 2 years ago

Hi @yztongzhan I'm sorry that I did not reply you in time. It is still a problem for us. Here is the training log log.txt TDN__somethingv2_RGB_resnet50_avg_segment8_e60.zip

yztongzhan commented 2 years ago

It seems no problem in the training log. Could you reproduce the results of other method like TSM and TEA on your SSV2?

yztongzhan commented 2 years ago

Besides, I want to know if you modify the validation code by yourself?

jwfanDL commented 2 years ago

I have reproduced TSM on somethingv2, and obtained promising results. In this version, we did not modify any codes. Thank you for your reply, and I will try to solve the problem.

jwfanDL commented 2 years ago

hi @yztongzhan, we have reproduce the results of TEA as well as TSM on sthv2. I would like to know, whether the official model use mixup method or flip on sthv2

yztongzhan commented 2 years ago

We use flip aug for SSV2 training, which also support in MMaction2 for such TSM. But we want to point out that our label generation script is different from MMaction2, you need to check our label carefully. I hope these could be useful for you.