search

Guocode / SlowFast-Networks

implement of slowfast networks
82 stars 19 forks source link

Result worse than 3dresnet50 #2

Open Guocode opened 5 years ago

Guocode commented 5 years ago

dataset:something-somethingv2 slowfast only get 31% acc 3d-resnet-50 get 36% acc under the same settings. If anyone has do tests on any other dataset, you can publish your score here.

nemonameless commented 5 years ago

There are no links between the fast and slow branches in your code , but this may not be the main reason to the relatively bad results. Do you think it's related to the number of GPU and pre-training? How about using pre-train model? maybe you just overfit？

SidHard commented 5 years ago

I test this code on my dataset，and the result worse than I3D. I guess the main reason is no pre-train model.

Guocode commented 5 years ago

@SidHard I find it's essential to pretrain for 3DCNN on imagenet, is your test of I3D and slowfastnet both trained from scratch?

95xueqian commented 5 years ago

I have trained this model on ucf101 dataset , the accuracy on test dataset is only 40%

dmenig commented 5 years ago

@SidHard I find it's essential to pretrain for 3DCNN on imagenet, is your test of I3D and slowfastnet both trained from scratch?

The paper exhibits results with random init, so It should be possible to have the same results with no imagenet pretraining. That's in fact, to me one of the impressive things in it.

hulianyuyy commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

idealboy commented 5 years ago

I test this code on my dataset，and the result worse than I3D. I guess the main reason is no pre-train model.

Yes, I think it converge very slow, and it have a swing about the acc and loss.

It is much worse than I3D with the same train datase and val dataset, and I have the same sampling strategy in the clip like my I3D training.

dmenig commented 5 years ago

I didn't have the ressources to test with the exact same parameters as the paper, but I reached 63% on kinetics in validation. However it does train from scratch which is very impressive. I do think it's because of my parameters because with the same parameters, I3D performs on par with this. But like @idealboy it's way worse than other models on my private dataset.

salmedina commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

Hello~ Which set of hyperparameters did you use for this? I was able to achieve 0.51 accuracy on UCF-101 by using the default parameters described on the paper. Thank you!

hulianyuyy commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

Hello~ Which set of hyperparameters did you use for this? I was able to achieve 0.51 accuracy on UCF-101 by using the default parameters described on the paper. Thank you!

i used batchsize 16, lr 0.01 and epoch 200 when trained with Pytorch. I think this network needs pretrained model to help it perform well otherwise it will have a bad performance.

salmedina commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

Hello~ Which set of hyperparameters did you use for this? I was able to achieve 0.51 accuracy on UCF-101 by using the default parameters described on the paper. Thank you!

i used batchsize 16, lr 0.01 and epoch 200 when trained with Pytorch. I think this network needs pretrained model to help it perform well otherwise it will have a bad performance.

All the way to 200 epochs? I ran it for only 80 epochs. Which optimizer did you use? Do you have any idea if anyone has released a pretrained model on Kinetics? I was thinking about pretraining it on Kinetics and release it if no one else has done it. But it is going to take me weeks...

hulianyuyy commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

Hello~ Which set of hyperparameters did you use for this? I was able to achieve 0.51 accuracy on UCF-101 by using the default parameters described on the paper. Thank you!

i used batchsize 16, lr 0.01 and epoch 200 when trained with Pytorch. I think this network needs pretrained model to help it perform well otherwise it will have a bad performance.

All the way to 200 epochs? I ran it for only 80 epochs. Which optimizer did you use? Do you have any idea if anyone has released a pretrained model on Kinetics? I was thinking about pretraining it on Kinetics and release it if no one else has done it. But it is going to take me weeks...

emmm i just used 200 epochs because i tried it when training on with other modes and it performs well. This model could bring out better if equipped with pretrained model but due to limits of GPUs i can't train it with Kinetics...Not having a pretrained model is a pity to it.

hulianyuyy commented 5 years ago

i just trained the SlowFast Model on UCF101 but i got ~ 55% accuracy on test split1. i think it's due to action complexities and video lengths. maybe another reason is hyperparameters. we shoule use our own hyperparameters on each dataset.

Hello~ Which set of hyperparameters did you use for this? I was able to achieve 0.51 accuracy on UCF-101 by using the default parameters described on the paper. Thank you!

i used batchsize 16, lr 0.01 and epoch 200 when trained with Pytorch. I think this network needs pretrained model to help it perform well otherwise it will have a bad performance.

All the way to 200 epochs? I ran it for only 80 epochs. Which optimizer did you use? Do you have any idea if anyone has released a pretrained model on Kinetics? I was thinking about pretraining it on Kinetics and release it if no one else has done it. But it is going to take me weeks...

oh i just forgot to tell i used standard SGD to optimize.