tsn-pytorch
tsn-pytorch copied to clipboard
yjxiong
Issue about training Kinetics400
Hi, is there anyone try training TSN on Kinetics400?
I train the TSN using default hyperparameter setting given in the paper (i.e., lr=0.001, lr_steps=[30,60], epochs=80, batch_size=128, 3 segments, backbone=bninception) but only get a validation accuracy 64.88% on split1, which is much lower than the accuracy reported on "http://yjxiong.me/others/kinetics_action/".
Why? Did I use the wrong hyperparameter setting? How to get such high validation accuracy?
I am training tsn-pytorch on kinetics400 data set. However, I get the accuracy 33% for train set while accuracy 5% for validate set. Did you complete the experiment for kinetics400?
My testing result on kinetics is
Testing Results: Prec@1 44.483 Prec@5 70.912 Loss 2.53040
It is also quite low accuracy compare to the reported results
hello, i need the kinetics dataset, can you send me that? thank you! [email protected]
The original dataset is so large, maybe you can try "https://pytorch.org/docs/1.2.0/torchvision/datasets.html?highlight=kinetics#torchvision.datasets.Kinetics400".
thanks for your reply! I want to know if I can get the extracted optical flows datasets and frames datasets of kinetics from you or elsewhere . I train or evaluate on a server with 12 GB memory and TITAN Xp GPU, Is it feasible for such a device to conduct pre_training of action recognition on kinetics dataset? If not, is it feasible to train model only on UCF101 and HMDB51 dataset without pre_ training on the kinetics dataset?
------------------ 原始邮件 ------------------ 发件人: "yjxiong/tsn-pytorch" @.>; 发送时间: 2021年5月26日(星期三) 下午2:38 @.>; @.@.>; 主题: Re: [yjxiong/tsn-pytorch] Issue about training Kinetics400 (#87)
hello, i need the kinetics dataset, can you send me that? thank you! @.***
The original dataset is so large, maybe you can try "https://pytorch.org/docs/1.2.0/torchvision/datasets.html?highlight=kinetics#torchvision.datasets.Kinetics400".
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.