search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.92k stars 933 forks source link

About fine-tuning resnet-18 on hmdb #114

Open Collinxz opened 6 years ago

Collinxz commented 6 years ago

Hi. I want to fine-tune resnet-18 on hmdb1 with lr=0.001 and weight_decay=0.00001,but after 300 epoch, the val only get 51% accuracy which below 56.4 in your paper.What do i miss? Thank you!

huangwei1995 commented 6 years ago

I have the same question when fine-tuning.

fengwuxuan commented 6 years ago

Me too, and I try to train from scratch on hmdb and ucf101, but the val accuracy is lower than paper's...

shuangshuangguo commented 5 years ago

@Collinxz @huangwei1995 @fengwuxuan You can try my scripts to finetune, which I got 53.45% clip-leval accuracy. https://github.com/shuangshuangguo/3D-ResNet-Pytorch