search

leftthomas / R2Plus1D-C3D

A PyTorch implementation of R2Plus1D and C3D based on CVPR 2017 paper "A Closer Look at Spatiotemporal Convolutions for Action Recognition" and CVPR 2014 paper "Learning Spatiotemporal Features with 3D Convolutional Networks"
58 stars 14 forks source link

Something about R2Pluse1D_model trained on Kinetics600 #4

Closed huamgmin closed 4 years ago

huamgmin commented 4 years ago

Hi, l know you haven't finished the R2Pluse1D_model train on Kinetics600, BUT have you saved the parameters model that trained 350 hours, l want to continue train this model, if l finished it ,i'll share it to you . THANKS! qq: 2458237010

huamgmin commented 4 years ago

And emmm, FIRST ,I don't know whether what i think is right, i find in R2Plus1D.py that when we choose the first_conv =False, you operate BN and RELU after conv, but in CLASS---SpatioTemporalResBlock, you do it again, which i think is wrong.
AND i am such a new hand in DL,so if i was wrong ,please tell me!!! THANK YOU VERY MUCH !!

daihu-ye commented 3 years ago

Hi, huamgmin. Have you done training the model and got the results?

huamgmin commented 3 years ago

I've got 60% on HMDB51 and 93% on UCF101