search

hx173149 / C3D-tensorflow

C3D is a modified version of BVLC tensorflow to support 3D ConvNets.
MIT License
587 stars 265 forks source link

ucf101 train from scratch #79

Open laura-wang opened 6 years ago

laura-wang commented 6 years ago

Use the train_c3d_ucf101.py to train the ucf101 dataset from scratch, I can only get ~30% acc, which is much more lower than the reported ~45% accuracy.

Can anyone get higher accuracy?

kjunhwa commented 6 years ago

i also got ~30% and i don't know what to do...

xmzhuxm commented 5 years ago

where i need to modify,i train from scratch and get a very low accuracy,maybe i modify error location,can you help me,thank you very much! i stacked here for many days.

i only set "use_pretrained_model = False" in the train_c3d_ucf101.py

xmzhuxm commented 5 years ago

i also got ~30% and i don't know what to do...

where i need to modify,i train from scratch and get a very low accuracy,maybe i modify error location,can you help me,thank you very much! i stacked here for many days.

i only set "use_pretrained_model = False" in the train_c3d_ucf101.py

MClarkTurner commented 5 years ago

Was anyone able to address this issue? It is concerning that we are unable to reproduce the same results that the GitHub presents. I have trained C3D from scratch using UCF-101 on split 1 at both 5k and 20k iterations (using the 5fps and 10 batch samples as suggested by the instructions in this GitHub). In both cases I am unable to generate an accuracy >30%.

laura-wang commented 5 years ago

I am able to reproduce the results of Du Tran's paper result, i.e., ~45% now. Our paper is accepted in CVPR 2019. And I will realse my code later.