search

moabitcoin / ig65m-pytorch

PyTorch 3D video classification models pre-trained on 65 million Instagram videos
MIT License
265 stars 30 forks source link

fine tune the pre-trained model on UCF101 #38

Open Yueeeeee-1 opened 4 years ago

Yueeeeee-1 commented 4 years ago

if I want to fine tune the pre-trained model on UCF101, how could I get the performance of 96.8%? In my settings, fine-tuning was only performed to train layer4 and the fully connected layer, and the learning rate is 0.0001, am I wrong? Cause the result I got just 78.5%, can you help me? Thank you!

bjuncek commented 4 years ago

Other than parameter fine-tuning (I'd say 15 epochs, LR:1e-4 for the convolutional blocks and 1e-3 for the FC layer), make sure the evaluation script is correct.

If I recall correctly, the results are reported as video level accuracy on 10 uniformly sampled clips per video, averaged over 3 cross validation splits. I'd expect the clip level accuracy (what you're likely measuring) to be around >80%.

Yueeeeee-1 commented 4 years ago

@bjuncek Thanks for your reply, I did test the clip level accuracy, and I didn't notice the difference between the video level and clip level, I will check it. and for the parameter fine-tuning, ''LR:1e-4 for the convolutional blocks '' , it means all convolutional blocks or just the last one? thanks for your patience!

bjuncek commented 4 years ago

Make sure you're averaging softmaxes rather than predictions (w.r.t. the video level acc).

For all conv blocks

paden118 commented 2 years ago

Hi, what is the accuracy of your final fine-tuning on ucf101?