Have you trained the model for the UCF101 from scratch?

irhum / R2Plus1D-PyTorch

PyTorch implementation of the R2Plus1D convolution based ResNet architecture described in the paper "A Closer Look at Spatiotemporal Convolutions for Action Recognition"

MIT License

363 stars 62 forks source link

Have you trained the model for the UCF101 from scratch? #5

Open lixianhang opened 6 years ago

lixianhang commented 6 years ago

I just trained for 45 epochs and set LR 0.01 for UCF101, but I got the really terrible result. Firstly the training loss is stable around 4, but cannot get lower(because the dataset is small, it should be overfitted). And second, the test accuracy is around 1% even after 45 epochs. I followed your instruction to process the dataset. I don't know where is wrong. Could you please tell me your result if you have?

qingxiaoli commented 5 years ago

I have met the same problem. How did you solve the problem? Can you please share?

yechanp commented 5 years ago

Hello , it might be too late. I found that learning rate is decreasing too fast. Change gamma 0.1 to big number such as 0.9. It will be better