search

yjxiong / tsn-pytorch

Temporal Segment Networks (TSN) in PyTorch
BSD 2-Clause "Simplified" License
1.07k stars 308 forks source link

about flow accuracy #56

Closed ll490187880 closed 5 years ago

ll490187880 commented 6 years ago

The accuracy of ucf101 (split 1) with flow is only 81%, but I used the official code to extract the flow. And the acc is 87% on the paper.Can you tell me what is wrong? @yjxiong

yjxiong commented 6 years ago

I can't. You have to add more details here.

imnotk commented 6 years ago

in my experiment，my experiment accuracy is only 81%. i use 5e-3 learning rate,5e-4 weight decay , 200 and 300 epoch decrease learning rate with factor 10,stop at 350 epoch, use tvl1 flow, and pretrained with inception v2 in tensorflow , and my training accuracy is about 97% without parital batch norm.The result is wierd

yjxiong commented 6 years ago

Not sure how you implemented it in TensorFlow. This codebase is for PyTorch.

Flyzanetothesky commented 5 years ago

Hi, @imnotk , I now reproduce TSN use Tensorflow, But I have no idea about how to copy the mean of Parameters to first convolution layer, could you give me some suggestion ?

imnotk commented 5 years ago

@Flyzanetothesky just repeat its mean value of channel