How to train on something dataset?

mzolfaghari / ECO-efficient-video-understanding

Code and models of paper " ECO: Efficient Convolutional Network for Online Video Understanding", ECCV 2018

MIT License

437 stars 96 forks source link

How to train on something dataset? #33

Closed kmyfoer closed 5 years ago

kmyfoer commented 5 years ago

I basically reproduce the result on hmdb51, but I can't reproduce same result on something dataset. Can you tell me how to set dropout, learning rate or other training trick?

Thank u very much！

mzolfaghari commented 5 years ago

@kmyfoer we set the dropout for this dataset to 0.3. You can find the information regarding SomethingSomething dataset here and here. Using the same setting (Caffe version), you should be able to reproduce the results of the paper. Please let me know if had further questions.

Best,

ntuyt commented 5 years ago

I also only achieve around 40 accuracy in something dataset using the provided caffe model. It is worse than reported 42.4. Is it because of the difference of the data? Does the author use the frames of something dataset downloaded from the official website?