Wrong Val acc?

[hawklucky]

Thanks for your excellent work! When I trained the model at around 10000 iters, the batch acc of training set is ~96%, but the batch acc of val set is ~ 29%, what's wrong? the training list and val list is the list i used for training in CNN which works well.

[hx173149]

Thanks for telling me this problem, I actually didn't make a validation in UCF101 dataset.
I just use this code to a binary classification problem, it works......
I will check this problem after 10.09, I am on my vocation now......
cheers~

[hx173149]

@hawklucky Hi I have also found this bug in my experienment, when I get this train accuracy(93.75%):

I can only get a low val accuracy(32.8%):

I think there must be some over fitting problem, maybe some bad parameter like L2 regular, ExponentialMovingAverage......
I will check where is the problem next days, if you find some bugs you can tell me too, thanks~

[hx173149]

@hawklucky recently, I have solved this problem. And if you want to get a high acc as same as the caffe version you must fine-tuning from the author's sports1m model. So I have transferred the caffe model to TF model, and fine-tuning this model from UCF101 dataset, and get a 72% acc.
you can download my pretrained model from here:
https://www.dropbox.com/sh/8wcjrcadx4r31ux/AAAkz3dQ706pPO8ZavrztRCca?dl=0