search

USTC-Video-Understanding / I3D_Finetune

TensorFlow code for finetuning I3D model on UCF101.
144 stars 43 forks source link

fc_out, relu, top_k_op #2

Open WuJunhui opened 6 years ago

WuJunhui commented 6 years ago

https://github.com/USTC-Video-Understanding/I3D_Finetune/blob/master/Demo_Transfer_rgb.py#L139 Hi, I think the last fc layer may should not use an activation unit. If the output of fc layer is all negative, after the ReLU unit, it would be all zero.
In this case, top_k_op = tf.nn.in_top_k(fc_out, label_holder, 1) will always return True.

Rhythmblue commented 6 years ago

I think you are right. When I use this code to train , the result will be all the same sometimes. You can just remove parameter of activation.

vra commented 6 years ago

Hi @WuJunhui , Thanks for your helpful advice, we have fix this issue in latest version of this repo. Please run git pull to download the newest code.