search

yfxc / pseudo-3d-tensorflow

Tensorflow implement for Pseudo-3d-residual network.
21 stars 5 forks source link

Training problem #5

Open leesky1c opened 5 years ago

leesky1c commented 5 years ago

My env: Tensorflow 1.8.0

Well, setting fps at 5 in ffmpeg will result in not-enough frames in some videos like 'UCF-101/HorseRiding/v_HorseRiding_g14_c02'. Therefore, when making the batch as input, some errors would happen like this.

ValueError: setting an array element with a sequence.

I think increasing the number of fps or checking about the number of video's frames may help.

leesky1c commented 5 years ago

Besides, could you share the final accuracy both on training set and testing set?

leesky1c commented 5 years ago

And why the NUM_CLASS is 3 in the setting.py file...

yfxc commented 5 years ago

And why the NUM_CLASS is 3 in the setting.py file...

My env: Tensorflow 1.8.0

Well, setting fps at 5 in ffmpeg will result in not-enough frames in some videos like 'UCF-101/HorseRiding/v_HorseRiding_g14_c02'. Therefore, when making the batch as input, some errors would happen like this.

ValueError: setting an array element with a sequence.

I think increasing the number of fps or checking about the number of video's frames may help.

Thank you for your advice.

yfxc commented 5 years ago

And why the NUM_CLASS is 3 in the setting.py file...

Because I apply the model for my own project where only exists there classes.I forget the accuracy for UCF dataset,but for my own project,accuracy for testing set is about 85% when it converged on training set(98+%),the problem is that my project dataset has medium quality.

wuchichung commented 5 years ago

hi @yfxc , can I have some ideas about the size of your dataset ?