search

gurkirt / realtime-action-detection

This repository host the code for real-time action detection paper
Other
319 stars 97 forks source link

Frame ratios for each class #15

Closed zhujiagang closed 6 years ago

zhujiagang commented 6 years ago

https://github.com/gurkirt/realtime-action-detection/blob/0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L82 How do you change the above class ratios to the following? https://github.com/gurkirt/realtime-action-detection/blob/0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L84

Is there any performance improvment after this change?

gurkirt commented 6 years ago

There shouldn't be substantial change as long as classes are balanced. However, it is about 1% video-mAP change if ratios are all ones.

On Mon, Feb 26, 2018 at 9:46 AM, zhujiagang notifications@github.com wrote:

https://github.com/gurkirt/realtime-action-detection/blob/ 0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L82 How do you change the above class ratios to the following? https://github.com/gurkirt/realtime-action-detection/blob/ 0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L84

Is there any performance improvment after this change?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/gurkirt/realtime-action-detection/issues/15, or mute the thread https://github.com/notifications/unsubscribe-auth/AGv-o_rEKNihdJuonKqgXb_FzQmkW6N1ks5tYn1kgaJpZM4SS8GX .

jiaozizhao commented 6 years ago

Hi, could you explain how did you get the ratios numbers? Is it for re-balancing the samples of different classes? Could you give more explanation, please? Thank you very much.

gurkirt commented 6 years ago

Please follow the code; If you uncomment line 85 https://github.com/gurkirt/realtime-action-detection/blob/0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L85

you will get newratios at https://github.com/gurkirt/realtime-action-detection/blob/0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L153

How many sample you want per class is defined at https://github.com/gurkirt/realtime-action-detection/blob/0e56ce769794e1e9d4526b7fb46d80ecdb796330/data/ucf24.py#L149

zhujiagang commented 6 years ago

@gurkirt If ratios are all ones, perhaps It would take more time than now, maybe more than four days with one gpu? How much time it takes you to finish one training?

gurkirt commented 6 years ago

I trained for 120k iterations. That is a constant time depending upon your hardware. I don't remember the excat difference in numbers but wasn't a lot (<1%) for 120k iterations with ratios equal to ones.

On Mon, 19 Mar 2018, 2:08 am zhujiagang, notifications@github.com wrote:

@gurkirt https://github.com/gurkirt If ratios are all ones, perhaps It would take more time than now, maybe more than four days with one gpu? How much time it takes you to finish one training?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/gurkirt/realtime-action-detection/issues/15#issuecomment-374080897, or mute the thread https://github.com/notifications/unsubscribe-auth/AGv-o0LZNy3z2fpP8uqmX3wR1uO9iZxbks5tfxMVgaJpZM4SS8GX .