search

yoyomimi / AS-Net

Code for one-stage adaptive set-based HOI detector AS-Net.
MIT License
48 stars 1 forks source link

Train time. #3

Open kingwmk opened 3 years ago

kingwmk commented 3 years ago

Dear Chen. Hello. May I ask how long do we need to train this model?

yoyomimi commented 3 years ago

Hi, if you set the batch size to 32, you will need about 11-12h to train 90 epochs on HICO-DET, following our paper.

kingwmk commented 3 years ago

OK, Thank you so much. I see you use 32 GPUs to train it. It is so expensive! LOL.

yoyomimi commented 3 years ago

Aha, HICO-DET is a large dataset that contains 3w+ images for training, a relatively large batch size will save some time in the training process. If you need a lighter model, you can try to cut off some encoder and decoder layers or feature channels of both branches. This will compromise on the performance, but it can effectively cut down the computational cost. Then you can trade-off the accuracy and cost depending on your own needs.(old story LOL~)

zhihou7 commented 3 years ago

Hi, have you ever check the least number of GPUs to train your network?