May I ask how long it will take to train the BP4D and DISFA data sets? I use two GeForce RTX 2080，it takes 14 days of training.

machanic / AU_R-CNN

The official implementation code of paper: "AU R-CNN：Encoding Expert Prior Knowledge into R-CNN for Action Unit Detection".

https://arxiv.org/abs/1812.05788

87 stars 24 forks source link

May I ask how long it will take to train the BP4D and DISFA data sets? I use two GeForce RTX 2080，it takes 14 days of training. #4

Open lucky-starh opened 4 years ago

machanic commented 4 years ago

If you use memcached and pylibmc to store the bounding box coordinates, the training will be accelarate by the cache. You must assure the cache is correctly used in the program. When I use the Nvidia Tesla P100 card, the training will be finished in 4 days at most.

lucky-starh commented 4 years ago

Thank you. I have one more question: OutOfMemoryError. The defualt batch_size =20, can only change the bitch_size ?

machanic commented 4 years ago

@lucky-starh if the input image is 512 x 512, batch size =20 is too large to the GPU memory, batch size = 5 will be better.

lucky-starh commented 4 years ago

Has been an error: Crop image error, always appear in the Img_RightCamera/SN029 or Img_RightCamera/SN029 folder

machanic commented 4 years ago

@lucky-starh Maybe the head is not frontal position, you need to check that reason. But I remember I have already added the exception code for handling this case....

lucky-starh commented 4 years ago

I haven't found the conversion to BGR->RGB image in get_example() function.

machanic commented 4 years ago

@lucky-starh I think you do not need to convert from BGR -> RGB. Have you fixed this issue?