About the GPU memory about refinedet HOT 4 CLOSED

@lz20061213 , Hi
By default, we use four GPUs (12G) with 32 batch size, so every GPU have 8 images. If you change the gpus for a single one, you should change the 'batch_size=32' correspondingly. Maybe 'batch_size=16' or 'batch_size=8' will run successfully. Because 32 images are too many for one 11G GPU.

from refinedet.

@sfzhang15 Hi, thanks for your reply! I wonder if i change the batch_size to solve the GPU memory problem, should i change the max_iter and stepvalue at the same time?

from refinedet.

@lz20061213 Hi, you do not need to change other parameters. Since the 'accum_batch_size' is unchanged. If you change the 'batch_size' form 32 to 8, the 'iter_size' of four GPUs is 1, while the 'iter_size' of one GPU will become 4. But the accum_batch_size is the same, i.e., accum_batch_size=32.

from refinedet.

@sfzhang15 Got it! Truly thanks!

from refinedet.