Hi, how many gpus did you use?

[TiankaiHang]

For example, PSPNet resnet50, cityscapes 1024 * 512

[CharlesPikachu]

By default, I use four Tesla-V100-32GB GPUs for training the models.

[TiankaiHang]

Got that, thanks~

---

for this setting, [For example, PSPNet resnet50, cityscapes 1024 * 512], if you set bs =16, will it run out of memory?

[CharlesPikachu]

I have not met this situation during training. If your GPU memory is limited, you can try to set output stride as 16 and use more GPUs since we use syncbn by default.

[TiankaiHang]

I used two 2080Ti (11G) GPU to train models.

While i set img_per_gpu = 1, it needs 3~4GB memory per GPU, however, when i try to double the batchsize, it ran out of the memory.

My setting:
PSPNet resnet50 os16, cityscapes 1024 * 512

[CharlesPikachu]

1. Our framework only supports setting batch_size rather than img_per_gpu, so img_per_gpu = batch_size / num_gpus.
2. I think 11GB * 2 is enough for PSPNet resnet50 os16, cityscapes 1024 * 512, batch_size=4. I will test this setting when I'm free.

[CharlesPikachu]

Hello, for PSPNet resnet50 os16, cityscapes 1024 * 512, batch_size=4，it requires 11374MiB GPU memory per GPU if you only own two GPUs.

[TiankaiHang]

Thanks for your kind answer!