some questions about GPU Devices

[Guzaiwang]

dear authors:
thanks for your great work. however, I have some concerns about training from GPU devices. On my custom GPU device ( Tesla V100 Memory - 32GB), I can only train the efficientdet-b3 with batch_size=4 on the single GPU card.

What is your training parameters setting, such as gpu_devices and memory, batch_size?

[mingxingtan]

Hi @Guzaiwang , all models are trained with TPUv3, which has 16GB memory per core. I am able to train EfficientDet-D7 with batch size 4 per core.

GPU memory seems to be an issue, I haven't looked into this. I am closing this bug for now, and use #85 to track GPU memory related issues.