About hyper-params and frozen bn

[guvcolie]

Thank you for your nice work! I have 2 questions, could you help me?

1. You said it's better to set batchsize=16, you mean total batchsize or each gpu batchsize? There's a hyper-param 'batch_size_per_gpu' in your code, if I use 2 GPU, should I set it 8 or 16?
2. Did you train on train_aug dataset, then frozen bn and finetuning on train dataset?

it's my hyperparam: python -u train_voc.py
--data_root_path=/home/work/dataset/VOCdevkit/VOC2012_aug
--checkpoint_dir=./checkpoints/
--result_filepath=./Results/
--backbone=resnet101
--output_stride=16
--gpu=0,1
--batch_size_per_gpu=16
--dataset=voc2012_aug
--base_size=513
--crop_size=513
--freeze_bn=False
--weight_decay=4e-5
--lr=0.007
--iter_max=30000
--poly_power=0.9