Do you train imagenet with batch size 64 too? It goes out of memory on 1080 Ti.

[weitingyuk]

@yxchng How is your training result?

[jianghaojun]

I set the batchsize=256/lr=0.1, but the training result(top1-acc: 77.64) is much lower than paper reported(top1-acc: 78.24)! More details about hyperparameters are listed as below. The epoch setting is converted from the iteration which is mentioned in paper. If we set the batchsize as 256, then there is 5k iteration in 1 epoch. According to the paper, we should decay the learning rate at 200k/5k=40, 400k/5k=80, 500k/5k=100 epoch, and terminate training at 530/5k=106 epoch.

The learning rate is divided by 10 at 200k, 400k, 500k iterations. We terminate training at 530k iterations.

Hyperparameter settings

args.epochs = 106
args.batch_size = 256
### data transform
args.autoaugment = False
args.colorjitter = False
args.change_light = True
### optimizer
args.optimizer = 'SGD'
args.lr = 0.1
args.momentum = 0.9
args.weigh_decay_apply_on_all = True  # TODO: weight decay apply on which params
args.weight_decay = 1e-4
args.nesterov = True
### criterion
args.labelsmooth = 0
### lr scheduler
args.scheduler = 'uneven_multistep'
args.lr_decay_rate = 0.1
args.lr_milestone = [40, 80, 100]