你好,结果不一致

[sjjdd]

你好,我在cifar10上运行您的train_search.py的时候,因为只有6G的gpu,没有改代码,运行python train_search.py --batch_size 96
50个epoch后得到的是
01/18 12:02:10 AM epoch 49 lr 0.000000e+00
01/18 12:02:10 AM genotype = Genotype(normal=[('skip_connect', 0), ('skip_connect', 1), ('skip_connect', 0), ('skip_connect', 2), ('max_pool_3x3', 0), ('sep_conv_3x3', 1), ('max_pool_3x3', 1), ('skip_connect', 0)], normal_concat=range(2, 6), reduce=[('max_pool_3x3', 1), ('sep_conv_5x5', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 2), ('sep_conv_5x5', 2), ('sep_conv_5x5', 0), ('skip_connect', 1), ('dil_conv_5x5', 2)], reduce_concat=range(2, 6))
如你所见,在normal cell中出现了许多的skip_connect,请问超参数一定要和您的设定的一样吗,是因为我的batch_size设置的太小了吗?

[yuhuixu1993]

Hi, I did not try batchsize 96, I used 256 as I claimed in the paper. If you changed batch-size, The learning rate should be changed corresponding.

[sjjdd]

谢谢,那意思就是我的batch_size更小的话,应当设置更小的学习率来进行网络结构搜索是吗?