mmlab-cu / polynet Goto Github PK

What is the configuration used in the paper for ResNet-500, i.e. #layers per stage?

Is the configuration carefully chosen using some empirical study?

I0726 09:27:07.791891 1097 net.cpp:98] Creating Layer data
F0726 09:27:07.791920 1097 net.cpp:394] Duplicate blobs produced by multiple sources.

We train very deep poly net with 10 GPUs and 30-batch-size. It leverages very slowly (almost no leverage).
Is there any solution?

After inspection of the model parameters, I noticed that each BatchNorm layer inside a poly-3 module has a different set of parameters. For example parameters for layers inception_resnet_b_1_path1_1x7_bn, inception_resnet_b_1_poly2_block1_path1_1x7_bn and inception_resnet_b_1_poly2_block2_path1_1x7_bn differ while inception_resnet_b_1_path1_1x7, inception_resnet_b_1_poly2_block1_path1_1x7 and inception_resnet_b_1_poly2_block2_path1_1x7 Convolutional layers have the same parameters.

Is my understanding correct that respective Convolutional layers are shared between all Inception blocks inside a poly-3 module while each BatchNorm layer inside an Inception block has its own set of parameters and therefore parameters of BatchNorm layers are not shared?