mobilenet structure omitts last pooling layer

[zhenglaizhang]

Have a check at the mobilenet structure, and most of the weights are in the final FC layer:

pre_fc1_weight parameter size=25690112, shape=(512L, 50176L)

and looking at the code, the last pooling layer is omitted ().

insightface/src/symbols/fmobilenet.py

Lines 75 to 80 in f1cd542

	conv_13_dw = Conv(conv_12, num_group=512, num_filter=512, kernel=(3, 3), pad=(1, 1), stride=(2, 2), name="conv_13_dw") # 14/7
	conv_13 = Conv(conv_13_dw, num_filter=1024, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_13") # 7/7
	conv_14_dw = Conv(conv_13, num_group=1024, num_filter=1024, kernel=(3, 3), pad=(1, 1), stride=(1, 1), name="conv_14_dw") # 7/7
	conv_14 = Conv(conv_14_dw, num_filter=1024, kernel=(1, 1), pad=(0, 0), stride=(1, 1), name="conv_14") # 7/7
	body = conv_14
	fc1 = symbol_utils.get_fc1(body, num_classes, fc_type)

Is this by design? Ignoring the last pooling layer leads to much larger model size -:(

[nttstar]

please refer to network structure section in our paper.

[zhenglaizhang]

@nttstar thanks for the info, I will read carefully the paper.

[zhenglaizhang]

@nttstar Hi, you are definitely right, I was using the version 'E' which removed the GP layer.

Then I tried the mobilenetv1 with D as version output, whose model size is around 15M, and started to train 15 hours ago, with 128 batch size, but found the test accuracy is still around 0.5.

Have you guys tried to train with such settings? maybe I need to tune other hyperparameters...

[nttstar]

@zhenglaizhang We all use 'E' in our recent experiments.

[mmxuan18]

@zhenglaizhang i also has this question, and in the code last conv directly connect with mx.sym.FullyConnected, is this func will do flatten inner?

which part of the paper explain this modify?