GoogLeNet

ph
(GoogleNet에서 넘어옴)
이동: 둘러보기, 검색

paper review

  • 17 Sep 2014
  • pdf
  • abstract

    We propose a deep convolutional neural network architecture codenamed "Inception", which was responsible for setting the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC 2014). The main hallmark of this architecture is the improved utilization of the computing resources inside the network. This was achieved by a carefully crafted design that allows for increasing the depth and width of the network while keeping the computational budget constant. To optimize quality, the architectural decisions were based on the Hebbian principle and the intuition of multi-scale processing. One particular incarnation used in our submission for ILSVRC 2014 is called GoogLeNet, a 22 layers deep network, the quality of which is assessed in the context of classification and detection.

  • Hebbian principle?


graphical view

mxnet의 tutorial에 나오는 그림. plot Previous Output Previous Output conv_in3a_1x1 Convolution 1x1/1x1, 64 conv_in3a_1x1->Previous Output 3x28x28 bn_in3a_1x1_gamma bn_in3a_1x1_gamma bn_in3a_1x1_beta bn_in3a_1x1_beta bn_in3a_1x1_moving_mean bn_in3a_1x1_moving_mean bn_in3a_1x1_moving_var bn_in3a_1x1_moving_var bn_in3a_1x1 bn_in3a_1x1 bn_in3a_1x1->conv_in3a_1x1 64x28x28 bn_in3a_1x1->bn_in3a_1x1_gamma bn_in3a_1x1->bn_in3a_1x1_beta bn_in3a_1x1->bn_in3a_1x1_moving_mean bn_in3a_1x1->bn_in3a_1x1_moving_var relu_in3a_1x1 Activation relu relu_in3a_1x1->bn_in3a_1x1 64x28x28 conv_in3a_3x3_reduce Convolution 1x1/1x1, 64 conv_in3a_3x3_reduce->Previous Output 3x28x28 bn_in3a_3x3_reduce_gamma bn_in3a_3x3_reduce_gamma bn_in3a_3x3_reduce_beta bn_in3a_3x3_reduce_beta bn_in3a_3x3_reduce_moving_mean bn_in3a_3x3_reduce_moving_mean bn_in3a_3x3_reduce_moving_var bn_in3a_3x3_reduce_moving_var bn_in3a_3x3_reduce bn_in3a_3x3_reduce bn_in3a_3x3_reduce->conv_in3a_3x3_reduce 64x28x28 bn_in3a_3x3_reduce->bn_in3a_3x3_reduce_gamma bn_in3a_3x3_reduce->bn_in3a_3x3_reduce_beta bn_in3a_3x3_reduce->bn_in3a_3x3_reduce_moving_mean bn_in3a_3x3_reduce->bn_in3a_3x3_reduce_moving_var relu_in3a_3x3_reduce Activation relu relu_in3a_3x3_reduce->bn_in3a_3x3_reduce 64x28x28 conv_in3a_3x3 Convolution 3x3/1x1, 64 conv_in3a_3x3->relu_in3a_3x3_reduce 64x28x28 bn_in3a_3x3_gamma bn_in3a_3x3_gamma bn_in3a_3x3_beta bn_in3a_3x3_beta bn_in3a_3x3_moving_mean bn_in3a_3x3_moving_mean bn_in3a_3x3_moving_var bn_in3a_3x3_moving_var bn_in3a_3x3 bn_in3a_3x3 bn_in3a_3x3->conv_in3a_3x3 64x28x28 bn_in3a_3x3->bn_in3a_3x3_gamma bn_in3a_3x3->bn_in3a_3x3_beta bn_in3a_3x3->bn_in3a_3x3_moving_mean bn_in3a_3x3->bn_in3a_3x3_moving_var relu_in3a_3x3 Activation relu relu_in3a_3x3->bn_in3a_3x3 64x28x28 conv_in3a_double_3x3_reduce Convolution 1x1/1x1, 64 conv_in3a_double_3x3_reduce->Previous Output 3x28x28 bn_in3a_double_3x3_reduce_gamma bn_in3a_double_3x3_reduce_gamma bn_in3a_double_3x3_reduce_beta bn_in3a_double_3x3_reduce_beta bn_in3a_double_3x3_reduce_moving_mean bn_in3a_double_3x3_reduce_moving_mean bn_in3a_double_3x3_reduce_moving_var bn_in3a_double_3x3_reduce_moving_var bn_in3a_double_3x3_reduce bn_in3a_double_3x3_reduce bn_in3a_double_3x3_reduce->conv_in3a_double_3x3_reduce 64x28x28 bn_in3a_double_3x3_reduce->bn_in3a_double_3x3_reduce_gamma bn_in3a_double_3x3_reduce->bn_in3a_double_3x3_reduce_beta bn_in3a_double_3x3_reduce->bn_in3a_double_3x3_reduce_moving_mean bn_in3a_double_3x3_reduce->bn_in3a_double_3x3_reduce_moving_var relu_in3a_double_3x3_reduce Activation relu relu_in3a_double_3x3_reduce->bn_in3a_double_3x3_reduce 64x28x28 conv_in3a_double_3x3_0 Convolution 3x3/1x1, 96 conv_in3a_double_3x3_0->relu_in3a_double_3x3_reduce 64x28x28 bn_in3a_double_3x3_0_gamma bn_in3a_double_3x3_0_gamma bn_in3a_double_3x3_0_beta bn_in3a_double_3x3_0_beta bn_in3a_double_3x3_0_moving_mean bn_in3a_double_3x3_0_moving_mean bn_in3a_double_3x3_0_moving_var bn_in3a_double_3x3_0_moving_var bn_in3a_double_3x3_0 bn_in3a_double_3x3_0 bn_in3a_double_3x3_0->conv_in3a_double_3x3_0 96x28x28 bn_in3a_double_3x3_0->bn_in3a_double_3x3_0_gamma bn_in3a_double_3x3_0->bn_in3a_double_3x3_0_beta bn_in3a_double_3x3_0->bn_in3a_double_3x3_0_moving_mean bn_in3a_double_3x3_0->bn_in3a_double_3x3_0_moving_var relu_in3a_double_3x3_0 Activation relu relu_in3a_double_3x3_0->bn_in3a_double_3x3_0 96x28x28 conv_in3a_double_3x3_1 Convolution 3x3/1x1, 96 conv_in3a_double_3x3_1->relu_in3a_double_3x3_0 96x28x28 bn_in3a_double_3x3_1_gamma bn_in3a_double_3x3_1_gamma bn_in3a_double_3x3_1_beta bn_in3a_double_3x3_1_beta bn_in3a_double_3x3_1_moving_mean bn_in3a_double_3x3_1_moving_mean bn_in3a_double_3x3_1_moving_var bn_in3a_double_3x3_1_moving_var bn_in3a_double_3x3_1 bn_in3a_double_3x3_1 bn_in3a_double_3x3_1->conv_in3a_double_3x3_1 96x28x28 bn_in3a_double_3x3_1->bn_in3a_double_3x3_1_gamma bn_in3a_double_3x3_1->bn_in3a_double_3x3_1_beta bn_in3a_double_3x3_1->bn_in3a_double_3x3_1_moving_mean bn_in3a_double_3x3_1->bn_in3a_double_3x3_1_moving_var relu_in3a_double_3x3_1 Activation relu relu_in3a_double_3x3_1->bn_in3a_double_3x3_1 96x28x28 avg_pool_in3a_pool Pooling avg, 3x3/1x1 avg_pool_in3a_pool->Previous Output 3x28x28 conv_in3a_proj Convolution 1x1/1x1, 32 conv_in3a_proj->avg_pool_in3a_pool 3x28x28 bn_in3a_proj_gamma bn_in3a_proj_gamma bn_in3a_proj_beta bn_in3a_proj_beta bn_in3a_proj_moving_mean bn_in3a_proj_moving_mean bn_in3a_proj_moving_var bn_in3a_proj_moving_var bn_in3a_proj bn_in3a_proj bn_in3a_proj->conv_in3a_proj 32x28x28 bn_in3a_proj->bn_in3a_proj_gamma bn_in3a_proj->bn_in3a_proj_beta bn_in3a_proj->bn_in3a_proj_moving_mean bn_in3a_proj->bn_in3a_proj_moving_var relu_in3a_proj Activation relu relu_in3a_proj->bn_in3a_proj 32x28x28 ch_concat_in3a_chconcat ch_concat_in3a_chconcat ch_concat_in3a_chconcat->relu_in3a_1x1 64x28x28 ch_concat_in3a_chconcat->relu_in3a_3x3 64x28x28 ch_concat_in3a_chconcat->relu_in3a_double_3x3_1 96x28x28 ch_concat_in3a_chconcat->relu_in3a_proj 32x28x28

위 net의 코드를 github에서 볼 수 있다. 아무리 봐도 tensorflow보다 간결하고 직관적이란 말이야...

a

원본 주소 "http://samediff.kr/wiki/index.php?title=GoogLeNet&oldid=12470"