caffe and build with "WITH_PYTHON_LAYER := 1"
MtripLoss are added to conv4_3, conv5_3, fc6 and fc7, and the MEucliLoss is added to conv1_1
This is a strange phenomenon. Using the 50% of larger response node to reconstruct input image,
roughly the content of the input image can be restored. But when usin the 20-30% of smaller response
node to reconstruct the input image, the roughly the content of the input image also can be restored.
