For the competition of TinyMind Multi-label classification, I implement the model with three methods:
a simple net
vgg16 to extra the feature of image, and three fullconnection layers for classification
ResNet50 to extra the feature of image, and three fullconnection layers for classification
Util 9.2, I trained one epoch by the second method, and submit the valid result to TinyMind, catch the score 16.08.
Because of the GPU memory size is 2G, something can't be tried, so I'll keep training the Network more epoch with CPU.