Pre training: the pre training stag

Pre training: the pre training stage of expert automatic encoder is divided into three parts, and each part has different purposes. First, we use all training samples to train the basic automatic encoder AEO to find the context independent initial compression feature map. Then, we perform context clustering on the initial compressed feature map of AEO to find ne context sensitive clusters. Finally, these clusters are used to train expert automatic coders initialized by basic automatic coders.The purpose of the basic automatic coder has two aspects: clustering the training samples using context free compressed feature mapping, and finding good initial weight parameters, from which the expert automatic coder can be fine tuned. The basic automatic encoder is trained by the original convolution feature map {XJ} MJ = 1, and the batch size is m. XJ is the output of the convolution layer involved in vggnet [3], which is provided by ij from a training image ij randomly selected from a large image database such as Imagenet [28].