The loss of vgg reaches the ln(1/n) threshold in one round, how to solve it

  1. When dealing with face classification tasks, prepare to use vggnet to implement face classification. The vggface data set is used. First, one hundred categories are classified, and the official vggnet model is used. Only the number of classifications in the last layer and the loss used are changed. The function is the cross-entropy loss function. The training is started, but after several trainings. The result is that the loss value directly reaches the threshold of ln(1/n), and the model will not continue to train.

  2. The second question is, when you use vgg to implement migration learning, what is the effect of using the parameters that have been given by the official? I use migration learning to freeze the previous parameters and modify only the last layer, and find the recognition accuracy in the back The improvement is not very obvious. Finally, it has been over-fitted, but after the test set accuracy rate is 7%

The idea behind transfer learning for vision tasks is to assume that convolution filters are more or less universally applicable to different tasks or different distributions. So you freeze everything related to convolutional part and keep training just fully connected part. VGG has multiple dense layers, you need to let them train too.