摘要：
1. requires_grad = False Set all parameters in the current model frozen: Filter some specific layers by name to be frozen: 2. Filter out unfrozen para 阅读全文
摘要：
Sigmoid Sigmoids saturate and kill gradients. Sigmoid outputs are not zero-centered. Exponential function is a little computational expensive. Tanh Ki 阅读全文
摘要：
Linear Classifier and Hing Loss (or Multiclass SVM Loss) Linear Mapping (Score function) Linear Classifier Hinge Loss Regularization Softmax Classifie 阅读全文