in fact , The following is what I began to think about halfway

• Error reporting response model nn.linear The parameters do not match and cannot be matmaul, After looking at the model parameters for a long time, there is no problem , Look at the error when taking out the incoming parameters .
• tensor The index says it can only be byte,long,bool, Then start with float Not good either. , Later, I found that it can actually be in autograd.Variable In the process tensor First convert to byte The reconversion of type into Variable—— I'm not sure , It turned out that it was just through direct cancellation variable, hold array Type passed in
• To determine the , The above method doesn't work , say mask[256] And what doesn't match
• “float() argument must be a string or a number, not ‘NoneType’”
  This is what I'm rewriting loss Produced in the process of . Then I tried to turn float The result is wrong , Strong go tensor And I failed . And found that ：NoneType Indicates that there are no parameters here , So my results didn't get in ？？
• Then I found out that actually my model No results ： That's the sentence ：（output Is directly None）
• It turned out to be my model Inside forward Forgot the return value
• Feel this output There's a problem, but I don't know how to solve it , After all, I return an individual value , But the other two values are entered . Look at the follow-up results , If not , Just take out output The third one in tensor. Sure enough ,loss All are nan
• Suddenly found that there were no three tensor？？？ On the other loss All are nan The problem of —— The gradient... Disappeared loss There should be something wrong with the function , Just... The first time -inf 了
  
• Look at the ,x and y It's written backwards mdzz
• Ah, the gradient disappeared in the first training , I don't know why this matrix increases and decreases , I decided to try the normalization operation and distribute it to 0~5
• Find it impossible to , that loss The function is designed to give 0~1 Between x The distribution of the , You can't zoom in . in addition , I found that even if the maximum and minimum normalization , It didn't work .
• Practice and try , This loss It's too big . I know why I'm old ,loss Function copied wrong , No sign , It must be smaller and smaller