Some problems with training and validation

[zi1zi]

Thanks for the great work, I have the following questions

1. I have trained 30 fonts, some images of the training process are as follows, I don’t think 65w iter is needed, as shown in the figure
   
2. Verification in training feels better than direct verification，The picture above is the verification during training, and the picture below is the direct verification
   
   
   3.The overall effect is not very good, I have the following guesses:
   (1) There are few types of training fonts or the number of characters in each font is small (I use 800)
   (2) The number of training 20w iter is too small
   (3) There is an error in the use of the verification code. I use inference.py for verification.But I don't know what's the difference between the two