1. The image generators batch_size set to sample, and then read the data into memory
2. The batch_size set to 200, when the fit into image generator, the former training quickly, but training about 20 times the acc can only about 0.5 to 1.0 valid_acc,
Every read bacth_size, the latter is the CPU reads data, influence on the speed of training, but training about 200 times the acc to around 0.9, valid_acc can reach more than 0.7,
, in 150 seconds to 9600 seconds,? The reason is this excuse me god, the model is the same, only is not the same as the way of data read
CodePudding user response:
No one no one?CodePudding user response:
Has been solved, the generator data generated every time there will be some bias, training need to read many times