Classification accuracy for real data in training is not consistent with post-evaluate

[lixiangideal]

Hi, I'm new on GAN.

When I use the program for my problem with classifier, the class_real_acc index in training process is high like 0.7. I think it means the classification accuracy for the real data in training is 70%, right?

However, after the training, I try to predict the class label for the real data, using similar codes as '' preds = np.argmax(model.predict([x[:10]])[1], -1) '' in xor.py. But the prediction is so wrong and the classification accuracy for the real data is very low. Looks like the model doesn't train at all.

I'm confused, and maybe it's because of the data type? I set all the real data as float64. Am I missing anything? Thanks for your attention!