Training config. (Training set size?)

[HOMGH]

@tranluan
Hi,
As your main code indicates:

flags.DEFINE_integer("train_size", 5000000, "The size of train images [np.inf]")
flags.DEFINE_integer("epoch", 1000, "Epoch to train [25]")

the training size is 5000000. But in your paper you mentioned that you are using 300W_LP dataset with 122,450 images. I trained your model with 1000 images as training images and 40 epochs. But the results are not reasonable! I'm just wondering is it because of my training size and epoch number? Should I use all 122,450 images as my training set?
And also how many epochs would be enough? Did you use 1000 epoch?
Thank you in advance