some relative details about implementation

[mr01101001]

1. What is the version of tensorflow you use?
2. When iteration reached to 10000, the loss value just converge to 2e-3 and relative L2 errors just converge to the value between 1~2, which is great different with the performance in your paper.
3. What is the specific parameters set in order to get your results in cylinder2D example, such like learning rate, epochs, batchsize etc?

[spawner9]

1. I have tried the version tensorflow1.12,and it does work.
2. I have got the same issue when iteration reached to about 10000, but if you run it for more time such as 12 hours (I tried), the result is similar with the result in paper.

[saoz233]

The gradient may be too small. The gradient of Velocity is always 10^-5, so maybe the model need more time (40hours used in the origin model).