deepmind / spectral_inference_networks Goto Github PK

Inorder to run hydrogen.py I had to make some changes, mainly stuff like -

tf.rand_uniform==> tf.rand.uniform

tf.<some_matrix_operation> ==> tf.compat.v1.<some_matrix_operation>

and at the top of hydrogen.py I added -
tf.compat.v1.disable_eager_execution()

Now it works, but the result I get seems quite different than yours.. It looks more like your beta=1 results :\ Guess some of the changes I made caused that.

Do you have an already updated working version?

Hi, I'm was running spin_test.py to get a basic understanding for how spin works and I noticed that it seems to be consistently finding the smallest magnitude eigenvalues rather than the largest. Is this expected behavior? I thought the paper specified that the goal of spin was to find the top N eigenfunctions? If it is expected behavior, is there a way to reverse the problem formulation to get the largest rather than smallest?

I was running spin_test with random seed 0 with matrices of size 10, and finding 5 eigenfunctions.

Actual eigenvalues:

Calculated eigenvalues:

Hey, we implemented the algorithm in jax here: https://github.com/Binbose/SpIN-Jax

We also made a colab notebook with some extra visualizations like animated training and phase diagrams here

Generally, the algorithm is really cool and runs well. We can recover the first 4 eigenfunctions of hydrogen reliably. However, we noticed that the variance of the energies of our implementation (and also your TensorFlow implementation here, which virtually behaves identically to our implementation) is much bigger compared to the graph in your paper. We are not quite sure how you achieved such stable training, do you remember if you used some additional tricks?
(The variance in the colab notebook is divided by 15 to make the graphs readable)
This is also in accordance with the case of 9 eigenfunctions, for which the training generally behaves a lot less stable than what we can see in the paper and the eigenfunctions look somewhat recognizable but are still far off.

Things we tried:
Reduce the learning rate up to 1e-6
Up the batch size to 512
Increase and decrease beta
Play around with different 'sparsifying' K (we couldn't find the parameter you used in your work)
Tried different decay rates for rmsprop and adam

Can you offer some additional insights?