Hawkes cumulants - Tensorflow v1

[claudio-ICL]

This PR proposes to break the class HawkesCumulantMatching in a base class with the general interface, and a derived class that implements the methods of the interface using tensorflow. A second derived class that will implement the same methods with pytorch is also prototyped. Moreover, this PR tries to migrate HawkesCumulantMatching to tensorflow v2 minimising the required changes: I used tf.compat.v1. Tests of the inference methodology are re-activated (no longer skipped by default).

[claudio-ICL]

Hi @Mbompr

I am trying to adapt the code of HawkesCumulantMatching to tensorflow v2. There are some tests that fail however, because the numerical precision is not achieved. I was wondering whether you can explain how the expected values were produced?

[Mbompr]

If I remember correctly, these values have been obtained by making the algorithm run. I think that if the new values are close enough, we can simply replace them. Maybe the optimization loop has changed with TF2.

[claudio-ICL]

If I remember correctly, these values have been obtained by making the algorithm run. I think that if the new values are close enough, we can simply replace them. Maybe the optimization loop has changed with TF2.

Understood. So, they are not ground truth. From a first run with v2 I get somewhat similar values, but the precision of 6 decimal places is certainly lost.

[Mbompr]

I think it makes sense to change the values in the test within the PR then

[claudio-ICL]

If I remember correctly, these values have been obtained by making the algorithm run. I think that if the new values are close enough, we can simply replace them. Maybe the optimization loop has changed with TF2.

Understood. So, they are not ground truth. From a first run with v2 I get somewhat similar values, but the precision of 6 decimal places is certainly lost.

I am thinking whether it would be more appropriate to design a test as follows:

1. Instantiate a parametric class (e.g. exponential) with given coefficients, and compute exact L1 norms of kernels.
2. Generate timestamps from simulation.
3. Calibrate HawkesCumulantMatching on the simulated timestamps.
4. Compare calibrated adjacency with exact L1 norms.

[Mbompr]

I agree that makes more sense. But I find it hard to suggest something big when I had been lazy before 😬

[claudio-ICL]

Hi @PhilipDeegan , Do you think we can merge this?

[claudio-ICL]

@PhilipDeegan I believe that this PR is ready to merge. Would it be possible to do it today?