Hawkes cumulants - Tensorflow v1 #505

claudio-ICL · 2022-12-14T09:56:10Z

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 · 2022-12-14T16:35:58Z

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 · 2022-12-14T16:50:19Z

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 · 2022-12-14T16:54:04Z

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 · 2022-12-14T16:55:12Z

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

claudio-ICL · 2022-12-14T16:59:02Z

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:

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

Mbompr · 2022-12-14T17:00:20Z

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

claudio-ICL · 2022-12-16T19:25:54Z

tick/hawkes/inference/tests/hawkes_cumulant_matching_test.py

+        np.random.seed(320982)
+
+    @staticmethod
+    def get_simulated_model():


@Mbompr This is the idea that I was suggesting in the conversation of this PR. I simulate a hawkes process with exponential kernels and I train HawkesCumulantMatching on five samples. Loading pickled data would be replaced by this, if you agree.
More importantly, the assertion of almost equality now are between the exact theoretical values from the simulated hawkes process and the calibrated coefficients of HawkesCumulantMatching. When these assertion succeed, they should guarantee approximate exactness of algorithm.

claudio-ICL · 2022-12-16T19:28:45Z

tick/hawkes/inference/tests/hawkes_cumulant_matching_test.py

+            np.allclose(
+                learner.solution,
+                expected_R_pred,
+                atol=0.1,  # TODO: explain why estimation is not so accurate


@Mbompr Unfortunately, accuracy is not so good here. Are you familiar with the tuning of hyperparameters of HawkesCumulantMatching? If yes, can you please see if a better choice of step, max_iter etc... gives more accurate estimations? I can see from the paper that the accuracy was really good.

Hi @Mbompr

Could you look into this please?

claudio-ICL · 2022-12-16T19:34:12Z

lib/include/tick/hawkes/inference/hawkes_cumulant.h

@@ -35,4 +33,29 @@ class DLL_PUBLIC HawkesCumulant : public ModelHawkesList {
  }
 };

+class DLL_PUBLIC HawkesTheoreticalCumulant {


@Mbompr @PhilipDeegan

This is a new c++ class that implements the theoretical formulae for mean intensity, covariance and skewness of hawkes processes. These are the formulae (7), (8), and (9) in the paper.
Currently, the class is only used to test the accuracy of the estimation of HawkesCumulantMatching. However, the formulae have intrinsic values beside such testing, and we might include these calculations in a more general Hawkes class in the future.

claudio-ICL · 2022-12-19T09:38:12Z

tick/hawkes/inference/tests/hawkes_cumulant_matching_test.py


-            np.testing.assert_array_almost_equal(
-                learner.objective(R=learner.solution), 149232.94041039888)
+    @unittest.skipIf(SKIP_TF, "Tensorflow not available")


@PhilipDeegan I have defined a global boolean variable SKIP_TF and used it with unitest.skipIf

claudio-ICL · 2023-01-18T09:41:58Z

tick/hawkes/inference/tests/hawkes_cumulant_matching_test.py

+            np.allclose(
+                learner.solution,
+                expected_R_pred,
+                atol=0.1,  # TODO: explain why estimation is not so accurate


Hi @Mbompr

Could you look into this please?

claudio-ICL · 2023-01-27T17:52:16Z

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

PhilipDeegan · 2023-01-29T20:45:54Z

lib/cpp/hawkes/inference/hawkes_cumulant.cpp

@@ -126,3 +124,66 @@ double HawkesCumulant::compute_E_ijk(ulong r, ulong i, ulong j, ulong k,
  res /= (*end_times)[r];
  return res;
 }
+
+HawkesTheoreticalCumulant::HawkesTheoreticalCumulant(int dim) : d(dim) {
+  Lambda = SArrayDouble::new_ptr(dim);


can these be part of the constuctor setup arguments like : d(dim) ?

PhilipDeegan · 2023-01-29T20:47:23Z

lib/cpp/hawkes/inference/hawkes_cumulant.cpp

@@ -126,3 +124,66 @@ double HawkesCumulant::compute_E_ijk(ulong r, ulong i, ulong j, ulong k,
  res /= (*end_times)[r];
  return res;
 }
+
+HawkesTheoreticalCumulant::HawkesTheoreticalCumulant(int dim) : d(dim) {
+  Lambda = SArrayDouble::new_ptr(dim);


is Lambda the best name?

variables names don't typically start with upper case

The variables Lambda, C, Kc, and R represent first integrated cumulant, second integrated cumulant, (reduced) third integrated cumulant, and inverse of $I - G$ (where $G$ is the matrix of $L^1$-norms of kernel functions), respectively. Their names are directly taken from the math symbols used in the paper: $\Lambda$, $C$, $K$, and $R$, respectively.

I agree that the upper case is uncommon. I can change the name as follows:
Lambda -> first_cumulant
C -> second_cumulant
Kc -> third_cumulant
R -> g_geometric

Done here: 11353f4

PhilipDeegan · 2023-01-29T20:49:15Z

lib/cpp/hawkes/simulation/simu_hawkes.cpp

@@ -56,15 +53,13 @@ bool Hawkes::update_time_shift_(double delay, ArrayDouble &intensity,
 void Hawkes::reset() {
  for (unsigned int i = 0; i < n_nodes; i++) {
    for (unsigned int j = 0; j < n_nodes; j++) {
-      if (kernels[i * n_nodes + j] != nullptr)
-        kernels[i * n_nodes + j]->rewind();
+      if (kernels[i * n_nodes + j] != nullptr) kernels[i * n_nodes + j]->rewind();


I know it was unsigned int before, I would generally prefer std::uint32_t or similar

Understood...
I will open a new PR about that because it is not related to HawkesCumulant or tensorflow

PhilipDeegan · 2023-01-29T20:51:29Z

lib/include/tick/hawkes/inference/hawkes_cumulant.h

+ public:
+  HawkesTheoreticalCumulant(int);
+  int get_dimension() { return d; }
+  void set_baseline(const SArrayDoublePtr mu) { this->mu = mu; }


We might not have it everywhere, but T const[&] n is the preferred order

Changed: fa43f6d

On branch tensorflow-v1-hawkes-cumulants Your branch is up-to-date with 'origin/tensorflow-v1-hawkes-cumulants'. Changes to be committed: modified: lib/cpp/hawkes/inference/hawkes_cumulant.cpp modified: lib/include/tick/hawkes/inference/hawkes_cumulant.h modified: lib/swig/tick/hawkes/inference/hawkes_cumulant.i modified: tick/hawkes/inference/hawkes_cumulant_matching.py

Conflicts: .github/workflows/build_nix.yml tick/hawkes/inference/hawkes_cumulant_matching.py

claudio-ICL · 2023-02-14T08:28:14Z

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

'tensorflow-v1-hawkes-cumulants' into 'pytorch-hawkes-cumulant'

claudio-ICL added 2 commits December 14, 2022 09:38

Hawkes cumulants - Tensorflow v1

f7b17ea

Merge branch 'master' into tensorflow-v1-hawkes-cumulants

e9ce6dd

PhilipDeegan requested a review from Mbompr December 14, 2022 16:31

Mbompr approved these changes Dec 14, 2022

View reviewed changes

claudio-ICL added 2 commits December 14, 2022 17:23

Install tensorflow-cpu

6d6b943

Hawkes Cumulant - theoretical quantities & testing

bee8a54

claudio-ICL commented Dec 16, 2022

View reviewed changes

Use unitest.skipIf when tf not available

55e8a2a

claudio-ICL mentioned this pull request Dec 19, 2022

Hawkes Cumulant Matching - PyTorch implementation #506

Draft

claudio-ICL commented Jan 18, 2023

View reviewed changes

HawkesCumulant - Add test of relative magnitudes

449fa52

PhilipDeegan reviewed Jan 29, 2023

View reviewed changes

claudio-tw and others added 5 commits February 2, 2023 11:13

HawkesTheoreticalCumulant - const specifier in setters

fa43f6d

Merge 'master' into tensorflow-v1-hawkes-cumulants

37844c3

Conflicts: .github/workflows/build_nix.yml tick/hawkes/inference/hawkes_cumulant_matching.py

HawkesCumulantMatching - Use new tensorflow class in example

837331e

Test Hawkes Cumulants - loging

9cc6090

claudio-ICL added a commit to claudio-ICL/tick that referenced this pull request Feb 14, 2023

Merge X-DataInitiative#505 into X-DataInitiative#506

1181645

'tensorflow-v1-hawkes-cumulants' into 'pytorch-hawkes-cumulant'

PhilipDeegan merged commit 4187902 into X-DataInitiative:master Feb 14, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Hawkes cumulants - Tensorflow v1 #505

Hawkes cumulants - Tensorflow v1 #505

claudio-ICL commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

claudio-ICL Dec 16, 2022

claudio-ICL Dec 16, 2022

claudio-ICL Jan 18, 2023

claudio-ICL Dec 16, 2022

claudio-ICL Dec 19, 2022

claudio-ICL Jan 18, 2023

claudio-ICL commented Jan 27, 2023

PhilipDeegan Jan 29, 2023

PhilipDeegan Jan 29, 2023

claudio-tw Feb 2, 2023

claudio-tw Feb 2, 2023

PhilipDeegan Jan 29, 2023

claudio-tw Feb 2, 2023

PhilipDeegan Jan 29, 2023 •

edited

Loading

claudio-tw Feb 2, 2023

claudio-ICL commented Feb 14, 2023

Hawkes cumulants - Tensorflow v1 #505

Hawkes cumulants - Tensorflow v1 #505

Conversation

claudio-ICL commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

claudio-ICL commented Dec 14, 2022

Mbompr commented Dec 14, 2022

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

claudio-ICL commented Jan 27, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

PhilipDeegan Jan 29, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

claudio-ICL commented Feb 14, 2023

PhilipDeegan Jan 29, 2023 •

edited

Loading