Hi David,

I applied your algorithm to my data set and found two issues:

1. I sometimes obtain the warning like this: /usr/lib/python2.7/dist-packages/numpy/lib/function_base.py:2476: RuntimeWarning: Degrees of freedom <= 0 for slice warnings.warn("Degrees of freedom <= 0 for slice", RuntimeWarning) Is it serious warning? should I stop or just ignore it.

2. Oscillation between solutions ('\n\n\nITERATION ###', 15) ('OPTIMIZATION for Cluster #', 0, 'DONE!!!') ('OPTIMIZATION for Cluster #', 2, 'DONE!!!') ('OPTIMIZATION for Cluster #', 3, 'DONE!!!') ('length of the cluster ', 0, '------>', 739) ('length of the cluster ', 1, '------>', 0) ('length of the cluster ', 2, '------>', 20) ('length of the cluster ', 3, '------>', 20) ('length of the cluster ', 4, '------>', 0) UPDATED THE OLD COVARIANCE beginning the smoothening ALGORITHM ('cluster that is zero is:', 1, 'selected cluster instead is:', 0) ('cluster that is zero is:', 4, 'selected cluster instead is:', 3) ('length of cluster #', 0, '-------->', 739) ('length of cluster #', 1, '-------->', 20) ('length of cluster #', 2, '-------->', 0) ('length of cluster #', 3, '-------->', 0) ('length of cluster #', 4, '-------->', 20) Done writing the figure

   ('\n\n\nITERATION ###', 16) ('OPTIMIZATION for Cluster #', 0, 'DONE!!!') ('OPTIMIZATION for Cluster #', 1, 'DONE!!!') ('OPTIMIZATION for Cluster #', 4, 'DONE!!!') ('length of the cluster ', 0, '------>', 739) ('length of the cluster ', 1, '------>', 20) ('length of the cluster ', 2, '------>', 0) ('length of the cluster ', 3, '------>', 0) ('length of the cluster ', 4, '------>', 20) UPDATED THE OLD COVARIANCE beginning the smoothening ALGORITHM ('cluster that is zero is:', 2, 'selected cluster instead is:', 0) ('cluster that is zero is:', 3, 'selected cluster instead is:', 4) ('length of cluster #', 0, '-------->', 739) ('length of cluster #', 1, '-------->', 0) ('length of cluster #', 2, '-------->', 20) ('length of cluster #', 3, '-------->', 20) ('length of cluster #', 4, '-------->', 0) Done writing the figure

   The runs continuously oscillate between two outcomes (739, 0, 20, 20, 0) and (739, 20, 0, 0, 20). Should I stop the run and choose one of them? What happens here and how to avoid this?

Thanks,

Hello, great work on your paper.

Results of the algorithm look promising. Unfortunately, it seems like I'll have to call solve() method with all training dataset to cluster any new data every time. That's not efficient.

Now I have few questions:

1. Is it possible to implement predict method on an already trained model?
2. Could we add load_model and save_model methods?
3. Will you be following scikit-learn interfaces in the future (__init__,fit, predict, etc)?

I could definitely contribute but after 2 hours of attempting to refactor the code, it seems like there are plenty of dependent variables and super long methods. How can I help?

Hi David,

Can we use TICC for online clustering time series? For instance, we want to identify the state the car is in during driving, given a set of learned states using TICC for batch learning.

Thanks,

I was following the readme.md and when I was executing this (cluster_assignment, cluster_MRFs) = TICC_solver.solve(window_size = 10,number_of_clusters = 5, lambda_parameter = 11e-2, beta = 400, maxIters = 70, threshold = 2e-5, write_out_file = True, input_file = "data.csv", prefix_string = "output_folder/") it throws me error: IndexError: only integers, slices (:), ellipsis (...), numpy.newaxis (None) and integer or boolean arrays are valid indices

I am using anaconda python 2.7 with cvxpy and snap just installed today.

Please give the name of paper or url, Thanks!

BTW: windows （64bit）after run generate_synthetic_data.py, run TICC.py, find this in output:

RuntimeError: 
            Attempt to start a new process before the current process
            has finished its bootstrapping phase.

            This probably means that you are on Windows and you have
            forgotten to use the proper idiom in the main module:

                if __name__ == '__main__':
                    freeze_support()

Hello! because I am trying to test my data in your example.py but I got an error: numpy.linalg.linalg.LinAlgError: Eigenvalues did not converge Do you know what's wrong with this? Thank you very much

Hi Could you please help me with this. As I tried to run example, I constantly received similar result: it printed ITERATION ### 0 and did not proceed. I will be beyond grateful if you could help me figuring out what is going wrong.

Hi,

I want to ask about how do yo generate synthetic data based on previous samples? In your paper, you mentioned that "The data is then drawn one sample at a time, conditioned on the values of the previous w − 1 samples.". I know you generate w samples (n * w array) using "numpy.random.multinomial", but how do you generate w samples based on previous w − 1 samples? It means how do I generate 1 sample given w - 1 samples?

Thank you for your time.

Hi, Could you please help me with this.

I have a problem running your code on my dataset. When I'm in Iteration 0 and calling the function "train_clusters", the return value "optRes" seems to contain some nan. So when the function "optimize_clusters" got the value "optRes" from the function "train_clusters", it raised the error.

I did print out the value of "val" every time. Here is the output and the error information: ################################################## ITERATION ### 0 /usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:497: RuntimeWarning: Degrees of freedom <= 0 for slice /usr/local/lib/python3.5/dist-packages/numpy/lib/function_base.py:3093: RuntimeWarning: divide by zero encountered in double_scalars c *= 1. / np.float64(fact) /usr/local/lib/python3.5/dist-packages/numpy/lib/function_base.py:3093: RuntimeWarning: invalid value encountered in multiply c *= 1. / np.float64(fact) val: [0.2301 0.1450 -0.0000 ..., 0.9513 -0.0000 0.3904] OPTIMIZATION for Cluster # 0 DONE!!! val: [0.7449 0.0000 -0.0000 ..., 1.0000 -0.0000 0.5509] OPTIMIZATION for Cluster # 1 DONE!!! val: [0.9931 0.0000 0.0000 ..., 0.7676 -0.0000 0.5487] OPTIMIZATION for Cluster # 2 DONE!!! val: [0.4735 0.0000 0.0000 ..., 0.7480 0.0000 0.3804] OPTIMIZATION for Cluster # 3 DONE!!! val: [0.1999 0.1088 0.0000 ..., 0.8651 0.0000 0.5592] OPTIMIZATION for Cluster # 4 DONE!!! val: [0.0202 0.0314 0.0143 ..., 0.7245 0.0000 0.1511] OPTIMIZATION for Cluster # 5 DONE!!! val: [nan nan nan ..., nan nan nan] OPTIMIZATION for Cluster # 6 DONE!!! /usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:567: RuntimeWarning: invalid value encountered in less /usr/local/lib/python3.5/dist-packages/ipykernel_launcher.py:567: RuntimeWarning: invalid value encountered in greater

LinAlgErrorTraceback (most recent call last) in fit(self, input_file) 274 #print("opt_res: " + str(opt_res)) 275 self.optimize_clusters(computed_covariance, len_train_clusters, log_det_values, opt_res, --> 276 train_cluster_inverse) 277 278 # update old computed covariance

in optimize_clusters(self, computed_covariance, len_train_clusters, log_det_values, optRes, train_cluster_inverse) 467 X2 = S_est 468 #print(S_est) --> 469 u, _ = np.linalg.eig(S_est) 470 cov_out = np.linalg.inv(X2) 471

/usr/local/lib/python3.5/dist-packages/numpy/linalg/linalg.py in eig(a) 1126 _assertRankAtLeast2(a) 1127 _assertNdSquareness(a) -> 1128 _assertFinite(a) 1129 t, result_t = _commonType(a) 1130

/usr/local/lib/python3.5/dist-packages/numpy/linalg/linalg.py in _assertFinite(*arrays) 215 for a in arrays: 216 if not (isfinite(a).all()): --> 217 raise LinAlgError("Array must not contain infs or NaNs") 218 219 def _isEmpty2d(arr):

LinAlgError: Array must not contain infs or NaNs ##################################################

I assume that there are some problems when the program calculate (use) the ADMM_solver. Do you know what's wrong with this? Thank you for your time.

Btw, I did normalize my dataset and set lambda to 0.1, 1, ...

I have a had a look at your updated code. Interestingly enough, when running the two tests cases in TestStringMethods together, they both pass.

However, when running TestStringMethods.test_multiExample, it fails. I'm not sure why exactly. Do you have an explanation for this?

I ran that using Mac Pro machine, Python 3.6.4 (miniconda). Which machine and Python version did you use?

When test case passes, this is a snippet of the output

`# PASSED`
lam_sparse 0.11
switch_penalty 600
num_cluster 5
num stacked 5
completed getting the data



ITERATION ### 0
OPTIMIZATION for Cluster # 0 DONE!!!
OPTIMIZATION for Cluster # 1 DONE!!!
OPTIMIZATION for Cluster # 2 DONE!!!
OPTIMIZATION for Cluster # 3 DONE!!!
OPTIMIZATION for Cluster # 4 DONE!!!
length of the cluster  0 ------> 2851.       ` # <---- these numbers are different than the failed one.`
length of the cluster  1 ------> 2698
length of the cluster  2 ------> 2094
length of the cluster  3 ------> 7522
length of the cluster  4 ------> 4438
UPDATED THE OLD COVARIANCE
beginning the smoothening ALGORITHM
length of cluster # 0 --------> 2131
length of cluster # 1 --------> 1886
length of cluster # 2 --------> 3072
length of cluster # 3 --------> 7915
length of cluster # 4 --------> 4599
Done writing the figure

And that is a snippet when it fails

 `# FAILED`
lam_sparse 0.11
switch_penalty 600
num_cluster 5
num stacked 5
completed getting the data



ITERATION ### 0
OPTIMIZATION for Cluster # 0 DONE!!!
OPTIMIZATION for Cluster # 1 DONE!!!
OPTIMIZATION for Cluster # 2 DONE!!!
OPTIMIZATION for Cluster # 3 DONE!!!
OPTIMIZATION for Cluster # 4 DONE!!!
length of the cluster  0 ------> 7851.   `# <-----  really higher than when it passed`
length of the cluster  1 ------> 2246
length of the cluster  2 ------> 2261
length of the cluster  3 ------> 2613
length of the cluster  4 ------> 4632
UPDATED THE OLD COVARIANCE
beginning the smoothening ALGORITHM
length of cluster # 0 --------> 7957
length of cluster # 1 --------> 1617
length of cluster # 2 --------> 3102
length of cluster # 3 --------> 3280
length of cluster # 4 --------> 3647
Done writing the figure

The failed test case has successfully convergred, but still fails

CONVERGED!!! BREAKING EARLY!!!



TRAINING F1 score: -1 -1 -1


Failure
Traceback (most recent call last):
  File "/Users/motaher/miniconda2/envs/pyliger/lib/python3.6/unittest/case.py", line 59, in testPartExecutor
    yield
  File "/Users/motaher/miniconda2/envs/pyliger/lib/python3.6/unittest/case.py", line 605, in run
    testMethod()
  File "/Users/motaher/dev/PycharmProjects/ORIGINAL_TICC/UnitTest.py", line 33, in test_multiExample
    self.assertEqual(sum(val), 0)
  File "/Users/motaher/miniconda2/envs/pyliger/lib/python3.6/unittest/case.py", line 829, in assertEqual
    assertion_func(first, second, msg=msg)
  File "/Users/motaher/miniconda2/envs/pyliger/lib/python3.6/unittest/case.py", line 822, in _baseAssertEqual
    raise self.failureException(msg)
AssertionError: 45059.0 != 0


Ran 1 test in 53.887s

FAILED (failures=1)

That could be due to the fact that states in stored file and states coming from the algorithms were in different values?

Switch TICC solver to python 3

The mammoth changes in TICC_helper are just converting all the indentation to spaces instead of a mix of tabs and spaces (which python 3 will break on)

For some reason this change breaks unit tests, but only slightly? The assignments stay the same, but the cluster MRFs are ever so slightly different: 5.9496 vs 5.9497 on the second diagonal entry (which shouldn't matter anyway?).

how to get betweenness centrality for each sensor based on MRF network, I’m trying to use networkx which is a Python library for studying graphs and networks to calculate betweenness centrality,but I’m confused about how to use it based on MRF network,at the same time, I couldn’t find the part in your code(https://github.com/davidhallac/TICC),could you be so kind as to give me the code about how to get betweenness centrality for each sensor based on MRF network?

Hi all,

I had a question regarding the smoothen_clusters function:

If I understand well, this function compute the LLE for all points. Hower, the follwoing condition avoid to compute the LLE of the last points... Any particular reason for that ? Since the vector LLE_all_points_clusters is of size [clustered_points_len, self.number_of_clusters], the last point have a LLE fixed to zero... which semms strange!

if point + self.window_size - 1 < complete_D_train.shape[0]:

Thanks in advance, Flo

It's not clear to me how to use TICC after training on test data.

Say I have a dataset X with 100 samples and I want to train on the first 50 and test it on the remaining 50. Could you elaborate on how to do that?

So basically when you have new points how are those assigned to the cluster? From what I understand I need to calculate the LLE matrix of dimension n_samples x n _clusters and then run the updateClusters function which is the dynamic programming algorithm that, given the trained inverse covariances for each cluster, works out the optimal cluster assignment for each point.

My question is do I need to concatenate the trained LLE with the new LLE deriving from the test data points? Because the results will change depending on if I do that or not.

Many thanks Gio

I noticed that the TICC_solver.py and the TICC.py in the paper_code folder have two different implementation of the optimization for the Graphical Lasso.

For TICC_solver.py you use a custom class called ADMMSolver and the process seems pretty straightforward whereas in the TICC.py you use a class called TGraphV which also uses the ADMMSolver but it's a much more complicated implementation and brings additional dependencies like cvxpy and snap

I was wondering if there is any major difference between the two or anything else the user should be aware of

Hello and thank you for the wonderful paper I noticed that the two implementation are quite different.

In the paper_code the implementation seems more complicated and it uses the SolverCrossTime and cvxpy which are not used in TICC_solver.py (as it only uses ADMM_solver.

Would please clarify on this aspect?

Many thanks Gio