Multi-Dimensional Feature Space

[Hongbo-UOM]

Multi-Dimensional Feature Space

[poypoyan]

@zhongbo-THU As of writing this, only one-dimensional data is supported, but I am now working on this. I will create a new branch once I make it work. Thanks!

[Hongbo-UOM]

@zhongbo-THU As of writing this, only one-dimensional data is supported, but I am now working on this. I will create a new branch once I make it work. Thanks!

Thank you very much for your reply and your open source code, very much look forward to your project to support multi-dimensional data operation, which will provide a more perfect reference for developers. I don't know if it's convenient for you to provide a contact or email address, which makes it easier to communicate with each other.

[poypoyan]

@zhongbo-THU Check out the multi-dim branch. The emissions are multivariate normal distribution. For this version:

• The observation must have shape of (num_of_samples, num_of_dim).
• The <GaussianHSMM object>.mean must have shape of (num_of_states, num_of_dim).
• The <GaussianHSMM object>.covmat must have shape of (num_of_states, num_of_dim, num_of_dim) (in other words, array of covariance matrices).

If there is no manual setting for mean and covmat, the default mean (for now) is just 'zeros' with shape as above (in the future, use K-means), and the default covmat is array of identity matrices per state.

The 'tutorial' in this branch is still a test of one-dimensional data. Actual testing of multi-dimensional data is appreciated.

You can contact me at [email protected]. Thanks!

[poypoyan]

Update: multi-dim branch is now merged to master, because I already did a preliminary "test" for 2-dimensional observations (see Notebooks).

I will now close this issue, but comments are, of course, still welcome. Thanks! 😄

[Hongbo-UOM]

very thanks～