training UBM with sidekit from custom data

I am trying to train GMM-UBM model from data that i have already extracted for emotion recognition with SIDEKIT(pretty much the same as speaker recognition. I also don't understand the HDF5 feature file system). My data is an ndarray with shape (1101,78) [78 are the number of acoustic features and 1101 the number of feature vectors(frames).

ubm = sidekit.Mixture()

llks = ubm.EM_uniform(anger, distribNb, iteration_min=3, iteration_max=10, llk_gain=0.01, do_init=True)

The error that is thrown is:

line 394, in _compute_all
    self.A = (numpy.square(self.mu) * self.invcov).sum(1) - 2.0 * (numpy.log(self.w) + numpy.log(self.cst))

ValueError: operands could not be broadcast together with shapes (512,78) (512,0)

which means that the covariance matrix is of shape (512,0). Is that wrong? Should it be like (512,78)? I may be wrong. Please give me a hint

回答1:

You might have figured it out already, but I thought I might as well post a possible solution to this.

The following code creates random data with dimensions (2,100) and tries to train a 128-mixture gmm using the EM_uniform algorithm:

import sidekit
import numpy as np
import random as rn

gmm = sidekit.Mixture()
data = np.array([[rn.random() for i in range(100)],[rn.random() for i in range(100)]])
gmm.EM_uniform(data,
               distrib_nb=128,
               iteration_min=3,
               iteration_max=10,
               llk_gain=0.01,
               do_init=True)

However, this results in the same error as you have reported: ValueError: operands could not be broadcast together with shapes (128,100) (128,0)

I suspect there is some bug in how gmm.invcov is calculated in Sidekit.Mixture._init_uniform(), so I have figured out a manual initialization of the mixture with code from Sidekit.Mixture._init() (the initialization function for the EM_split()-algorithm).

The following code ran without errors on my computer:

import sidekit
import numpy as np
import random as rn
import copy

gmm = sidekit.Mixture()
data = np.array([[rn.random() for i in range(100)],[rn.random() for i in range(100)]])

# Initialize the Mixture with code from Sidekit.Mixture._init()
mu = data.mean(0)
cov = (data**2).mean(0)
gmm.mu = mu[None]
gmm.invcov = 1./cov[None]
gmm.w = np.asarray([1.0])
gmm.cst = np.zeros(gmm.w.shape)
gmm.det = np.zeros(gmm.w.shape)
gmm.cov_var_ctl = 1.0 / copy.deepcopy(gmm.invcov)
gmm._compute_all()

# Now run EM without initialization
gmm.EM_uniform(data,
               distrib_nb=128,
               iteration_min=3,
               iteration_max=10,
               llk_gain=0.01,
               do_init=False)

This gave the following output: [-31.419146414931213, 54.759037708692404, 54.759037708692404, 54.759037708692404], which is the log-likelihood values after each iteration (convergence after 4 iterations. Do note that this example data is way to small to train a gmm on.)

I cannot guarantee this leads to any errors later on, leave a comment if that is the case!

As for HDF5-files, check out the the h5py documentation for tutorials. Also, hdfview allows you to look into contents of the h5-files, which is pretty convenient for debugging later on when you get to scoring.

回答2:

What's the content of the parameter 'feature_list' that the sidekit.UBM takes?

ubm=sk.Mixture()
 llk = ubm.EM_split(features_server=server,
                     feature_list=ubm_list,
                     distrib_nb=512,
                     num_thread=8,
                     save_partial=False)