Skip to content
Snippets Groups Projects
Commit 31ae6f70 authored by Yannick DAYER's avatar Yannick DAYER
Browse files

IVector test for machine with update_sigma

parent d3c33d84
No related branches found
No related tags found
1 merge request!60Port of I-Vector to python
Hide whitespace changes
Inline Side-by-side
bob/learn/em/ivector.py
+ 31
− 37
View file @ 31ae6f70
......@@ -186,27 +186,6 @@ def compute_tt_sigma_inv_fnorm(
return output
def forward(
ubm_means: np.ndarray, stats: GMMStats, T: np.ndarray, sigma: np.ndarray
):
# void bob::learn::em::IVectorMachine::forward_(const bob::learn::em::GMMStats& gs,
# blitz::Array<double,1>& ivector) const
# {
# // Computes \f$(Id + \sum_{c=1}^{C} N_{i,j,c} T^{T} \Sigma_{c}^{-1} T)\f$
# computeIdTtSigmaInvT(gs, m_tmp_tt);
# // Computes \f$T^{T} \Sigma^{-1} \sum_{c=1}^{C} (F_c - N_c ubmmean_{c})\f$
# computeTtSigmaInvFnorm(gs, m_tmp_t1);
# // Solves m_tmp_tt.ivector = m_tmp_t1
# bob::math::linsolve(m_tmp_tt, m_tmp_t1, ivector);
# }
return np.linalg.solve(
compute_id_tt_sigma_inv_t(stats, T, sigma),
compute_tt_sigma_inv_fnorm(ubm_means, stats, T, sigma),
)
class IVectorMachine(BaseEstimator):
"""Trains and projects data using I-Vector.
......@@ -397,26 +376,39 @@ class IVectorMachine(BaseEstimator):
logger.info(f"Did not converge after {step+1} steps.")
return self
def project(self, stats: GMMStats) -> IVectorStats:
if not isinstance(stats, GMMStats):
return [self.project(s) for s in stats] # TODO yes?
return forward(self.ubm.means, stats, self.T, self.sigma)
def project(self, stats: GMMStats) -> np.ndarray:
"""Projects the GMMStats on the IVectorMachine.
This takes data already projected onto the UBM.
**Returns:**
The IVector of the input stats.
"""
return np.linalg.solve(
compute_id_tt_sigma_inv_t(stats, self.T, self.sigma),
compute_tt_sigma_inv_fnorm(
self.ubm.means, stats, self.T, self.sigma
),
)
def transform(self, data: np.ndarray) -> List[IVectorStats]:
def transform(self, data: List[np.ndarray]) -> List[np.ndarray]:
"""Transforms the data using the trained IVectorMachine.
Parameters
----------
data : np.ndarray
The data to transform.
This takes MFCC data, will project them onto the ubm, and compute the IVector
statistics.
Returns
-------
np.ndarray
The transformed data.
**Parameters:**
data
The data (MFCC features) to transform. Arrays of shape (n_samples, n_features).
**Returns:**
The IVector for each sample. Shape: (dim_t,)
"""
stats = self.project(self.ubm.acc_stats(data))
return stats.t
return [self.project(self.ubm.acc_stats(d)) for d in data]
def enroll(self, stats: List[GMMStats]) -> IVectorStats:
"""Enrolls a new speaker.
......@@ -433,7 +425,9 @@ class IVectorMachine(BaseEstimator):
"""
return self.project(stats)
def score(self, model: IVectorStats, probes: List[GMMStats]) -> List[float]:
def score(
self, model: IVectorStats, probes: List[np.ndarray]
) -> List[float]:
return linear_scoring(
model,
ubm=self.ubm,
......
bob/learn/em/test/test_ivector.py
+ 113
− 124
View file @ 31ae6f70
......@@ -2,8 +2,9 @@
# @author: Yannick Dayer <yannick.dayer@idiap.ch>
# @date: Fri 06 May 2022 12:59:21 UTC+02
import copy
import numpy as np
import pytest
from bob.learn.em import GMMMachine, GMMStats, IVectorMachine
......@@ -58,29 +59,24 @@ def test_ivector_machine_projection():
def test_ivector_machine_transformer():
dim_t = 2
ubm = GMMMachine(n_gaussians=2)
ubm.means = np.array([[1, 7, 4], [4, 5, 3]], dtype=float)
ubm.variances = np.array([[0.5, 1.0, 1.5], [1.0, 1.5, 2.0]], dtype=float)
machine = IVectorMachine(ubm=ubm, dim_t=2)
machine = IVectorMachine(ubm=ubm, dim_t=dim_t)
machine.T = np.array(
[[[1, 2], [4, 1], [0, 3]], [[5, 8], [7, 10], [11, 1]]], dtype=float
)
assert hasattr(machine, "fit")
assert hasattr(machine, "transform")
assert hasattr(machine, "enroll")
assert hasattr(machine, "score")
transformed = machine.transform([np.ndarray([1, 2, 3])])[0]
assert isinstance(transformed, IVectorMachine)
nij_sigma_wij2_ref = np.array([[0.5, 1.0, 1.5], [1.0, 1.5, 2.0]]) # TODO
nij_ref = np.array([[1, 2, 3], [4, 5, 6]])
fnorm_sigma_wij_ref = np.array([[0.5, 1.0, 1.5], [1.0, 1.5, 2.0]]) # TODO
snormij_ref = np.array([[1, 2, 3], [4, 5, 6]])
np.testing.assert_almost_equal(
transformed.nij_sigma_wij2, nij_sigma_wij2_ref
)
np.testing.assert_almost_equal(transformed.nij, nij_ref)
np.testing.assert_almost_equal(
transformed.fnorm_sigma_wij, fnorm_sigma_wij_ref
transformed = machine.transform([np.array([1, 2, 3])])[0]
assert isinstance(transformed, np.ndarray)
np.testing.assert_array_equal(
transformed, np.array([-0.04213415, 0.21463343])
)
np.testing.assert_almost_equal(transformed.snormij, snormij_ref)
def test_ivector_machine_training():
......@@ -324,6 +320,7 @@ def test_trainer_nosigma():
data = [gs1, gs2]
# Reference values TODO: load from hdf5 file
acc_Nij_Sigma_wij2_ref1 = np.array(
[
[[0.03202305, -0.02947769], [-0.02947769, 0.0561132]],
......@@ -411,7 +408,7 @@ def test_trainer_nosigma():
assert acc_Nij_Sigma_wij2_ref1.shape == (2, 2, 2)
assert acc_Nij_Sigma_wij2_ref2.shape == (2, 2, 2)
# Python implementation
# Reference implementation
# Machine
m = IVectorMachine(ubm, dim_t=2)
t = np.array([[[1.0, 2], [4, 1], [0, 3]], [[5, 8], [7, 10], [11, 1]]])
......@@ -444,16 +441,15 @@ def test_trainer_nosigma():
# New implementation
# Machine
m = IVectorMachine(ubm, dim_t=2)
m = IVectorMachine(ubm, dim_t=2, update_sigma=False)
t = np.array([[[1.0, 2], [4, 1], [0, 3]], [[5, 8], [7, 10], [11, 1]]])
sigma = np.array([[1.0, 2.0, 1.0], [3.0, 2.0, 4.0]])
# Initialization
m.T = t
m.sigma = sigma
m.sigma = copy.deepcopy(sigma)
stats = None
for it in range(2):
print("Iteration:", it)
# E-Step
stats = m.e_step(data)
np.testing.assert_almost_equal(
......@@ -470,10 +466,7 @@ def test_trainer_nosigma():
# M-Step
m.m_step(stats)
np.testing.assert_almost_equal(t_ref[it], m.T, decimal=5)
np.testing.assert_equal(sigma, m.sigma) # sigma is not updated
# testing exceptions
pytest.raises(RuntimeError, m.e_step, [1, 2, 2])
np.testing.assert_equal(sigma, m.sigma) # sigma should not be updated
def test_trainer_update_sigma():
......@@ -512,90 +505,90 @@ def test_trainer_update_sigma():
data = [gs1, gs2]
# Reference values
acc_Nij_Sigma_wij2_ref1 = {
0: np.array([[0.03202305, -0.02947769], [-0.02947769, 0.0561132]]),
1: np.array([[0.07953279, -0.07829414], [-0.07829414, 0.13814242]]),
}
acc_Fnorm_Sigma_wij_ref1 = {
0: np.array(
# Reference values TODO: load from hdf5 file
acc_Nij_Sigma_wij2_ref1 = np.array(
[
[[0.03202305, -0.02947769], [-0.02947769, 0.0561132]],
[[0.07953279, -0.07829414], [-0.07829414, 0.13814242]],
]
)
acc_Fnorm_Sigma_wij_ref1 = np.array(
[
[
[-0.29622691, 0.61411796],
[0.09391764, -0.27955961],
[-0.39014455, 0.89367757],
]
),
1: np.array(
],
[
[0.04695882, -0.13977981],
[-0.05718673, 0.24159665],
[-0.17098161, 0.47326585],
]
),
}
acc_Snorm_ref1 = np.array([16.6, 22.4, 16.6, 61.4, 55.0, 97.4])
],
]
)
acc_Snorm_ref1 = np.array([[16.6, 22.4, 16.6], [61.4, 55.0, 97.4]])
N_ref1 = np.array([0.6, 1.4])
t_ref1 = np.array(
[
[1.59543739, 11.78239235],
[-3.20130371, -6.66379081],
[4.79674111, 18.44618316],
[-0.91765407, -1.5319461],
[2.26805901, 3.03434944],
[2.76600031, 4.9935962],
[
[1.59543739, 11.78239235],
[-3.20130371, -6.66379081],
[4.79674111, 18.44618316],
],
[
[-0.91765407, -1.5319461],
[2.26805901, 3.03434944],
[2.76600031, 4.9935962],
],
]
)
sigma_ref1 = np.array(
[
16.39472121,
34.72955353,
3.3108037,
43.73496916,
38.85472445,
68.22116903,
[16.39472121, 34.72955353, 3.3108037],
[43.73496916, 38.85472445, 68.22116903],
]
)
acc_Nij_Sigma_wij2_ref2 = {
0: np.array([[0.50807426, -0.11907756], [-0.11907756, 0.12336544]]),
1: np.array([[1.18602399, -0.2835859], [-0.2835859, 0.39440498]]),
}
acc_Fnorm_Sigma_wij_ref2 = {
0: np.array(
acc_Nij_Sigma_wij2_ref2 = np.array(
[
[[0.50807426, -0.11907756], [-0.11907756, 0.12336544]],
[[1.18602399, -0.2835859], [-0.2835859, 0.39440498]],
]
)
acc_Fnorm_Sigma_wij_ref2 = np.array(
[
[
[0.07221453, 1.1189786],
[-0.08681275, -0.35396112],
[0.15902728, 1.47293972],
]
),
1: np.array(
],
[
[-0.04340637, -0.17698056],
[0.10662127, 0.21484933],
[0.13116645, 0.64474271],
]
),
}
acc_Snorm_ref2 = np.array([16.6, 22.4, 16.6, 61.4, 55.0, 97.4])
],
]
)
acc_Snorm_ref2 = np.array([[16.6, 22.4, 16.6], [61.4, 55.0, 97.4]])
N_ref2 = np.array([0.6, 1.4])
t_ref2 = np.array(
[
[2.93105054, 11.89961223],
[-1.08988119, -3.92120757],
[4.02093173, 15.82081981],
[-0.17376634, -0.57366984],
[0.26585634, 0.73589952],
[0.60557877, 2.07014704],
[
[2.93105054, 11.89961223],
[-1.08988119, -3.92120757],
[4.02093173, 15.82081981],
],
[
[-0.17376634, -0.57366984],
[0.26585634, 0.73589952],
[0.60557877, 2.07014704],
],
]
)
sigma_ref2 = np.array(
[
5.12154025e00,
3.48623823e01,
1.00000000e-05,
4.37792350e01,
3.91525332e01,
6.85613258e01,
[5.12154025e00, 3.48623823e01, 1.00000000e-05],
[4.37792350e01, 3.91525332e01, 6.85613258e01],
]
)
......@@ -609,70 +602,66 @@ def test_trainer_update_sigma():
t_ref = [t_ref1, t_ref2]
sigma_ref = [sigma_ref1, sigma_ref2]
# Python implementation
# Reference implementation
# Machine
m = IVectorMachine(ubm, 2)
t = np.array([[[1.0, 2], [4, 1], [0, 3]], [[5, 8], [7, 10], [11, 1]]])
sigma = np.array([1.0, 2.0, 1.0, 3.0, 2.0, 4.0])
sigma = np.array([[1.0, 2.0, 1.0], [3.0, 2.0, 4.0]])
# Initialization
trainer = IVectorTrainerPy(sigma_update=True)
trainer.initialize(m, data)
m.t = t
m.T = t
m.sigma = sigma
for it in range(2):
# E-Step
trainer.e_step(m, data)
for k in acc_Nij_Sigma_wij2_ref[it]:
assert np.allclose(
acc_Nij_Sigma_wij2_ref[it][k],
trainer.m_acc_Nij_Sigma_wij2[k],
1e-5,
)
for k in acc_Fnorm_Sigma_wij_ref[it]:
assert np.allclose(
acc_Fnorm_Sigma_wij_ref[it][k],
trainer.m_acc_Fnorm_Sigma_wij[k],
1e-5,
)
assert np.allclose(acc_Snorm_ref[it], trainer.m_acc_Snorm, 1e-5)
assert np.allclose(N_ref[it], trainer.m_N, 1e-5)
np.testing.assert_almost_equal(
acc_Nij_Sigma_wij2_ref[it], trainer.m_acc_Nij_Sigma_wij2, decimal=5
)
np.testing.assert_almost_equal(
acc_Fnorm_Sigma_wij_ref[it],
trainer.m_acc_Fnorm_Sigma_wij,
decimal=5,
)
np.testing.assert_almost_equal(
acc_Snorm_ref[it], trainer.m_acc_Snorm, decimal=5
)
np.testing.assert_almost_equal(N_ref[it], trainer.m_N, decimal=5)
# M-Step
trainer.m_step(m, data)
assert np.allclose(t_ref[it], m.t, 1e-5)
assert np.allclose(sigma_ref[it], m.sigma, 1e-5)
trainer.m_step(m)
np.testing.assert_almost_equal(t_ref[it], m.T, decimal=5)
np.testing.assert_almost_equal(sigma_ref[it], m.sigma, decimal=5)
# C++ implementation TODO
# New implementation
# Machine
m = IVectorMachine(ubm, 2)
m.variance_threshold = 1e-5
m = IVectorMachine(
ubm, dim_t=2, variance_floor=1e-5
) # update_sigma is True by default
# Initialization
# trainer = IVectorTrainer(update_sigma=True)
# trainer.initialize(m)
# m.t = t
# m.sigma = sigma
# for it in range(2):
# # E-Step
# trainer.e_step(m, data)
# for k in acc_Nij_Sigma_wij2_ref[it]:
# assert np.allclose(
# acc_Nij_Sigma_wij2_ref[it][k], trainer.acc_nij_wij2[k], 1e-5
# )
# for k in acc_Fnorm_Sigma_wij_ref[it]:
# assert np.allclose(
# acc_Fnorm_Sigma_wij_ref[it][k], trainer.acc_fnormij_wij[k], 1e-5
# )
# assert np.allclose(
# acc_Snorm_ref[it].reshape(dim_c, dim_d), trainer.acc_snormij, 1e-5
# )
# assert np.allclose(N_ref[it], trainer.acc_nij, 1e-5)
# # M-Step
# trainer.m_step(m)
# assert np.allclose(t_ref[it], m.t, 1e-5)
# assert np.allclose(sigma_ref[it], m.sigma, 1e-5)
# Manual Initialization
m.T = t
m.sigma = sigma
for it in range(2):
# E-Step
stats = m.e_step(data)
np.testing.assert_almost_equal(
acc_Nij_Sigma_wij2_ref[it], stats.nij_sigma_wij2, decimal=5
)
np.testing.assert_almost_equal(
acc_Fnorm_Sigma_wij_ref[it], stats.fnorm_sigma_wij, decimal=5
)
np.testing.assert_almost_equal(
acc_Snorm_ref[it], stats.snormij, decimal=5
)
np.testing.assert_almost_equal(N_ref[it], stats.nij, decimal=5)
# M-Step
m.m_step(stats)
np.testing.assert_almost_equal(t_ref[it], m.T, decimal=5)
np.testing.assert_almost_equal(sigma_ref[it], m.sigma, decimal=5)
def test_trainer_update_sigma_parallel():
......@@ -814,8 +803,8 @@ def test_trainer_update_sigma_parallel():
# C++ implementation TODO
# Machine
serial_m = IVectorMachine(ubm, 2)
serial_m.variance_threshold = 1e-5
# serial_m = IVectorMachine(ubm, 2)
# serial_m.variance_threshold = 1e-5
# SERIAL TRAINER
# serial_trainer = IVectorTrainer(update_sigma=True)
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment