def test_meanvar_incremental():
np.random.seed(1234)
N = 32
X = np.random.randn(N, 100, 24)
lengths = [len(x) for x in X]
X_mean = np.mean(X, axis=(0, 1))
X_var = np.var(X, axis=(0, 1))
X_std = np.sqrt(X_var)
# Check consistency with numpy
X_mean_inc, X_var_inc = P.meanvar(X)
assert np.allclose(X_mean, X_mean_inc)
assert np.allclose(X_var, X_var_inc)
# Split dataset and compute meanvar incrementaly
X_a = X[:N // 2]
X_b = X[N // 2:]
X_mean_a, X_var_a, last_sample_count = P.meanvar(
X_a, return_last_sample_count=True)
assert last_sample_count == np.sum(lengths[:N // 2])
X_mean_b, X_var_b = P.meanvar(
X_b, mean_=X_mean_a, var_=X_var_a,
last_sample_count=last_sample_count)
assert np.allclose(X_mean, X_mean_b)
assert np.allclose(X_var, X_var_b)
# meanstd
X_mean_a, X_std_a, last_sample_count = P.meanstd(
X_a, return_last_sample_count=True)
assert last_sample_count == np.sum(lengths[:N // 2])
X_mean_b, X_std_b = P.meanstd(
X_b, mean_=X_mean_a, var_=X_std_a**2,
last_sample_count=last_sample_count)
assert np.allclose(X_mean, X_mean_b)
assert np.allclose(X_std, X_std_b)
def __init__(self,input_file_source,output_file_source, mfcc_only, num_frames=128, norm_calc=True):
self.input_file_source = input_file_source
self.output_file_source = output_file_source
self.num_frames = num_frames
self.mfcc_only = mfcc_only
self.input_meanstd = None
self.output_meanstd = None
if norm_calc:
print("Performing speaker 1 normalization...")
SF1_lengths = [len(y) for y in self.input_file_source]
self.input_meanstd = meanstd(self.input_file_source, SF1_lengths)
print("Performing speaker 2 normalization...")
TF2_lengths = [len(y) for y in self.output_file_source]
self.output_meanstd = meanstd(self.output_file_source, TF2_lengths)
def test_meanvar():
# Pick acoustic features for testing
_, X = example_file_data_sources_for_acoustic_model()
X = FileSourceDataset(X)
lengths = [len(x) for x in X]
D = X[0].shape[-1]
X_mean, X_var = P.meanvar(X)
X_std = np.sqrt(X_var)
assert np.isfinite(X_mean).all()
assert np.isfinite(X_var).all()
assert X_mean.shape[-1] == D
assert X_var.shape[-1] == D
_, X_std_hat = P.meanstd(X)
assert np.allclose(X_std, X_std_hat)
x = X[0]
x_scaled = P.scale(x, X_mean, X_std)
assert np.isfinite(x_scaled).all()
# For padded dataset
_, X = example_file_data_sources_for_acoustic_model()
X = PaddedFileSourceDataset(X, 1000)
# Should get same results with padded features
X_mean_hat, X_var_hat = P.meanvar(X, lengths)
assert np.allclose(X_mean, X_mean_hat)
assert np.allclose(X_var, X_var_hat)
# Inverse transform
x = X[0]
x_hat = P.inv_scale(P.scale(x, X_mean, X_std), X_mean, X_std)
assert np.allclose(x, x_hat, atol=1e-5)