def test_kmeans_quantization_gaussian_mixture(self):
"""Check approximate recovery for a well-separated mixture of Gaussian.
"""
gaussian_0 = np.random.randn(100, 4) + (-5) * np.eye(4)[0, :]
gaussian_1 = np.random.randn(100, 4)
gaussian_2 = np.random.randn(100, 4) + 5 * np.eye(4)[0, :]
matrix = np.concatenate((gaussian_0, gaussian_1, gaussian_2), axis=0)
matrix = np.reshape(matrix, (-1, 20))
[codebook,
encoding] = kmeans_quantize.kmeans_quantize_block(matrix,
levels=3,
pruning_factor=0.0,
block_size=4,
is_padded=True)
self.assertEqual(codebook.shape, (3, 4))
self.assertEqual(encoding.shape,
(matrix.shape[0], matrix.shape[1] // 4))
recovered_matrix = codebook[encoding].reshape(matrix.shape)
# Assert recovered matrix is close to original matrix.
self.assertLess(
np.linalg.norm((recovered_matrix - matrix).flatten()) / 300, 0.2)
def test_basic_kmeans_quantization(self):
matrix = np.random.randn(20, 20)
[codebook,
encoding] = kmeans_quantize.kmeans_quantize_block(matrix,
levels=7,
pruning_factor=0.0,
block_size=4,
is_padded=True)
self.assertEqual(codebook.shape, (7, 4))
self.assertEqual(encoding.shape, (100, ))
def static_matrix_compressor(self, a_matrix):
"""K-means decomposition of a_matrix.
Args:
a_matrix: input matrix
Returns:
[codebook, a_matrix_encoding]: rows of codebook are centroid vectors, and
a_matrix_encoding is an array of centroid indices for blocks in a_matrix.
"""
[codebook, a_matrix_encoding] = kmeans_quantize.kmeans_quantize_block(
a_matrix,
levels=self._spec.rank,
pruning_factor=self._spec.pruning_fraction,
block_size=self._spec.block_size,
is_padded=True)
return [codebook, a_matrix_encoding]
def test_basic_kmeans_quantization(self):
"""Test return shapes are as expected.
"""
n_rows = 20
n_cols = 20
block_size = 4
levels = 7
matrix = np.random.randn(n_rows, n_cols)
[codebook, encoding
] = kmeans_quantize.kmeans_quantize_block(matrix,
levels=levels,
pruning_factor=0.0,
block_size=block_size,
is_padded=True)
self.assertEqual(codebook.shape, (levels, block_size))
self.assertEqual(encoding.shape, (n_rows, n_cols // block_size))
def test_block_quantize_with_zero_blocks(self):
"""Check zero blocks remains zero after clustering.
"""
n_rows = 100
n_cols = 4
n_zero_rows = 2
matrix = np.random.randn(n_rows, n_cols)
matrix[:n_zero_rows, :] = 0.0
[codebook,
encoding] = kmeans_quantize.kmeans_quantize_block(matrix,
levels=3,
pruning_factor=0.0,
block_size=n_cols,
is_padded=True)
recovered_matrix = codebook[encoding].reshape(matrix.shape)
# Zero blocks should remain zero.
self.assertEqual(np.linalg.norm(recovered_matrix[:n_zero_rows, :]),
0.0)