def static_matrix_compressor(self, a_matrix):
"""Simhash decomposition of a_matrix.
Args:
a_matrix: input matrix
Returns:
List [b_matrix] which is the simhash approximation of a_matrix. Rank is
taken from spec.rank and interpreted to be a compression factor.
"""
# Tag used for all print statements within this method.
logging_tag = 'Inside simhash static_matrix_compressor:'
rank = ((np.min(a_matrix.shape) * 100) // self._spec.rank) + 1
tf.logging.info('%s compression factor, old rank, and new rank '
'are %s %s %s',
logging_tag,
self._spec.rank,
a_matrix.shape[1], rank)
r, s, d = decompose_matrix.np_simhash_decompose(
a_matrix, rank, seed=self._seed)
self.uncompressed_size = np.size(a_matrix)
self.compressed_size = np.size(r)
logging.info(
'%s r,s,d shapes are: %s, %s, %s, compressed and uncompressed size are %s %s',
logging_tag, r.shape, s.shape, d.shape, self.uncompressed_size,
self.compressed_size)
a_matrix_approx = np.dot(r, np.dot(s, d))
logging.info('%s a_matrix_approx norm: %s', logging_tag,
np.linalg.norm(a_matrix_approx))
return [a_matrix_approx.astype(np.float32)]
def static_matrix_compressor(self, a_matrix):
"""Simhash decomposition of a_matrix.
Args:
a_matrix: input matrix.
Returns:
List [b_matrix] which is the simhash approximation of a_matrix. Rank is
taken from spec.rank and interpreted to be a compression factor.
"""
# Tag used for all print statements within this method.
logging_tag = 'Inside simhash static_matrix_compressor:'
# self._spec.rank can be considered a compression factor out of a 100,
# where 100 corresponds to the full size of the original matrix. For
# example, self._spec.rank of 200 means that the original rank is
# compressed by a factor of 2.
rank = np.int(np.floor(a_matrix.shape[1] * 100 / self._spec.rank))
logging.info('%s compression factor, old rank, and new rank are %s, %s, %s',
logging_tag, self._spec.rank, a_matrix.shape[1], rank)
r, s, d = decompose_matrix.np_simhash_decompose(
a_matrix, rank, seed=self._seed)
logging.info('%s r,s,d shapes are: %s, %s, %s', logging_tag, r.shape,
s.shape, d.shape)
a_matrix_approx = np.dot(r, np.dot(s, d))
logging.info('%s a_matrix_approx norm: %s', logging_tag,
np.linalg.norm(a_matrix_approx))
return [a_matrix_approx.astype(np.float32)]