swh:1:snp:baaf5c1e6241789b5d3986b85175366cc976774a
perm.py
import torch
import numpy as np
import src.debfly.tree
import src.partition
import src.debfly.product
from src.cluster_tree import ClusterTree
def col_permute(matrix, permutation):
""" permute columns of matrix according to permutation """
assert len(matrix.shape) == 2
if not isinstance(permutation, torch.LongTensor):
permutation = torch.LongTensor(permutation)
assert len(permutation.shape) == 1
return matrix[:, permutation]
def row_permute(matrix, permutation):
""" permute rows of matrix according to permutation """
assert len(matrix.shape) == 2
if not isinstance(permutation, torch.LongTensor):
permutation = torch.LongTensor(permutation)
assert len(permutation.shape) == 1
return matrix[permutation, :]
def permutation_to_matrix(permutation):
""" permutation is a tuple """
n = len(permutation)
return col_permute(torch.eye(n), permutation)
def inverse_permutation(permutation):
""" compute the inverse permutation """
permutation = np.array(permutation)
inverse_perm = np.argsort(permutation)
return tuple(inverse_perm)
def composition(a, b):
""" permute according to a, then permute according to b """
if not isinstance(a, torch.LongTensor):
a = torch.LongTensor(a)
if not isinstance(b, torch.LongTensor):
b = torch.LongTensor(b)
return a[b]
def perm_col_partition_to_monarch_canonical(col_partition, tree):
""" return permutation to go from col_partition to monarch canonical """
src.partition.check_col_partition_is_monarch_compatible(col_partition, tree)
p = torch.LongTensor(col_partition)
p = p.view(-1)
return p
def _perm_row_canonical_to_block(diag, row):
p = torch.LongTensor(src.partition.canonical_row_partition_monarch(diag, row))
return p.view(-1)
def perm_row_partition_to_monarch_canonical(row_partition, tree):
""" permutation to go from row_partition to monarch canoncial """
src.partition.check_row_partition_is_monarch_compatible(row_partition, tree)
diag = len(row_partition)
row = len(row_partition[0])
p1 = torch.LongTensor(row_partition)
p1 = p1.view(-1)
perm = composition(p1, inverse_permutation(_perm_row_canonical_to_block(diag, row)))
return perm
def perm_row_tree_to_butterfly_canonical(row_cluster_tree):
""" permutation to go from the row cluster tree to the canonical butterfly row cluster tree """
assert isinstance(row_cluster_tree, ClusterTree)
size = len(row_cluster_tree.node)
log_n = int(np.log2(size))
assert size == 2**log_n
leaves = np.squeeze(np.array(row_cluster_tree.nodes_at_level(log_n)), axis=1).tolist()
canonical_tree = src.cluster_tree.even_odd_split_cluster_tree(list(range(size)))
canonical_leaves = np.squeeze(np.array(canonical_tree.nodes_at_level(log_n)), axis=1).tolist()
return composition(leaves, inverse_permutation(canonical_leaves))
def perm_col_tree_to_butterfly_canonical(col_cluster_tree):
""" permutation to go from the column cluster tree to the canonical butterfly column cluster tree """
assert isinstance(col_cluster_tree, ClusterTree)
size = len(col_cluster_tree.node)
log_n = int(np.log2(size))
assert size == 2 ** log_n
leaves = np.squeeze(np.array(col_cluster_tree.nodes_at_level(log_n)), axis=1).tolist()
canonical_tree = src.cluster_tree.middle_split_cluster_tree(list(range(size)))
canonical_leaves = np.squeeze(np.array(canonical_tree.nodes_at_level(log_n)), axis=1).tolist()
return composition(leaves, inverse_permutation(canonical_leaves))
def perm_DFT(num_factors):
result = []
size = 2 ** num_factors
for i in range(num_factors):
if i == 0:
continue
for j in range(3):
z = perm_type(i + 1, j)
result.append(np.kron(np.identity(2 ** (num_factors - 1 - i)), z))
return result
def perm_type(i, type):
"""
Type 0 is c in paper. Type 1 is b in paper. Type 2 is a in paper.
:param i:
:param type:
:return:
"""
size = 2 ** i
result = np.zeros((size,size))
if type == 0:
result[np.arange(size//2), np.arange(size//2)] = 1
result[size//2 + np.arange(size//2), size - 1 - np.arange(size//2)] = 1
elif type == 1:
result[size // 2 - 1 - np.arange(size//2), np.arange(size//2)] = 1
result[size // 2 + np.arange(size//2), size//2 + np.arange(size//2)] = 1
else:
result[np.arange(size//2), np.arange(size//2) * 2] = 1
result[size//2 + np.arange(size//2), np.arange(size//2) * 2 + 1] = 1
return result
def bit_reversal_permutation(log_n):
perm_mat = src.debfly.product.matrix_product(perm_DFT(log_n))
return perm_mat
if __name__ == '__main__':
p = (3, 2, 0, 1)
mat = torch.randn(4, 4)
print(mat @ permutation_to_matrix(p) == col_permute(mat, p))
print(permutation_to_matrix(p).t() @ mat == row_permute(mat, p))
print(inverse_permutation(p))
print(col_permute(col_permute(mat, inverse_permutation(p)), p) == mat)
print(col_permute(col_permute(mat, p), inverse_permutation(p)) == mat)
print(row_permute(row_permute(mat, inverse_permutation(p)), p) == mat)
print(row_permute(row_permute(mat, p), inverse_permutation(p)) == mat)
p1 = (0, 1, 3, 2)
p2 = (0, 2, 1, 3)
p1_then_p2 = composition(p1, p2)
print(p1_then_p2)
print(col_permute(col_permute(mat, p1), p2) == col_permute(mat, p1_then_p2))