File size: 6,249 Bytes
114594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 |
import operator
from functools import reduce, singledispatch
from sympy.core.expr import Expr
from sympy.core.singleton import S
from sympy.matrices.expressions.hadamard import HadamardProduct
from sympy.matrices.expressions.inverse import Inverse
from sympy.matrices.expressions.matexpr import (MatrixExpr, MatrixSymbol)
from sympy.matrices.expressions.special import Identity, OneMatrix
from sympy.matrices.expressions.transpose import Transpose
from sympy.combinatorics.permutations import _af_invert
from sympy.matrices.expressions.applyfunc import ElementwiseApplyFunction
from sympy.tensor.array.expressions.array_expressions import (
_ArrayExpr, ZeroArray, ArraySymbol, ArrayTensorProduct, ArrayAdd,
PermuteDims, ArrayDiagonal, ArrayElementwiseApplyFunc, get_rank,
get_shape, ArrayContraction, _array_tensor_product, _array_contraction,
_array_diagonal, _array_add, _permute_dims, Reshape)
from sympy.tensor.array.expressions.from_matrix_to_array import convert_matrix_to_array
@singledispatch
def array_derive(expr, x):
"""
Derivatives (gradients) for array expressions.
"""
raise NotImplementedError(f"not implemented for type {type(expr)}")
@array_derive.register(Expr)
def _(expr: Expr, x: _ArrayExpr):
return ZeroArray(*x.shape)
@array_derive.register(ArrayTensorProduct)
def _(expr: ArrayTensorProduct, x: Expr):
args = expr.args
addend_list = []
for i, arg in enumerate(expr.args):
darg = array_derive(arg, x)
if darg == 0:
continue
args_prev = args[:i]
args_succ = args[i+1:]
shape_prev = reduce(operator.add, map(get_shape, args_prev), ())
shape_succ = reduce(operator.add, map(get_shape, args_succ), ())
addend = _array_tensor_product(*args_prev, darg, *args_succ)
tot1 = len(get_shape(x))
tot2 = tot1 + len(shape_prev)
tot3 = tot2 + len(get_shape(arg))
tot4 = tot3 + len(shape_succ)
perm = list(range(tot1, tot2)) + \
list(range(tot1)) + list(range(tot2, tot3)) + \
list(range(tot3, tot4))
addend = _permute_dims(addend, _af_invert(perm))
addend_list.append(addend)
if len(addend_list) == 1:
return addend_list[0]
elif len(addend_list) == 0:
return S.Zero
else:
return _array_add(*addend_list)
@array_derive.register(ArraySymbol)
def _(expr: ArraySymbol, x: _ArrayExpr):
if expr == x:
return _permute_dims(
ArrayTensorProduct.fromiter(Identity(i) for i in expr.shape),
[2*i for i in range(len(expr.shape))] + [2*i+1 for i in range(len(expr.shape))]
)
return ZeroArray(*(x.shape + expr.shape))
@array_derive.register(MatrixSymbol)
def _(expr: MatrixSymbol, x: _ArrayExpr):
m, n = expr.shape
if expr == x:
return _permute_dims(
_array_tensor_product(Identity(m), Identity(n)),
[0, 2, 1, 3]
)
return ZeroArray(*(x.shape + expr.shape))
@array_derive.register(Identity)
def _(expr: Identity, x: _ArrayExpr):
return ZeroArray(*(x.shape + expr.shape))
@array_derive.register(OneMatrix)
def _(expr: OneMatrix, x: _ArrayExpr):
return ZeroArray(*(x.shape + expr.shape))
@array_derive.register(Transpose)
def _(expr: Transpose, x: Expr):
# D(A.T, A) ==> (m,n,i,j) ==> D(A_ji, A_mn) = d_mj d_ni
# D(B.T, A) ==> (m,n,i,j) ==> D(B_ji, A_mn)
fd = array_derive(expr.arg, x)
return _permute_dims(fd, [0, 1, 3, 2])
@array_derive.register(Inverse)
def _(expr: Inverse, x: Expr):
mat = expr.I
dexpr = array_derive(mat, x)
tp = _array_tensor_product(-expr, dexpr, expr)
mp = _array_contraction(tp, (1, 4), (5, 6))
pp = _permute_dims(mp, [1, 2, 0, 3])
return pp
@array_derive.register(ElementwiseApplyFunction)
def _(expr: ElementwiseApplyFunction, x: Expr):
assert get_rank(expr) == 2
assert get_rank(x) == 2
fdiff = expr._get_function_fdiff()
dexpr = array_derive(expr.expr, x)
tp = _array_tensor_product(
ElementwiseApplyFunction(fdiff, expr.expr),
dexpr
)
td = _array_diagonal(
tp, (0, 4), (1, 5)
)
return td
@array_derive.register(ArrayElementwiseApplyFunc)
def _(expr: ArrayElementwiseApplyFunc, x: Expr):
fdiff = expr._get_function_fdiff()
subexpr = expr.expr
dsubexpr = array_derive(subexpr, x)
tp = _array_tensor_product(
dsubexpr,
ArrayElementwiseApplyFunc(fdiff, subexpr)
)
b = get_rank(x)
c = get_rank(expr)
diag_indices = [(b + i, b + c + i) for i in range(c)]
return _array_diagonal(tp, *diag_indices)
@array_derive.register(MatrixExpr)
def _(expr: MatrixExpr, x: Expr):
cg = convert_matrix_to_array(expr)
return array_derive(cg, x)
@array_derive.register(HadamardProduct)
def _(expr: HadamardProduct, x: Expr):
raise NotImplementedError()
@array_derive.register(ArrayContraction)
def _(expr: ArrayContraction, x: Expr):
fd = array_derive(expr.expr, x)
rank_x = len(get_shape(x))
contraction_indices = expr.contraction_indices
new_contraction_indices = [tuple(j + rank_x for j in i) for i in contraction_indices]
return _array_contraction(fd, *new_contraction_indices)
@array_derive.register(ArrayDiagonal)
def _(expr: ArrayDiagonal, x: Expr):
dsubexpr = array_derive(expr.expr, x)
rank_x = len(get_shape(x))
diag_indices = [[j + rank_x for j in i] for i in expr.diagonal_indices]
return _array_diagonal(dsubexpr, *diag_indices)
@array_derive.register(ArrayAdd)
def _(expr: ArrayAdd, x: Expr):
return _array_add(*[array_derive(arg, x) for arg in expr.args])
@array_derive.register(PermuteDims)
def _(expr: PermuteDims, x: Expr):
de = array_derive(expr.expr, x)
perm = [0, 1] + [i + 2 for i in expr.permutation.array_form]
return _permute_dims(de, perm)
@array_derive.register(Reshape)
def _(expr: Reshape, x: Expr):
de = array_derive(expr.expr, x)
return Reshape(de, get_shape(x) + expr.shape)
def matrix_derive(expr, x):
from sympy.tensor.array.expressions.from_array_to_matrix import convert_array_to_matrix
ce = convert_matrix_to_array(expr)
dce = array_derive(ce, x)
return convert_array_to_matrix(dce).doit()
|