introvoyz041's picture
Migrated from GitHub
6288873 verified
from dataclasses import dataclass
from functools import partial
from os import PathLike
import h5py
import jax.numpy as jnp
from jax import jit
from varipeps import varipeps_config
import varipeps.config
from varipeps.peps import PEPS_Tensor, PEPS_Unit_Cell
from varipeps.contractions import apply_contraction, Definitions
from varipeps.expectation.model import Expectation_Model
from varipeps.expectation.one_site import calc_one_site_multi_gates
from varipeps.expectation.two_sites import _two_site_workhorse
from varipeps.expectation.spiral_helpers import apply_unitary
from varipeps.typing import Tensor
from varipeps.mapping import Map_To_PEPS_Model
from varipeps.utils.random import PEPS_Random_Number_Generator
from typing import (
Sequence,
Union,
List,
Callable,
TypeVar,
Optional,
Tuple,
Type,
Dict,
Any,
)
T_Honeycomb_Map_To_Square = TypeVar(
"T_Honeycomb_Map_To_Square", bound="Honeycomb_Map_To_Square"
)
@dataclass
class Honeycomb_Expectation_Value(Expectation_Model):
"""
Class to calculate expectation values for a mapped Honeycomb
structure.
.. figure:: /images/honeycomb_structure.*
:align: center
:width: 100%
:alt: Structure of the Honeycomb/Brickwall lattice with links marked.
Structure of the Honeycomb/Brickwall lattice with links marked.
\\
Args:
x_gates (:term:`sequence` of :obj:`jax.numpy.ndarray`):
Sequence with the gates that should be applied to x links of the lattice.
y_gates (:term:`sequence` of :obj:`jax.numpy.ndarray`):
Sequence with the gates that should be applied to y links of the lattice.
z_gates (:term:`sequence` of :obj:`jax.numpy.ndarray`):
Sequence with the gates that should be applied to z links of the lattice.
real_d (:obj:`int`):
Physical dimension of a single site before mapping.
normalization_factor (:obj:`int`):
Factor which should be used to normalize the calculated values.
Likely will be 2 for the a single layer structure.
is_spiral_peps (:obj:`bool`):
Flag if the expectation value is for a spiral iPEPS ansatz.
spiral_unitary_operator (:obj:`jax.numpy.ndarray`):
Operator used to generate unitary for spiral iPEPS ansatz. Required
if spiral iPEPS ansatz is used.
"""
x_gates: Sequence[jnp.ndarray]
y_gates: Sequence[jnp.ndarray]
z_gates: Sequence[jnp.ndarray]
real_d: int
normalization_factor: int = 2
is_spiral_peps: bool = False
spiral_unitary_operator: Optional[jnp.ndarray] = None
def __post_init__(self) -> None:
if isinstance(self.x_gates, jnp.ndarray):
self.x_gates = (self.x_gates,)
if isinstance(self.y_gates, jnp.ndarray):
self.y_gates = (self.y_gates,)
if isinstance(self.z_gates, jnp.ndarray):
self.z_gates = (self.z_gates,)
if (
(
len(self.x_gates) > 0
and len(self.y_gates) > 0
and len(self.x_gates) != len(self.y_gates)
)
or (
len(self.x_gates) > 0
and len(self.z_gates) > 0
and len(self.x_gates) != len(self.z_gates)
)
or (
len(self.y_gates) > 0
and len(self.z_gates) > 0
and len(self.y_gates) != len(self.z_gates)
)
):
raise ValueError("Lengths of gate lists mismatch.")
self._y_tuple = tuple(self.y_gates)
self._z_tuple = tuple(self.z_gates)
self._result_type = (
jnp.float64
if all(jnp.allclose(g, g.T.conj()) for g in self.x_gates)
and all(jnp.allclose(g, g.T.conj()) for g in self.y_gates)
and all(jnp.allclose(g, g.T.conj()) for g in self.z_gates)
else jnp.complex128
)
if self.is_spiral_peps:
self._spiral_D, self._spiral_sigma = jnp.linalg.eigh(
self.spiral_unitary_operator
)
def __call__(
self,
peps_tensors: Sequence[jnp.ndarray],
unitcell: PEPS_Unit_Cell,
spiral_vectors: Optional[Union[jnp.ndarray, Sequence[jnp.ndarray]]] = None,
*,
normalize_by_size: bool = True,
only_unique: bool = True,
return_single_gate_results: bool = False,
) -> Union[jnp.ndarray, List[jnp.ndarray]]:
result = [
jnp.array(0, dtype=self._result_type)
for _ in range(
max(
len(self.x_gates),
len(self.y_gates),
len(self.z_gates),
)
)
]
if self.is_spiral_peps:
if isinstance(spiral_vectors, jnp.ndarray):
spiral_vectors = (spiral_vectors,)
if len(spiral_vectors) != 1:
raise ValueError("Length mismatch for spiral vectors!")
working_h_gates = tuple(
apply_unitary(
h,
jnp.array((0, 1)),
spiral_vectors,
self._spiral_D,
self._spiral_sigma,
self.real_d,
2,
(1,),
varipeps_config.spiral_wavevector_type,
)
for h in self._y_tuple
)
working_v_gates = tuple(
apply_unitary(
v,
jnp.array((1, 0)),
spiral_vectors,
self._spiral_D,
self._spiral_sigma,
self.real_d,
2,
(1,),
varipeps_config.spiral_wavevector_type,
)
for v in self._z_tuple
)
else:
working_h_gates = self._y_tuple
working_v_gates = self._z_tuple
for x, iter_rows in unitcell.iter_all_rows(only_unique=only_unique):
for y, view in iter_rows:
# On site x term
if len(self.x_gates) > 0:
onsite_tensor = peps_tensors[view.get_indices((0, 0))[0][0]]
onsite_tensor_obj = view[0, 0][0][0]
step_result_x = calc_one_site_multi_gates(
onsite_tensor,
onsite_tensor_obj,
self.x_gates,
)
for sr_i, sr in enumerate(step_result_x):
result[sr_i] += sr
# y term
if len(self.y_gates) > 0:
horizontal_tensors_i = view.get_indices((0, slice(0, 2, None)))
horizontal_tensors = [
peps_tensors[i] for j in horizontal_tensors_i for i in j
]
horizontal_tensor_objs = [t for tl in view[0, :2] for t in tl]
for ti in range(2):
t = horizontal_tensors[ti]
horizontal_tensors[ti] = t.reshape(
t.shape[0],
t.shape[1],
self.real_d,
self.real_d,
t.shape[3],
t.shape[4],
)
density_matrix_left = apply_contraction(
"honeycomb_density_matrix_left",
[horizontal_tensors[0]],
[horizontal_tensor_objs[0]],
[],
disable_identity_check=True,
)
density_matrix_right = apply_contraction(
"honeycomb_density_matrix_right",
[horizontal_tensors[1]],
[horizontal_tensor_objs[1]],
[],
disable_identity_check=True,
)
step_result_y = _two_site_workhorse(
density_matrix_left,
density_matrix_right,
working_h_gates,
self._result_type is jnp.float64,
)
for sr_i, sr in enumerate(step_result_y):
result[sr_i] += sr
# z term
if len(self.z_gates) > 0:
vertical_tensors_i = view.get_indices((slice(0, 2, None), 0))
vertical_tensors = [
peps_tensors[i] for j in vertical_tensors_i for i in j
]
vertical_tensor_objs = [t for tl in view[:2, 0] for t in tl]
for ti in range(2):
t = vertical_tensors[ti]
vertical_tensors[ti] = t.reshape(
t.shape[0],
t.shape[1],
self.real_d,
self.real_d,
t.shape[3],
t.shape[4],
)
density_matrix_top = apply_contraction(
"honeycomb_density_matrix_top",
[vertical_tensors[0]],
[vertical_tensor_objs[0]],
[],
disable_identity_check=True,
)
density_matrix_bottom = apply_contraction(
"honeycomb_density_matrix_bottom",
[vertical_tensors[1]],
[vertical_tensor_objs[1]],
[],
disable_identity_check=True,
)
step_result_z = _two_site_workhorse(
density_matrix_top,
density_matrix_bottom,
working_v_gates,
self._result_type is jnp.float64,
)
for sr_i, sr in enumerate(step_result_z):
result[sr_i] += sr
if normalize_by_size:
if only_unique:
size = unitcell.get_len_unique_tensors()
else:
size = unitcell.get_size()[0] * unitcell.get_size()[1]
size = size * self.normalization_factor
result = [r / size for r in result]
if len(result) == 1:
return result[0]
else:
return result
def save_to_group(self, grp: h5py.Group):
cls = type(self)
grp.attrs["class"] = f"{cls.__module__}.{cls.__qualname__}"
grp_gates = grp.create_group("gates", track_order=True)
grp_gates.attrs["len"] = len(self.x_gates)
for i, (x_g, y_g, z_g) in enumerate(
zip(self.x_gates, self.y_gates, self.z_gates, strict=True)
):
grp_gates.create_dataset(
f"x_gate_{i:d}",
data=x_g,
compression="gzip",
compression_opts=6,
)
grp_gates.create_dataset(
f"y_gate_{i:d}", data=y_g, compression="gzip", compression_opts=6
)
grp_gates.create_dataset(
f"z_gate_{i:d}", data=z_g, compression="gzip", compression_opts=6
)
grp.attrs["real_d"] = self.real_d
grp.attrs["normalization_factor"] = self.normalization_factor
grp.attrs["is_spiral_peps"] = self.is_spiral_peps
if self.is_spiral_peps:
grp.create_dataset(
"spiral_unitary_operator",
data=self.spiral_unitary_operator,
compression="gzip",
compression_opts=6,
)
@classmethod
def load_from_group(cls, grp: h5py.Group):
if not grp.attrs["class"] == f"{cls.__module__}.{cls.__qualname__}":
raise ValueError(
"The HDF5 group suggests that this is not the right class to load data from it."
)
x_gates = tuple(
jnp.asarray(grp["gates"][f"x_gate_{i:d}"])
for i in range(grp["gates"].attrs["len"])
)
y_gates = tuple(
jnp.asarray(grp["gates"][f"y_gate_{i:d}"])
for i in range(grp["gates"].attrs["len"])
)
z_gates = tuple(
jnp.asarray(grp["gates"][f"z_gate_{i:d}"])
for i in range(grp["gates"].attrs["len"])
)
is_spiral_peps = grp.attrs["is_spiral_peps"]
if is_spiral_peps:
spiral_unitary_operator = jnp.asarray(grp["spiral_unitary_operator"])
else:
spiral_unitary_operator = None
return cls(
x_gates=x_gates,
y_gates=y_gates,
z_gates=z_gates,
real_d=grp.attrs["real_d"],
normalization_factor=grp.attrs["normalization_factor"],
is_spiral_peps=is_spiral_peps,
spiral_unitary_operator=spiral_unitary_operator,
)
@dataclass
class Honeycomb_Map_To_Square(Map_To_PEPS_Model):
"""
Map the Honeycomb/Brickwall TN structure to a square PEPS unitcell.
The convention for the input tensors is:
Convention for physical site tensors:
* t1: [left, bottom, phys, right]
* t2: [left, phys, right, top]
The both tensors will be contracted along the right bond of t1 and the
left bond of t2.
Args:
unitcell_structure (:term:`sequence` of :term:`sequence` of :obj:`int` or 2d array):
Two dimensional array modeling the structure of the unit cell. For
details see the description of :obj:`~varipeps.peps.PEPS_Unit_Cell`.
chi (:obj:`int`):
Bond dimension of environment tensors which should be used for the
unit cell generated.
max_chi (:obj:`int`):
Maximal allowed bond dimension of environment tensors which should be
used for the unit cell generated.
"""
unitcell_structure: Sequence[Sequence[int]]
chi: int
max_chi: Optional[int] = None
@staticmethod
def _map_single_structure(site1: jnp.ndarray, site2: jnp.ndarray):
result = jnp.tensordot(site1, site2, ((3,), (0,)))
return result.reshape(
result.shape[0],
result.shape[1],
-1,
result.shape[4],
result.shape[5],
)
def __call__(
self,
input_tensors: Sequence[jnp.ndarray],
*,
generate_unitcell: bool = True,
) -> Union[List[jnp.ndarray], Tuple[List[jnp.ndarray], PEPS_Unit_Cell]]:
num_peps_sites = len(input_tensors) // 2
if num_peps_sites * 2 != len(input_tensors):
raise ValueError(
"Input tensors seems not be a list for a Honeycomb TN system."
)
peps_tensors = [
self._map_single_structure(*(input_tensors[(i * 2) : (i * 2 + 2)]))
for i in range(num_peps_sites)
]
if generate_unitcell:
peps_tensor_objs = [
PEPS_Tensor.from_tensor(
i,
i.shape[2],
(i.shape[0], i.shape[1], i.shape[3], i.shape[4]),
self.chi,
self.max_chi,
)
for i in peps_tensors
]
unitcell = PEPS_Unit_Cell.from_tensor_list(
peps_tensor_objs, self.unitcell_structure
)
return peps_tensors, unitcell
return peps_tensors
@classmethod
def random(
cls: Type[T_Honeycomb_Map_To_Square],
structure: Sequence[Sequence[int]],
d: int,
D: int,
chi: Union[int, Sequence[int]],
dtype: Type[jnp.number],
max_chi: int,
*,
seed: Optional[int] = None,
destroy_random_state: bool = True,
) -> Tuple[List[jnp.ndarray], T_Honeycomb_Map_To_Square]:
structure_arr = jnp.asarray(structure)
structure_arr, tensors_i = PEPS_Unit_Cell._check_structure(structure_arr)
# Check the inputs
if not isinstance(d, int):
raise ValueError("d has to be a single integer.")
if not isinstance(D, int):
raise ValueError("D has to be a single integer.")
if not isinstance(chi, int):
raise ValueError("chi has to be a single integer.")
# Generate the PEPS tensors
if destroy_random_state:
PEPS_Random_Number_Generator.destroy_state()
rng = PEPS_Random_Number_Generator.get_generator(seed, backend="jax")
result_tensors = []
for i in tensors_i:
result_tensors.append(rng.block((D, D, d, D), dtype=dtype)) # t1
result_tensors.append(rng.block((D, d, D, D), dtype=dtype)) # t2
return result_tensors, cls(
unitcell_structure=structure, chi=chi, max_chi=max_chi
)
@classmethod
def save_to_file(
cls,
path: PathLike,
tensors: List[jnp.ndarray],
unitcell: PEPS_Unit_Cell,
*,
store_config: bool = True,
auxiliary_data: Optional[Dict[str, Any]] = None,
) -> None:
"""
Save Honeycomb PEPS tensors and unit cell to a HDF5 file.
This function creates a single group "honeycomb_peps" in the file
and pass this group to the method
:obj:`~Honeycomb_Map_To_Square.save_to_group` then.
Args:
path (:obj:`os.PathLike`):
Path of the new file. Caution: The file will overwritten if existing.
tensors (:obj:`list` of :obj:`jax.numpy.ndarray`):
List with the PEPS tensors which should be stored in the file.
unitcell (:obj:`~varipeps.peps.PEPS_Unit_Cell`):
Full unit cell object which should be stored in the file.
Keyword args:
store_config (:obj:`bool`):
Store the current values of the global config object into the HDF5
file as attrs of an extra group.
auxiliary_data (:obj:`dict` with :obj:`str` to storable objects, optional):
Dictionary with string indexed auxiliary HDF5-storable entries which
should be stored along the other data in the file.
"""
with h5py.File(path, "w", libver=("earliest", "v110")) as f:
grp = f.create_group("honeycomb_peps")
cls.save_to_group(grp, tensors, unitcell, store_config=store_config)
if auxiliary_data is not None:
grp_aux = f.create_group("auxiliary_data")
grp_aux.attrs["keys"] = list(auxiliary_data.keys())
for key, val in auxiliary_data.items():
if key == "keys":
raise ValueError(
"Name 'keys' forbidden as name for auxiliary data"
)
if isinstance(
val, (jnp.ndarray, np.ndarray, collections.abc.Sequence)
):
try:
if val.ndim == 0:
val = val.reshape(1)
except AttributeError:
pass
grp_aux.create_dataset(
key,
data=jnp.asarray(val),
compression="gzip",
compression_opts=6,
)
else:
grp_aux.attrs[key] = val
@staticmethod
def save_to_group(
grp: h5py.Group,
tensors: List[jnp.ndarray],
unitcell: PEPS_Unit_Cell,
*,
store_config: bool = True,
) -> None:
"""
Save Honeycomb PEPS tensors and unit cell to a HDF5 group which is be
passed to the method.
Args:
grp (:obj:`h5py.Group`):
HDF5 group object to store the data into.
tensors (:obj:`list` of :obj:`jax.numpy.ndarray`):
List with the PEPS tensors which should be stored in the file.
unitcell (:obj:`~varipeps.peps.PEPS_Unit_Cell`):
Full unit cell object which should be stored in the file.
Keyword args:
store_config (:obj:`bool`):
Store the current values of the global config object into the HDF5
file as attrs of an extra group.
"""
num_peps_sites = len(tensors) // 2
if num_peps_sites * 2 != len(tensors):
raise ValueError(
"Input tensors seems not be a list for a Honeycomb system."
)
grp_peps = grp.create_group("honeycomb_peps_tensors", track_order=True)
grp_peps.attrs["num_peps_sites"] = num_peps_sites
for i in range(num_peps_sites):
(
t1,
t2,
) = tensors[(i * 2) : (i * 2 + 2)]
grp_peps.create_dataset(
f"site{i}_t1", data=t1, compression="gzip", compression_opts=6
)
grp_peps.create_dataset(
f"site{i}_t2", data=t2, compression="gzip", compression_opts=6
)
grp_unitcell = grp.create_group("unitcell")
unitcell.save_to_group(grp_unitcell, store_config=store_config)
@classmethod
def load_from_file(
cls: Type[T_Honeycomb_Map_To_Square],
path: PathLike,
*,
return_config: bool = False,
return_auxiliary_data: bool = False,
) -> Union[
Tuple[List[jnp.ndarray], PEPS_Unit_Cell],
Tuple[List[jnp.ndarray], PEPS_Unit_Cell, varipeps.config.VariPEPS_Config],
]:
"""
Load Honeycomb PEPS tensors and unit cell from a HDF5 file.
This function read the group "honeycomb_peps" from the file and pass
this group to the method
:obj:`~Honeycomb_Map_To_Square.load_from_group` then.
Args:
path (:obj:`os.PathLike`):
Path of the HDF5 file.
Keyword args:
return_config (:obj:`bool`):
Return a config object initialized with the values from the HDF5
files. If no config is stored in the file, just the data is returned.
Missing config flags in the file uses the default values from the
config object.
return_auxiliary_data (:obj:`bool`):
Return dictionary with string indexed auxiliary data which has been
should be stored along the other data in the file.
Returns:
:obj:`tuple`\ (:obj:`list`\ (:obj:`jax.numpy.ndarray`), :obj:`~varipeps.peps.PEPS_Unit_Cell`) or :obj:`tuple`\ (:obj:`list`\ (:obj:`jax.numpy.ndarray`), :obj:`~varipeps.peps.PEPS_Unit_Cell`, :obj:`~varipeps.config.VariPEPS_Config`):
The tuple with the list of the PESS tensors and the PEPS unitcell
is returned. If ``return_config = True``. the config is returned
as well. If ``return_auxiliary_data = True``. the auxiliary data is
returned as well.
"""
with h5py.File(path, "r") as f:
out = cls.load_from_group(f["honeycomb_peps"], return_config=return_config)
auxiliary_data = {}
auxiliary_data_grp = f.get("auxiliary_data")
if auxiliary_data_grp is not None:
for k in auxiliary_data_grp.attrs["keys"]:
aux_d = auxiliary_data_grp.get(k)
if aux_d is None:
aux_d = auxiliary_data_grp.attrs[k]
else:
aux_d = jnp.asarray(aux_d)
auxiliary_data[k] = aux_d
else:
max_trunc_error_list = f.get("max_trunc_error_list")
if max_trunc_error_list is not None:
auxiliary_data["max_trunc_error_list"] = jnp.asarray(
max_trunc_error_list
)
if return_config and return_auxiliary_data:
return out[0], out[1], out[2], auxiliary_data
elif return_config:
return out[0], out[1], out[2]
elif return_auxiliary_data:
return out[0], out[1], auxiliary_data
return out[0], out[1]
@staticmethod
def load_from_group(
grp: h5py.Group,
*,
return_config: bool = False,
) -> Union[
Tuple[List[jnp.ndarray], PEPS_Unit_Cell],
Tuple[List[jnp.ndarray], PEPS_Unit_Cell, varipeps.config.VariPEPS_Config],
]:
"""
Load the unit cell from a HDF5 group which is be passed to the method.
Args:
grp (:obj:`h5py.Group`):
HDF5 group object to load the data from.
Keyword args:
return_config (:obj:`bool`):
Return a config object initialized with the values from the HDF5
files. If no config is stored in the file, just the data is returned.
Missing config flags in the file uses the default values from the
config object.
Returns:
:obj:`tuple`\ (:obj:`list`\ (:obj:`jax.numpy.ndarray`), :obj:`~varipeps.peps.PEPS_Unit_Cell`) or :obj:`tuple`\ (:obj:`list`\ (:obj:`jax.numpy.ndarray`), :obj:`~varipeps.peps.PEPS_Unit_Cell`, :obj:`~varipeps.config.VariPEPS_Config`):
The tuple with the list of the PESS tensors and the PEPS unitcell
is returned. If ``return_config = True``. the config is returned
as well.
"""
grp_peps = grp["honeycomb_peps_tensors"]
num_peps_sites = grp_peps.attrs["num_peps_sites"]
tensors = []
for i in range(num_peps_sites):
tensors.append(jnp.asarray(grp_peps[f"site{i}_t1"]))
tensors.append(jnp.asarray(grp_peps[f"site{i}_t2"]))
out = PEPS_Unit_Cell.load_from_group(
grp["unitcell"], return_config=return_config
)
if return_config:
return tensors, out[0], out[1]
return tensors, out
@classmethod
def autosave_wrapper(
cls: Type[T_Honeycomb_Map_To_Square],
filename: PathLike,
tensors: jnp.ndarray,
unitcell: PEPS_Unit_Cell,
counter: Optional[int] = None,
auxiliary_data: Optional[Dict[str, Any]] = None,
) -> None:
if counter is not None:
cls.save_to_file(
f"{str(filename)}.{counter}",
tensors,
unitcell,
auxiliary_data=auxiliary_data,
)
else:
cls.save_to_file(filename, tensors, unitcell, auxiliary_data=auxiliary_data)