Source code for pde.backends.torch.backend

"""Defines base class of backends that implement computations.

.. codeauthor:: David Zwicker <david.zwicker@ds.mpg.de>
"""

from __future__ import annotations

import numbers
import warnings
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Literal

import numpy as np
import torch

from ...fields import VectorField
from ...grids import GridBase
from ..base import BackendBase, OperatorInfo, TFunc
from .utils import NUMPY_TO_TORCH_DTYPE, TORCH_TO_NUMPY_DTYPE

if TYPE_CHECKING:
    from collections.abc import Callable

    from numpy.typing import DTypeLike

    from ...fields import DataFieldBase
    from ...grids import GridBase
    from ...grids.boundaries.axes import BoundariesBase
    from ...pdes import PDEBase
    from ...solvers.base import SolverBase
    from ...tools.config import Config
    from ...tools.expressions import ExpressionBase
    from ...tools.typing import (
        NumberOrArray,
        NumericArray,
        TArray,
        TField,
    )
    from ..base import TFunc
    from ..numpy.backend import OperatorInfo
    from .operators.common import TorchDifferentialOperator


[docs] class TorchBackend(BackendBase): """Defines :mod:`torch` backend.""" implementation = "torch" compile_options = { "fullgraph": True, # force compilation of entire graph (no graph breaks) "dynamic": False, # compiled functions do not support changing shapes "options": {"epilogue_fusion": True, "max_autotune": True}, } """dict: defines options that affect compilation by torch""" _dtype_cache: dict[str, dict[DTypeLike, torch.dtype]] = defaultdict(dict) """dict: contains information about the dtypes available for the current device""" _emitted_downcast_warning: bool = False """bool: global flag to track whether we already warned about downcasting""" def __init__( self, config: Config | None = None, *, name: str | None = None, device: str = "config", ): """Initialize the torch backend. Args: config (:class:`~pde.tools.config.Config`): Configuration data for the backend name (str): The name of the backend device (str): The torch device to use. Special values are "config" (read from configuration) and "auto" (use CUDA if available, otherwise CPU) """ if config is None: from .config import DEFAULT_CONFIG as config # type: ignore super().__init__(config, name=name) self.device = device @property def device(self) -> torch.device: """The currently assigned torch device.""" return self._device @device.setter def device(self, device: str) -> None: """Set a new torch device.""" # determine which device we need to use if device == "config": device = self.config["device"] if device == "auto": device = "cuda" if torch.cuda.is_available() else "cpu" # check whether the device is available if device.startswith("cuda") and not torch.cuda.is_available(): msg = "cuda device is not available" raise RuntimeError(msg) if device.startswith("mps") and not torch.backends.mps.is_available(): msg = "mps device is not available" raise RuntimeError(msg) # set the actual device self._device = torch.device(device)
[docs] def get_torch_dtype(self, dtype: DTypeLike) -> torch.dtype: """Convert dtype to torch dtype. Args: dtype: numpy dtype to convert to corresponding torch dtype Returns: :class:`torch.dtype`: A proper dtype for torch """ if isinstance(dtype, torch.dtype): return dtype # load the dtype cache of the current device type_cache = self._dtype_cache[self.device.type] np_dtype = np.dtype(dtype) try: # try returning type from cache return type_cache[np_dtype] except KeyError: pass # convert numpy dtype to corresponding torch dtype torch_dtype = NUMPY_TO_TORCH_DTYPE[np_dtype] try: # Try to create a tensor of this dtype on the device torch.empty(1, dtype=torch_dtype, device=self.device) except TypeError: # dtype is not supported, so we see whether we need to use downcasting if self.config["dtype_downcasting"] and torch_dtype == torch.float64: if not self._emitted_downcast_warning: self._logger.warning( " %s device doesn't support float64, so we use float32 instead", self.device.type, ) self._emitted_downcast_warning = True torch_dtype = torch.float32 else: raise # store dtype in cache type_cache[np_dtype] = torch_dtype return torch_dtype
[docs] def get_numpy_dtype(self, dtype: DTypeLike) -> np.dtype: """Determine numpy dtype suitable for the torch backend. Args: dtype: numpy dtype to convert to supported dtype Returns: :class:`torch.dtype`: A numpy dtype that is compatible with the torch backend """ # convert numpy dtype to torch dtype to support cases where the torch device # only supports narrower types return TORCH_TO_NUMPY_DTYPE[self.get_torch_dtype(dtype)]
[docs] def from_numpy(self, value: Any) -> Any: """Convert values from numpy to torch representation. This method also ensures that the value is copied to the selected device. """ if isinstance(value, torch.Tensor): return value.to(self.device) # move tensor to device if isinstance(value, (np.ndarray, numbers.Number)): value_arr = np.asarray(value) # convert numbers to arrays for torch arr_torch = torch.from_numpy(value_arr) # convert to torch.Tensor dtype = self.get_torch_dtype(value_arr.dtype) return arr_torch.to(self.device, dtype=dtype) # move tensor to device msg = f"Unsupported type `{type(value).__name__}" raise TypeError(msg)
[docs] def to_numpy(self, value: Any) -> Any: """Convert native values to numpy representation.""" if isinstance(value, torch.Tensor): return value.cpu().numpy() return value
[docs] def compile_function(self, func: TFunc, **compile_options) -> TFunc: r"""General method that compiles a user function. Args: func (callable): The function that needs to be compiled for this backend **compile_options: Additional keyword arguments will be forwarded to :func:`torch.compile` """ if not self.config["compile"]: return func # compile the function using the torch backend opts = self.compile_options | compile_options return torch.compile(func, **opts) # type: ignore
[docs] def make_operator_no_bc( self, grid: GridBase, operator: str | OperatorInfo, *, dtype: DTypeLike | None = None, native: bool = False, **kwargs, ) -> TorchDifferentialOperator: """Return a compiled function applying an operator without boundary conditions. A function that takes the discretized full data as an input and an array of valid data points to which the result of applying the operator is written. Note: The resulting function does not check whether the ghost cells of the input array have been supplied with sensible values. It is the responsibility of the user to set the values of the ghost cells beforehand. Use this function only if you absolutely know what you're doing. In all other cases, :meth:`make_operator` is probably the better choice. Args: grid (:class:`~pde.grid.base.GridBase`): Grid for which the operator is needed operator (str): Identifier for the operator. Some examples are 'laplace', 'gradient', or 'divergence'. The registered operators for this grid can be obtained from the :attr:`~pde.grids.base.GridBase.operators` attribute. dtype (numpy dtype): The data type of the field. native (bool): If True, the returned functions expects the native data representation of the backend. Otherwise, the input and output are expected to be :class:`~numpy.ndarray`. **kwargs: Specifies extra arguments influencing how the operator is created. Returns: callable: the function that applies the operator. This function has the signature (arr: NumericArray, out: NumericArray), so they `out` array need to be supplied explicitly. """ # obtain details about the operator operator_info = self.get_operator_info(grid, operator) dtype = self.get_numpy_dtype(dtype or np.double) # create an operator with or without BCs torch_operator = operator_info.factory(grid, bcs=None, dtype=dtype, **kwargs) torch_operator.eval() # type: ignore # compile the function and move it to the device torch_operator_jitted = self.compile_function(torch_operator) torch_operator_jitted.to(self.device) # type: ignore if native: return torch_operator_jitted # type: ignore def operator_no_bc(arr: NumericArray, out: NumericArray) -> None: arr_torch = self.from_numpy(arr) out_torch = torch_operator_jitted(arr_torch) # type: ignore out[...] = self.to_numpy(out_torch) return operator_no_bc # type: ignore
[docs] def make_operator( self, grid: GridBase, operator: str | OperatorInfo, *, bcs: BoundariesBase, dtype: DTypeLike | None = None, native: bool = False, **kwargs, ) -> TorchDifferentialOperator: """Return a torch function applying an operator with boundary conditions. Args: grid (:class:`~pde.grid.base.GridBase`): Grid for which the operator is needed operator (str): Identifier for the operator. Some examples are 'laplace', 'gradient', or 'divergence'. The registered operators for this grid can be obtained from the :attr:`~pde.grids.base.GridBase.operators` attribute. bcs (:class:`~pde.grids.boundaries.axes.BoundariesBase`, optional): The boundary conditions used before the operator is applied dtype (numpy dtype): The data type of the field. native (bool): If True, the returned functions expects the native data representation of the backend. Otherwise, the input and output are expected to be :class:`~numpy.ndarray`. **kwargs: Specifies extra arguments influencing how the operator is created. Warning: The same operator should not be assigned to different variables that are used in the same code, because :mod:`torch` has problems compiling the resulting code. This particularly precludes caching the operators, since they then might be reused, e.g., if boundary conditions agree between different operators. Returns: callable: the function that applies the operator. This function has the signature (arr: NumericArray, out: NumericArray = None, args=None). """ # obtain details about the operator operator_info = self.get_operator_info(grid, operator) dtype = self.get_numpy_dtype(dtype or np.double) bcs = grid.get_boundary_conditions(bcs, rank=operator_info.rank_in) # create an operator with or without BCs torch_operator = operator_info.factory(grid, bcs, dtype=dtype, **kwargs) # type: ignore torch_operator.eval() # type: ignore # compile the function and move it to the device torch_operator_jitted = self.compile_function(torch_operator) torch_operator_jitted.to(self.device) # type: ignore if native: # return the native representation if requested return torch_operator_jitted # type: ignore # wrap the operator such that it can be called from numpy shape_out = (grid.dim,) * operator_info.rank_out + grid.shape # define numpy version of the operator def apply_op( arr: NumericArray, out: NumericArray | None = None, args=None ) -> NumericArray: """Set boundary conditions and apply operator.""" arr_torch = self.from_numpy(arr) if out is None: out = np.empty(shape_out, dtype=arr.dtype) elif out.shape != shape_out: msg = f"Incompatible shapes {out.shape} != {shape_out}" raise ValueError(msg) out_torch = torch_operator_jitted(arr_torch) # type: ignore out[:] = self.to_numpy(out_torch) return out # return the compiled versions of the operator return apply_op # type: ignore
[docs] def make_integrator( self, grid: GridBase, *, dtype: DTypeLike = np.double ) -> Callable[[NumericArray], NumberOrArray]: """Return function that integrates discretized data over a grid. If this function is used in a multiprocessing run (using MPI), the integrals are performed on all subgrids and then accumulated. Each process then receives the same value representing the global integral. Args: grid (:class:`~pde.grid.base.GridBase`): Grid for which the integrator is defined dtype: The data type of the field that is being integrated Returns: A function that takes a numpy array and returns the integral with the correct weights given by the cell volumes. """ from .operators.common import TorchIntegralOperator # create the torch operator integrate_torch = self.compile_function( TorchIntegralOperator(grid, dtype=self.get_numpy_dtype(dtype)) ) integrate_torch.to(self.device) def integrate_global(arr: NumericArray) -> NumberOrArray: """Integrate data. Args: arr (:class:`~numpy.ndarray`): discretized data on grid """ # move data to device arr_torch = self.from_numpy(arr) # integrate on device res = integrate_torch(arr_torch) # return result res_np = self.to_numpy(res) if res_np.ndim == 0: return res_np[()] # type: ignore return res_np # type: ignore return integrate_global
[docs] def make_inner_prod_operator( self, field: DataFieldBase, *, conjugate: bool = True ) -> Callable[[TArray, TArray, TArray | None], TArray]: """Return operator calculating the dot product between two fields. This supports both products between two vectors as well as products between a vector and a tensor. Args: field (:class:`~pde.fields.datafield_base.DataFieldBase`): Field for which the inner product is defined conjugate (bool): Whether to use the complex conjugate for the second operand Returns: function that takes two instance of :class:`~numpy.ndarray`, which contain the discretized data of the two operands. An optional third argument can specify the output array to which the result is written. """ num_axes = field.grid.num_axes def dot( a: torch.Tensor, b: torch.Tensor, out: torch.Tensor | None = None ) -> torch.Tensor: """Numpy implementation to calculate dot product between two fields.""" rank_a = a.ndim - num_axes rank_b = b.ndim - num_axes if rank_a < 1 or rank_b < 1: msg = "Fields in dot product must have rank >= 1" raise TypeError(msg) if a.shape[rank_a:] != b.shape[rank_b:]: msg = "Shapes of fields are not compatible for dot product" raise ValueError(msg) if out is not None: msg = "torch implementation of inner product does not allow `out` arg." raise TypeError(msg) if conjugate: b = b.conj() if rank_a == 1 and rank_b == 1: # result is scalar field return torch.einsum("i...,i...->...", a, b) if rank_a == 2 and rank_b == 1: # result is vector field return torch.einsum("ij...,j...->i...", a, b) if rank_a == 1 and rank_b == 2: # result is vector field return torch.einsum("i...,ij...->j...", a, b) if rank_a == 2 and rank_b == 2: # result is tensor-2 field return torch.einsum("ij...,jk...->ik...", a, b) msg = f"Unsupported shapes ({a.shape}, {b.shape})" raise TypeError(msg) return dot # type: ignore
[docs] def make_outer_prod_operator( self, field: DataFieldBase ) -> Callable[[TArray, TArray, TArray | None], TArray]: """Return operator calculating the outer product between two fields. This supports typically only supports products between two vector fields. Args: field (:class:`~pde.fields.datafield_base.DataFieldBase`): Field for which the outer product is defined Returns: function that takes two instance of :class:`~numpy.ndarray`, which contain the discretized data of the two operands. An optional third argument can specify the output array to which the result is written. """ if not isinstance(field, VectorField): msg = "Can only define outer product between vector fields" raise TypeError(msg) def outer( a: NumericArray, b: NumericArray, out: NumericArray | None = None ) -> NumericArray: """Calculate the outer product using numpy.""" if out is not None: msg = "torch implementation of inner product does not allow `out` arg." raise TypeError(msg) return torch.einsum("i...,j...->ij...", a, b) # type: ignore return outer # type: ignore
[docs] def make_pde_rhs( self, eq: PDEBase, state: TField, *, native: bool = False ) -> Callable[[TArray, float], TArray]: """Return a function for evaluating the right hand side of the PDE. Args: eq (:class:`~pde.pdes.base.PDEBase`): The object describing the differential equation state (:class:`~pde.fields.FieldBase`): An example for the state from which information can be extracted native (bool): If True, the returned functions expects the native data representation of the backend. Otherwise, the input and output are expected to be :class:`~numpy.ndarray`. Returns: Function returning deterministic part of the right hand side of the PDE. """ # the following method is deprecated since 2026-03-02 try: make_rhs = eq.make_pde_rhs_torch # type: ignore except AttributeError: # method is not implemented, which should be the default rhs_native = None else: warnings.warn( "`eq.make_pde_rhs_torch` method is deprecated. Implement " "`eq.make_evolution_rate` instead.", DeprecationWarning, stacklevel=2, ) rhs_native = make_rhs(state) if rhs_native is None: try: make_rhs = eq.make_evolution_rate except AttributeError as err: msg = ( "The right-hand side of the PDE is not implemented using the " f"`{self.name}` backend. To add the implementation, provide the " "method `make_evolution_rate`, which should return a compilable " "function calculating the evolution rate." ) raise NotImplementedError(msg) from err else: rhs_native = make_rhs(state, backend=self) # get the compiled right hand side rhs_torch = self.compile_function(rhs_native) if native: return rhs_torch def rhs(arr: NumericArray, t: float = 0) -> NumericArray: """Helper wrapping function working with torch tensors.""" arr_torch = self.from_numpy(arr) # We wrap the scalar time into a tensor, so torch correctly identifies it as # a value that is modified each time we call the function. t_torch = self.from_numpy(t) res_torch = rhs_torch(arr_torch, t_torch) return self.to_numpy(res_torch) # type: ignore return rhs # type: ignore
[docs] def make_inner_stepper( self, solver: SolverBase, stepper_style: Literal["fixed", "adaptive"], state: TField, dt: float, ) -> Callable: """Return a stepper function using an explicit scheme. Args: solver (:class:`~pde.solvers.base.SolverBase`): The solver instance, which determines how the stepper is constructed stepper_style (str): The style of the stepper, either "fixed" or "adaptive" state (:class:`~pde.fields.base.FieldBase`): An example for the state from which the grid and other information can be extracted dt (float): Time step used (Uses :attr:`SolverBase.dt_default` if `None`) Returns: Function that can be called to advance the `state` from time `t_start` to time `t_end`. The function call signature is `(state: numpy.ndarray, t_start: float, t_end: float)` """ solver.info["backend"]["device"] = self.device.type solver.info["backend"]["compile"] = self.config["compile"] return super().make_inner_stepper(solver, stepper_style, state, dt)
[docs] def make_expression_function( self, expression: ExpressionBase, *, single_arg: bool = False, user_funcs: dict[str, Callable] | None = None, ) -> Callable[..., NumberOrArray]: """Return a function evaluating an expression. Args: expression (:class:`~pde.tools.expression.ExpressionBase`): The expression that is converted to a function single_arg (bool): Determines whether the returned function accepts all variables in a single argument as an array or whether all variables need to be supplied separately. user_funcs (dict): Additional functions that can be used in the expression. Returns: function: the function """ import sympy from sympy.printing.pycode import PythonCodePrinter from .utils import SPECIAL_FUNCTIONS_TORCH # collect all the user functions user_functions = expression.user_funcs.copy() if user_funcs is not None: user_functions.update(user_funcs) user_functions.update(SPECIAL_FUNCTIONS_TORCH) user_functions = { k: self.compile_function(v) for k, v in user_functions.items() } # initialize the printer that deals with numpy arrays correctly class ListArrayPrinter(PythonCodePrinter): """Special sympy printer returning arrays as lists.""" def _print_ImmutableDenseNDimArray(self, arr): arrays = ", ".join(f"{self._print(expr)}" for expr in arr) return f"[{arrays}]" printer = ListArrayPrinter( { "fully_qualified_modules": False, "inline": True, "allow_unknown_functions": True, "user_functions": {k: k for k in user_functions}, } ) # determine the list of variables that the function depends on variables = (expression.vars,) if single_arg else tuple(expression.vars) constants = tuple(expression.consts) # turn the expression into a callable function self._logger.info("Parse sympy expression `%s`", expression._sympy_expr) func = sympy.lambdify( variables + constants, expression._sympy_expr, modules=[user_functions, "torch"], printer=printer, ) # Apply the constants if there are any. Note that we use this pattern of a # partial function instead of replacing the constants in the sympy expression # directly since sympy does not work well with numpy arrays. if constants: const_values = tuple( self.from_numpy(expression.consts[c]) for c in constants ) func = self.compile_function(func) def result(*args): return func(*args, *const_values) else: result = func return self.compile_function(result)