#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Author: Benjamin Vial
# This file is part of nannos
# License: GPLv3
# See the documentation at nannos.gitlab.io
import os
import warnings
from packaging.version import parse as parse_version
from .__about__ import __author__, __description__, __version__
from .__about__ import data as _data
from .log import *
warnings.filterwarnings(
action="ignore",
category=RuntimeWarning,
message=".*deadlock.",
)
available_backends = ["numpy", "scipy", "autograd"]
def print_info():
print(f"nannos v{__version__}")
print("=============")
print(__description__)
print(f"Author: {__author__}")
print(f"Licence: {_data['License']}")
def has_torch():
try:
import torch
return True
except ModuleNotFoundError:
return False
def has_jax():
try:
import jax
return True
except ModuleNotFoundError:
return False
def _has_cuda():
try:
import torch
return torch.cuda.is_available()
except ModuleNotFoundError:
return False
HAS_TORCH = has_torch()
HAS_CUDA = _has_cuda()
HAS_JAX = has_jax()
if HAS_TORCH:
available_backends.append("torch")
if HAS_JAX:
available_backends.append("jax")
def use_gpu(boolean):
global DEVICE
global _CPU_DEVICE
global _GPU_DEVICE
global _FORCE_GPU
_FORCE_GPU = 1
if boolean:
if BACKEND not in ["torch", "jax"]:
logger.debug(f"Cannot use GPU with {BACKEND} backend.")
_delvar("_GPU_DEVICE")
_CPU_DEVICE = True
elif BACKEND == "torch" and not HAS_TORCH:
logger.warning("pytorch not found. Cannot use GPU.")
_delvar("_GPU_DEVICE")
_CPU_DEVICE = True
elif BACKEND == "torch" and not HAS_CUDA:
logger.warning("cuda not found. Cannot use GPU.")
_delvar("_GPU_DEVICE")
_CPU_DEVICE = True
elif BACKEND == "jax" and not HAS_JAX:
logger.warning("jax not found. Cannot use GPU.")
_delvar("_GPU_DEVICE")
_CPU_DEVICE = True
else:
DEVICE = "cuda"
logger.debug("Using GPU.")
_delvar("_CPU_DEVICE")
_GPU_DEVICE = True
else:
_CPU_DEVICE = True
_delvar("_GPU_DEVICE")
DEVICE = "cpu"
logger.debug("Using CPU.")
_reload_package()
def jit(fun, **kwargs):
# if get_backend() == "jax":
# import jax
# return jax.jit(fun)
return fun
def _delvar(VAR):
if VAR in globals():
del globals()[VAR]
def _del_vars(VARS):
for VAR in VARS:
_delvar(VAR)
[docs]def set_backend(backend):
"""Set the numerical backend.
Parameters
----------
backend : str
Either ``numpy``, ``scipy``, ``autograd``, ``torch`` or ``jax``.
"""
global _NUMPY
global _SCIPY
global _AUTOGRAD
global _JAX
global _TORCH
global _FORCE_BACKEND
_FORCE_BACKEND = 1
if backend == get_backend():
return
#
# _backend_env_var = os.environ.get("NANNOS_BACKEND")
# if _backend_env_var is not None:
# if _backend_env_var in available_backends:
# if backend != _backend_env_var:
# # _delvar("_FORCE_BACKEND")
# pass
# else:
# backend = _backend_env_var
if backend == "autograd":
logger.debug("Setting autograd backend")
_AUTOGRAD = True
_del_vars(["_JAX", "_TORCH", "_SCIPY"])
elif backend == "scipy":
logger.debug("Setting scipy backend")
_SCIPY = True
_del_vars(["_JAX", "_TORCH", "_AUTOGRAD"])
elif backend == "jax":
logger.debug("Setting jax backend")
_JAX = True
_del_vars(["_SCIPY", "_TORCH", "_AUTOGRAD"])
elif backend == "torch":
_TORCH = True
logger.debug("Setting torch backend")
_del_vars(["_SCIPY", "_JAX", "_AUTOGRAD"])
elif backend == "numpy":
_NUMPY = True
logger.debug("Setting numpy backend")
_del_vars(["_SCIPY", "_JAX", "_AUTOGRAD", "_TORCH"])
else:
raise ValueError(
f"Unknown backend '{backend}'. Please choose between 'numpy', 'scipy', 'jax', 'torch' and 'autograd'."
)
_reload_package()
def _reload_package():
import importlib
import sys
import nannos
importlib.reload(nannos)
its = [s for s in sys.modules.items() if s[0].startswith("nannos")]
for k, v in its:
importlib.reload(v)
def get_backend():
if "_SCIPY" in globals():
return "scipy"
elif "_AUTOGRAD" in globals():
return "autograd"
elif "_JAX" in globals():
return "jax"
elif "_TORCH" in globals():
return "torch"
else:
return "numpy"
def _grad(f):
raise NotImplementedError(f"grad is not implemented for {BACKEND} backend.")
if "_SCIPY" in globals():
import numpy
grad = _grad
backend = numpy
elif "_AUTOGRAD" in globals():
from autograd import grad, numpy
backend = numpy
elif "_JAX" in globals():
if HAS_JAX:
import jax
if parse_version(jax.__version__) >= parse_version("0.4.16"):
jax.numpy.trapz = jax.scipy.integrate.trapezoid
if DEVICE == "cpu":
os.environ["CUDA_VISIBLE_DEVICES"] = ""
jax.config.update("jax_platform_name", "cpu")
else:
# os.environ.pop("CUDA_VISIBLE_DEVICES", None)
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
jax.config.update("jax_platform_name", "gpu")
jax.config.update("jax_enable_x64", True)
# TODO: jax eig not implemented on GPU
# see https://github.com/google/jax/issues/1259
# TODO: support jax properly (is it faster than autograd? use jit?)
# jax does not support eig
# for autodif wrt eigenvectors yet.
# see: https://github.com/google/jax/issues/2748
# from jax import grad, numpy
from jax import numpy
grad = _grad
backend = numpy
else:
logger.warning("jax not found. Falling back to default numpy backend.")
set_backend("numpy")
elif "_TORCH" in globals():
if HAS_TORCH:
import numpy
import torch
# torch.set_default_tensor_type(torch.cuda.FloatTensor)
backend = torch
def _array(a, **kwargs):
if isinstance(a, backend.Tensor):
return a.to(torch.device(DEVICE))
else:
return backend.tensor(a, **kwargs).to(torch.device(DEVICE))
backend.array = _array
def grad(f):
def df(x, *args, **kwargs):
x = backend.array(x, dtype=bk.float64)
_x = x.clone().detach().requires_grad_(True)
out = backend.autograd.grad(
f(_x, *args, **kwargs), _x, allow_unused=True
)[0]
return out
return df
else:
logger.warning("pytorch not found. Falling back to default numpy backend.")
set_backend("numpy")
else:
import numpy
grad = _grad
backend = numpy
def get_device():
return "cuda" if "_GPU_DEVICE" in globals() else "cpu"
def use_32_bit(yes):
global _USE_32
if yes:
_USE_32 = True
else:
_delvar("_USE_32")
_reload_package()
def get_types():
return (
(backend.float32, backend.complex64)
if "_USE_32" in globals()
else (backend.float64, backend.complex128)
)
BACKEND = get_backend()
DEVICE = get_device()
FLOAT, COMPLEX = get_types()
_backend_env_var = os.environ.get("NANNOS_BACKEND")
_gpu_env_var = os.environ.get("NANNOS_GPU")
if (
_backend_env_var in available_backends
and _backend_env_var is not None
and BACKEND != _backend_env_var
and "_FORCE_BACKEND" not in globals()
):
logger.debug(f"Found environment variable NANNOS_BACKEND={_backend_env_var}")
set_backend(_backend_env_var)
if _gpu_env_var is not None and "_FORCE_GPU" not in globals():
logger.debug(f"Found environment variable NANNOS_GPU={_gpu_env_var}")
use_gpu(True)
from . import optimize
from .constants import *
from .excitation import *
from .lattice import *
from .parallel import *
from .sample import *
from .simulation import *
from .utils import *
array = backend.array