# Copyright 2021 The NetKet Authors - All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import jax
import jax.numpy as jnp
import numpy as np
from numba import jit
from numba4jax import njit4jax
from netket.operator import AbstractOperator, DiscreteJaxOperator
from netket.utils import struct
from .base import MetropolisRule
class HamiltonianRuleBase(MetropolisRule):
"""
Rule proposing moves according to the terms in an operator.
"""
operator: AbstractOperator = struct.field(pytree_node=False)
"""The (hermitian) operator giving the transition amplitudes."""
def __init__(self, operator: AbstractOperator):
self.operator = operator
def init_state(rule, sampler, machine, params, key):
if sampler.hilbert != rule.operator.hilbert:
raise ValueError(
f"""
The hilbert space of the sampler ({sampler.hilbert}) and the hilbert space
of the operator ({rule.operator.hilbert}) for HamiltonianRule must be the same.
"""
)
return super().init_state(sampler, machine, params, key)
def __post_init__(self):
# Raise errors if hilbert is not an Hilbert
if not isinstance(self.operator, AbstractOperator):
raise TypeError(
"Argument to HamiltonianRule must be a valid operator, "
f"but operator is a {type(self.operator)}."
)
def __repr__(self):
return f"HamiltonianRuleNumba({self.operator})"
@struct.dataclass
class HamiltonianRuleNumba(HamiltonianRuleBase):
"""
Rule proposing moves according to the terms in an operator.
In this case, the transition matrix is taken to be:
.. math::
T( \\mathbf{s} \\rightarrow \\mathbf{s}^\\prime) = \\frac{1}{\\mathcal{N}(\\mathbf{s})}\\theta(|H_{\\mathbf{s},\\mathbf{s}^\\prime}|),
This rule only works on CPU! If you want to use it on GPU, you
must use the numpy variant :class:`netket.sampler.rules.HamiltonianRuleNumpy`
together with the numpy metropolis sampler :class:`netket.sampler.MetropolisSamplerNumpy`.
"""
def transition(rule, sampler, machine, parameters, state, key, σ):
"""
This implements the transition rule for `DiscreteOperator`s that are
implemented in Numba, relying on the `_get_conn_flattened_closure`
hack to make it work in numba.
"""
get_conn_flattened = rule.operator._get_conn_flattened_closure()
n_conn_from_sections = rule.operator._n_conn_from_sections
@njit4jax(
(
jax.core.ShapedArray(σ.shape, σ.dtype),
jax.core.ShapedArray((σ.shape[0],), σ.dtype),
)
)
def _transition(args):
# unpack arguments
v_proposed, log_prob_corr, v, rand_vec = args
log_prob_corr.fill(0)
sections = np.empty(v.shape[0], dtype=np.int32)
vp, _ = get_conn_flattened(v, sections)
_choose(vp, sections, rand_vec, v_proposed, log_prob_corr)
# TODO: n_conn(v_proposed, sections) implemented below, but
# might be slower than fast implementations like ising
get_conn_flattened(v_proposed, sections)
n_conn_from_sections(sections)
log_prob_corr -= np.log(sections)
# ideally we would pass the key to python/numba in _choose, initialise a
# np.random.default_rng(key) and use it to generate random uniform integers.
# However, numba dose not support np states, and reseeding it's MT1998 implementation
# would be slow so we generate floats in the [0,1] range in jax and pass those
# to python
rand_vec = jax.random.uniform(key, shape=(σ.shape[0],))
σp, log_prob_correction = _transition(σ, rand_vec)
return σp, log_prob_correction
def __repr__(self):
return f"HamiltonianRuleNumba({self.operator})"
@jit(nopython=True)
def _choose(vp, sections, rand_vec, out, w):
low_range = 0
for i, s in enumerate(sections):
n_rand = low_range + int(np.floor(rand_vec[i] * (s - low_range)))
out[i] = vp[n_rand]
w[i] = math.log(s - low_range)
low_range = s
@struct.dataclass
class HamiltonianRuleJax(HamiltonianRuleBase):
"""
Rule proposing moves according to the terms in an operator.
In this case, the transition matrix is taken to be:
.. math::
T( \\mathbf{s} \\rightarrow \\mathbf{s}^\\prime) = \\frac{1}{\\mathcal{N}(\\mathbf{s})}\\theta(|H_{\\mathbf{s},\\mathbf{s}^\\prime}|),
This rule only works with operators which are written in jax.
"""
def transition(self, _0, _1, _2, _3, key, x):
xp, mels = self.operator.get_conn_padded(x)
n_conn = self.operator.n_conn(x)
# def _n_conn(mels):
# nonzeros = jnp.abs(mels) > 0
# return nonzeros.sum(axis=-1)
# n_conn = _n_conn(mels)
rand_i = jax.random.randint(key, shape=(x.shape[0],), minval=0, maxval=n_conn)
# we need to shift rand_i so that we only select the nonzeros
#
# if we were to assume that the nonzero mels are gathered at the beginning of each
# row we could avoid this, i.e. just do
# x_proposed = xp[jnp.arange(xp.shape[0]), rand_i]
#
# instead of actually shifting we build a mask for the selected element in each row
# and sum over the zeros at the end
# TODO let the operator supply the nonzeros or nonzero_i_plus1 directly
nonzeros = jnp.abs(mels) > 0
nonzero_i_plus1 = (jnp.cumsum(nonzeros, axis=-1)) * nonzeros
rand_i_mask = nonzero_i_plus1 == jnp.expand_dims(rand_i + 1, -1)
x_proposed = (xp * jnp.expand_dims(rand_i_mask, -1)).sum(axis=1)
n_conn_proposed = self.operator.n_conn(x_proposed)
# _, mels_proposed = self.operator.get_conn_padded(x_proposed)
# n_conn_proposed = _n_conn(mels_proposed)
# if there are no connected elements x_proposed might
# contain illegal states (i.e. zeros)
# and log_prob_corr will be nan
# TODO is it safe to assume that n_conn >=1 ?
#
# no_nonzeros = nonzeros.sum(axis=-1) == 0
# x_proposed = jax.lax.select(
# jax.lax.broadcast_in_dim(no_nonzeros, x.shape, (0,)), x, x_proposed
# )
# n_conn = jnp.maximum(n_conn, 1)
# n_conn_proposed = jnp.maximum(n_conn_proposed, 1)
log_prob_corr = jnp.log(n_conn) - jnp.log(n_conn_proposed)
return x_proposed, log_prob_corr
def __repr__(self):
return f"HamiltonianRuleJax({self.operator})"
[docs]
def HamiltonianRule(operator):
r"""
Rule proposing moves according to the terms in an operator.
In this case, the transition matrix is taken to be:
.. math::
T( \mathbf{s} \rightarrow \mathbf{s}^\prime) =
\frac{1}{\mathcal{N}(\mathbf{s})}\theta(|H_{\mathbf{s},\mathbf{s}^\prime}|),
This is a thin wrapper on top of the constructors of :class:`netket.sampler.rules.HamiltonianRuleJax` and
:class:`netket.sampler.rules.HamiltonianRuleNumba`, which dispatches on one of the two implementations
depending on whether the operator specified is jax-compatible (:class:`netket.operator.DiscreteJaxOperator`)
or not.
"""
if isinstance(operator, DiscreteJaxOperator):
return HamiltonianRuleJax(operator)
else:
return HamiltonianRuleNumba(operator)
