# Copyright (c) 2020 zfit
import copy
import functools
import math as _mt
from collections import defaultdict
from typing import Any, Callable
import numpy as np
import tensorflow as tf
from ..settings import ztypes
[docs]def constant(value, dtype=ztypes.float, shape=None, name="Const", verify_shape=None):
# TODO(tf2): remove this legacy thing below
if verify_shape is not None:
raise RuntimeError("'verify_shape' is not a valid argument anymore. It's always true. Please remove.")
return tf.constant(value, dtype=dtype, shape=shape, name=name)
pi = np.float64(_mt.pi)
[docs]def to_complex(number, dtype=ztypes.complex):
return tf.cast(number, dtype=dtype)
[docs]def to_real(x, dtype=ztypes.float):
return tf.cast(x, dtype=dtype)
[docs]def abs_square(x):
return tf.math.real(x * tf.math.conj(x))
[docs]def nth_pow(x, n, name=None):
"""Calculate the nth power of the complex Tensor x.
Args:
x (tf.Tensor, complex):
n (int >= 0): Power
name (str): No effect, for API compatibility with tf.pow
"""
if not n >= 0:
raise ValueError("n (power) has to be >= 0. Currently, n={}".format(n))
power = to_complex(1.)
for _ in range(n):
power *= x
return power
[docs]def unstack_x(value: Any, num: Any = None, axis: int = -1, always_list: bool = False, name: str = "unstack_x"):
"""Unstack a Data object and return a list of (or a single) tensors in the right order.
Args:
value ():
num (Union[]):
axis (int):
always_list (bool): If True, also return a list if only one element.
name (str):
Returns:
Union[List[tensorflow.python.framework.ops.Tensor], tensorflow.python.framework.ops.Tensor, None]:
"""
if isinstance(value, list):
if len(value) == 1 and not always_list:
value = value[0]
return value
try:
return value.unstack_x(always_list=always_list)
except AttributeError:
unstacked_x = tf.unstack(value=value, num=num, axis=axis, name=name)
if len(unstacked_x) == 1 and not always_list:
unstacked_x = unstacked_x[0]
return unstacked_x
[docs]def stack_x(values, axis: int = -1, name: str = "stack_x"):
return tf.stack(values=values, axis=axis, name=name)
# random sampling
[docs]def convert_to_tensor(value, dtype=None, name=None, preferred_dtype=None):
return tf.convert_to_tensor(value=value, dtype=dtype, name=name, dtype_hint=preferred_dtype)
[docs]def safe_where(condition: tf.Tensor, func: Callable, safe_func: Callable, values: tf.Tensor,
value_safer: Callable = tf.ones_like) -> tf.Tensor:
"""Like :py:func:`tf.where` but fixes gradient `NaN` if func produces `NaN` with certain `values`.
Args:
condition (:py:class:`tf.Tensor`): Same argument as to :py:func:`tf.where`, a boolean :py:class:`tf.Tensor`
func (Callable): Function taking `values` as argument and returning the tensor _in case
condition is True_. Equivalent `x` of :py:func:`tf.where` but as function.
safe_func (Callable): Function taking `values` as argument and returning the tensor
_in case the condition is False_, Equivalent `y` of :py:func:`tf.where` but as function.
values (:py:class:`tf.Tensor`): Values to be evaluated either by `func` or `safe_func` depending on
`condition`.
value_safer (Callable): Function taking `values` as arguments and returns "safe" values
that won't cause troubles when given to`func` or by taking the gradient with respect
to `func(value_safer(values))`.
Returns:
:py:class:`tf.Tensor`:
"""
safe_x = tf.where(condition=condition, x=values, y=value_safer(values))
result = tf.where(condition=condition, x=func(safe_x), y=safe_func(values))
return result
[docs]def run_no_nan(func, x):
from zfit.core.data import Data
value_with_nans = func(x=x)
if value_with_nans.dtype in (tf.complex128, tf.complex64):
value_with_nans = tf.math.real(value_with_nans) + tf.math.imag(value_with_nans) # we care only about NaN or not
finite_bools = tf.math.is_finite(tf.cast(value_with_nans, dtype=tf.float64))
finite_indices = tf.compat.v1.where(finite_bools)
new_x = tf.gather_nd(params=x, indices=finite_indices)
new_x = Data.from_tensor(obs=x.obs, tensor=new_x)
vals_no_nan = func(x=new_x)
result = tf.scatter_nd(indices=finite_indices, updates=vals_no_nan,
shape=tf.shape(input=value_with_nans, out_type=finite_indices.dtype))
return result
# tf_function_deco = tf.function(autograph=False, experimental_relax_shapes=True)
[docs]class FunctionWrapperRegistry:
wrapped_functions = []
registries = []
do_jit = True
[docs] @classmethod
def check_wrapped_functions_registered(cls):
return all((func.zfit_graph_cache_registered for func in cls.wrapped_functions))
def __init__(self, **kwargs_user) -> None:
"""`tf.function`-like decorator with additional cache-invalidation functionality.
Args:
**kwargs_user: arguments to `tf.function`
"""
super().__init__()
self._initial_user_kwargs = kwargs_user
self.registries.append(self)
self.reset(**self._initial_user_kwargs)
# self.inside_tracing = False
self.currently_traced = set()
[docs] def reset(self, **kwargs_user):
kwargs = dict(autograph=False, experimental_relax_shapes=True)
kwargs.update(self._initial_user_kwargs)
kwargs.update(kwargs_user)
self.tf_function = tf.function(**kwargs)
self.function_cache = defaultdict(list)
def __call__(self, func):
wrapped_func = self.tf_function(func)
cache = self.function_cache[func]
from zfit.util.cache import FunctionCacheHolder
def call_correct_signature(func, args, kwargs):
if args == [] and kwargs != {}:
return func(**kwargs)
elif args != [] and kwargs == {}:
return func(*args)
elif args == [] and kwargs == {}:
return func()
elif args != [] and kwargs != {}:
return func(*args, **kwargs)
def concrete_func(*args, **kwargs):
if not self.do_jit or func in self.currently_traced:
return call_correct_signature(func, args, kwargs)
# self.inside_tracing = True
self.currently_traced.add(func)
nonlocal wrapped_func
function_holder = FunctionCacheHolder(func, wrapped_func, args, kwargs)
try:
func_holder_index = cache.index(function_holder)
except ValueError: # not in cache
cache.append(function_holder)
else:
func_holder_cached = cache[func_holder_index]
if func_holder_cached.is_valid:
function_holder = func_holder_cached
else:
wrapped_func = self.tf_function(func) # update nonlocal wrapped function
function_holder = FunctionCacheHolder(func, wrapped_func, args, kwargs)
cache[func_holder_index] = function_holder
func_to_run = function_holder.wrapped_func
result = call_correct_signature(func_to_run, args, kwargs)
self.currently_traced.remove(func)
return result
return concrete_func
FunctionWrapperRegistry2 = copy.deepcopy(FunctionWrapperRegistry)
# FunctionWrapperRegistry2.do_jit = True
# FunctionWrapperRegistry.do_jit = False
tf_function = FunctionWrapperRegistry()
function_tf = FunctionWrapperRegistry2() # for only tensorflow inside
function_sampling = tf_function
# py_function = tf.py_function
@functools.wraps(tf.py_function)
def py_function(func, inp, Tout, name=None):
from .. import settings
if not settings.options['numerical_grad']:
raise RuntimeError("Running a py_function without using the numerical gradient will result in wrong gradient"
" calculation. Will be more fine-grained in the future. To switch to numerical calculation"
" (even if the gradients are not calculated at all), do"
" `zfit.settings.options['numerical_grad'] = True`")
return tf.py_function(func=func, inp=inp, Tout=Tout, name=name)