# Copyright (c) 2020 zfit
import contextlib
import multiprocessing
import os
import sys
import warnings
from typing import List, Union, Optional
import numpy as np
import tensorflow as tf
from .legacy import deprecated
from .container import DotDict, is_container
[docs]class RunManager:
DEFAULT_MODE = {'graph': 'auto',
'autograd': True}
def __init__(self, n_cpu='auto'):
"""Handle the resources and runtime specific options. The `run` method is equivalent to `sess.run`"""
self.MAX_CHUNK_SIZE = sys.maxsize
self.chunking = DotDict()
self._cpu = []
self._n_cpu = None
self._inter_cpus = None
self._intra_cpus = None
self._strict = False
self.numeric_checks = True
self._mode = self.DEFAULT_MODE.copy()
self.set_n_cpu(n_cpu=n_cpu)
# HACK
self._enable_parameter_autoconversion = True
# HACK END
# set default values
self.chunking.active = False # not yet implemented the chunking...
self.chunking.max_n_points = 1000000
@property
def mode(self):
return self._mode
@property
def chunksize(self):
if self.chunking.active:
return self.chunking.max_n_points
else:
return self.MAX_CHUNK_SIZE
@property
def n_cpu(self):
return len(self._cpu)
[docs] def set_n_cpu(self, n_cpu: Union[str, int] = 'auto', strict: bool = False) -> None:
"""Set the number of cpus to be used by zfit. For more control, use `set_cpus_explicit`.
Args:
n_cpu: Number of cpus, will be the number for inter-op parallelism
strict: If strict, sets intra parallelism to 1
"""
if n_cpu == 'auto':
try:
cpu = sorted(os.sched_getaffinity(0))
except AttributeError:
cpu = range(multiprocessing.cpu_count())
warnings.warn("Not running on Linux. Determining available cpus for thread can fail"
"and be overestimated. Workaround (only if too many cpus are used):"
"`zfit.run.set_n_cpu(your_cpu_number)`")
elif isinstance(n_cpu, int):
cpu = range(n_cpu)
self._cpu = ['dummy_cpu{}'.format(i) for i in cpu]
n_cpu = len(cpu)
if strict ^ self._strict:
intra = 1 if strict else 2
inter = n_cpu
self.set_cpus_explicit(intra=intra, inter=inter)
[docs] def set_cpus_explicit(self, intra: int, inter: int) -> None:
"""Set the number of threads (cpus) used for inter-op and intra-op parallelism
Args:
intra: Number of threads used to perform an operation. For larger operations, e.g. large Tensors, this
is usually beneficial to have >= 2.
inter: Parallelization on the level of ops. This is beneficial, if many operations can be computed
independently in parallel.
"""
try:
tf.config.threading.set_intra_op_parallelism_threads(intra)
tf.config.threading.set_inter_op_parallelism_threads(inter)
self._n_cpu = inter + intra
except RuntimeError as err:
raise RuntimeError("Cannot set the number of cpus after initialization, has to be at the beginning."
f" Original message: {err}")
[docs] @contextlib.contextmanager
def aquire_cpu(self, max_cpu: int = -1) -> List[str]:
if isinstance(max_cpu, int):
if max_cpu < 0:
max_cpu = max((self.n_cpu + 1 + max_cpu, 0)) # -1 means all
if max_cpu == 0:
cpu = []
else:
n_cpu = min((max_cpu, self.n_cpu))
cpu = self._cpu[-n_cpu:]
self._cpu = self._cpu[:-n_cpu]
yield cpu
self._cpu.extend(cpu)
def __call__(self, *args, **kwargs):
if kwargs:
raise RuntimeError("Why kwargs provided?")
flattened_args = tf.nest.flatten(args)
evaluated_args = [eval_object(arg) for arg in flattened_args]
values = tf.nest.pack_sequence_as(args, flat_sequence=evaluated_args)
was_container = is_container(args[0]) and not isinstance(args[0], np.ndarray, )
if not was_container and values:
values = values[0]
return values
[docs] @staticmethod
@deprecated(date=None, instructions="Use `set_mode(graph=False)`")
def experimental_enable_eager(eager: bool = False):
"""DEPRECEATED! Enable eager makes tensorflow run like numpy. Useful for debugging.
Do NOT directly mix it with Numpy (and if, also enable the numberical gradient).
This can BREAK in the future.
"""
from .graph import jit
jit._set_all(not eager)
[docs] def set_mode(self, graph: Optional[Union[bool, str, dict]] = None, autograd: Optional[bool] = None):
"""Set the policy for graph building and the usage of automatic vs numerical gradients.
zfit runs on top of TensorFlow, a modern, powerful computing engine very similar in design to Numpy.
An interactive tutorial can be found at https://github.com/zfit/zfit-tutorials
**Graph building**
It has two ways to be run where the first
defaults to the normal mode we are in except inside a :py:func:`~zfit.z.function` decorated function. Setting
the mode allows to control the behavior of decorated functions to not always trigger a graph building.
- **numpy-like/eager**: in this mode, the syntax slightly differs from pure numpy but is similar. For example,
`tf.sqrt`, `tf.math.log` etc. The return values are `EagerTensors` that represent "wrapped Numpy arrays" and
can directly be used with any Numpy function. They can explicitly be converted to a Numpy array with
`zfit.run(EagerTensor)`, which takes also care of nested structures and already existing `np.ndarrays`,
or just a `.numpy()` method.
The difference to Numpy is that TensorFlow tries to optimize the calculation slightly beforehand and may
also executes on the GPU. This will result in a slight performance penalty for *very small* computations
compared to Numpy, on the other hand an improved performance for larger computations.
- **graph**: a function can be decorated with :py:func:`~zfit.z.function`, which will *not* execute its
content immediately, but first trace it and build a graph. This is done by *recording all `tf.*` operations
and adding them to the graph while any Python operation, e.g. `np.random.*` will use a fixed value added
to the graph. Building a graph greatly reduces the flexibility, since only `tf.*` operations can
effectively be used to have dynamics in there, on the other hand it can greatly increase the performance.
When the graph is built, it is cached (for later re-use), optimized and then executed.
Calling a `tf.function` decorated
function does therefore not make an actualy difference *for the caller*. But it is a difference on how the
function behaves.
.. code-block:: python
@z.function
def add_rnd(x):
res1 = x + np.random.uniform() # returns a Python scalar. This exact scalar will be constant
res2 = x + z.random.uniform(shape=()) # returns a tf.Tensor. This will be flexible
return res1, res2
res_np1, res_tf1 = add_rnd(5)
res_np2, res_tf2 = add_rnd(5)
assert res_np1 == res_np2 # they will be the same!
assert res_tf1 != res_tf2 # these differ
While writing TensorFlow is just like Numpy, if we build a graph, only `tf.*` dynamics "survives".
Important: while values are usually constant, changing a :py:class:`zfit.Parameter` value with
:py:meth:`~zfit.Parameter.set_value(...)` *will* change the value in the graph as well.
.. code-block:: python
@z.function
def add(x, param):
return x + param
param = zfit.Parameter('param1', 36)
assert add_rnd(5, param) == 41
param.set_value(6)
assert add_rnd(5, param) == 42 # the value changed!
Every graph generation takes some additional time and is stored, consuming memory and slowing down the
overall execution process.
To clear all caches and force a rebuild of the graph, `zfit.run.clear_graph_cache()` can be used.
If a function is not decorated with `z.function`, this does not guarantee that it is executed in eager,
as an outer function may uses a decorator. A typical case is the loss, which is decorated. Therefore,
any Model called inside will be evaluated with a graph building first.
**When to use what**:
- Any repeated call (as a typical call to the loss function in the minimization process) is usually
better suited within a `z.function`.
- A single call (e.g. for plotting) or repeated calls *with different arguments* should rather be run
*without* a graph built first
- Debugging is usually way easier without graph building. Therefore, set the graph mode to `False`
- If the minimization fails but the pdf works without graph, maybe the graph mode can be switched on
for everything to have the same behavior in the pdf as when the loss is called.
**automatic gradient**
Another strong feature of TensorFlow is the possibility to derive an analytic expression for the gradient
by successively applying the chain rule to all of its operations. This is *independent* of whether the code
is run in graph or eager execution, but requires all operations that are dynamic to be `tf.*` operations.
For example, multiplying by a constant (constant as in *not chaning ever*) does not require the constant to
be a `tf.constant(...)` but can be a Python scalar. For example, it is also fine to use a fixed template shape
using Numpy (Scipy etc), as the template shape will stay constant (this requires though to use a
`z.py_function` to work, but this is another story about graph mode or not).
To allow to have dynamic numpy operations in a component, preferably wrapped with `z.py_function` instead of
forced eager, and to still retrieve a meaningful gradient, a numerical gradient has to be used.
In general, this can be achieved by setting the `autograd` to False. Any derivative received will then be
numerically computed. Furthermore, some minimizers (e.g. :py:class:`~zfit.minimize.Minuit`) have their own way
of calculating gradients, which can be faster.
Disabling `autograd` and using the zfit builting numerical way of calculating the gradients and hessian can
be less stable and may raises errors.
Args:
graph: Policy for when to build a graph with which function. Currently allowed values are
- `True`: this will make all :py:func:`zfit.z.function` decorated function to be traced. Useful
to have a consistent behavior overall, as e.g. a PDF may not be traced if `pdf` or `integrate` is
called, but may be traced when inside a loss.
- `False`: this will make everything execute immediately, like Numpy (this is **not enough** to be
fully Numpy compatible in the sense of using `, also see the `autograd` option)
- 'auto': Something in between, where sampling (currently) and the loss builds a graph but
all model methods, such as `pdf`, `integrate` (except of `*sample*`) do not and are executed
eagerly.
- (**advanced and experimental!**): a dictionary containing the string of a wrapped function identifier
(see also :py:func:`~zfit.z.function` for more information about this) with a boolean that switches
explicitly on/off the graph building for this type of decorated functions.
autograd: Whether the automatic gradient feature of TensorFlow should be used or a numerical procedure
instead. If any non-constant Python (numpy, scipy,...) code is used inside, this should be switched on.
"""
jit_mode = graph
from .graph import jit as jit_obj
if graph is None and autograd is None:
raise ValueError("Both graph and autograd are None. Specify at least one.")
if jit_mode is True:
jit_obj._set_all(True)
elif jit_mode is False:
jit_obj._set_all(False)
elif jit_mode == 'auto':
jit_obj._set_default()
elif isinstance(jit_mode, dict):
jit_obj._update_allowed(jit_mode)
elif jit_mode is not None:
raise ValueError(f"{jit_mode} is not a valid keyword to the `jit` behavior. Use either "
f"True, False, 'default' or a dict. You can read more about it in the docs.")
if jit_mode is not None:
self._mode['graph'] = graph
if autograd is not None:
from zfit import settings
settings.options.numerical_grad = not autograd
self._mode['autograd'] = autograd
[docs] def current_policy_graph(self) -> Union[bool, str]:
"""Return the current policy for graph building.
Returns:
bool, str: The current policy. For more information, check :py:meth:`~zfit.run.set_mode`.
"""
return self.mode['graph']
[docs] def current_policy_autograd(self) -> bool:
"""The current policy for using the automatic gradient or falling back to the numerical
Returns:
bool: If autograd is being used.
"""
return self.mode['autograd']
[docs] def set_mode_default(self):
"""Reset the mode to the default of `graph` = 'auto' and `autograd` = True."""
self.set_mode(**self.DEFAULT_MODE)
[docs] def clear_graph_cache(self):
"""Clear all generated graphs and effectively reset. Should not affect execution, only performance.
In a simple fit scenario, this is not used. But if several fits are performed with different python objects
such as a scan over a range (by changing the norm_range and creating a new dataset), doing minimization and
therefore invoking the loss (by default creating a graph) will leave the graphs in the cache, even tough
the already scanned ranges are not needed anymore.
To clean, this function can be invoked. The only effect should be to speed up things, but should not have
any side-effects other than that.
"""
from zfit.util.cache import clear_graph_cache
clear_graph_cache()
[docs] def assert_executing_eagerly(self):
"""Assert that the execution is eager and Python side effects are taken into account.
This can be placed inside a model _in case python side-effects are necessary_ and no other way is possible.
"""
if not tf.executing_eagerly():
raise RuntimeError("This code is not supposed to run inside a graph.")
@property
@deprecated(None, "Use `current_policy_graph() is False`")
def experimental_is_eager(self):
return not self.mode['graph']
[docs] @deprecated(date=None, instructions="Use clear_graph_caches instead.")
def experimental_clear_caches(self):
"""DEPRECATED! Use `clear_graph_caches` instead.
"""
self.clear_graph_cache()
[docs]def eval_object(obj: object) -> object:
from zfit.core.parameter import BaseComposedParameter
if isinstance(obj, BaseComposedParameter): # currently no numpy attribute. Should we add this?
obj = obj.value()
if tf.is_tensor(obj):
return obj.numpy()
else:
return obj