A simple decorator to make your classes hickle-able.
hickleable
provides a simple decorator for your classes that will almost always
make them serilalize well using the excellent
hickle package. By default, custom classes are
not supported by hickle
-- instead, they are written to the HDF5 file as a binary
dataset that is serialized using the standard Python pickle
. This obviously negates
much of the benefit of hickle
, for example, the fact that pickle-serialized data is
only readable using Python.
hickle
provides a way to serialize your custom classes using the HDF5 format, via
defining a few hooks for loading/dumping. However, it can be a little tricky to
implement these hooks, as they are quite general.
hickleable
provides a "default implementation" of these hooks that should satisfy
the requirements of most custom classes, and can be applied as a simple decorator.
This makes it a one-liner to transform your class into a well-supported data format.
Check out the docs at ReadTheDocs.
Simply pip install hickleable
. Conda-installable dependencies include h5py
.
Simply:
from hickleable import hickleable @hickleable() class MyClass: def __init__(self, a=1, b='foo', c={'a': 'dict'}): self.a = a self.b = b self.c = c
Now, MyClass
can be hickled without any pickling:
import hickle my_obj = MyClass() hickle.dump(my_obj, 'temporary_file.h5') # Note: no warnings about having to pickle new_obj = hickle.load('temporary_file.h5')
One super cool thing is that @cached_property
attributes are respected, and
dataclasses are also supported:
from dataclass import dataclass from functools import cached_property @hickleable() @dataclass class CachedClass: foo: str bar: int @cached_property def foobar(self) -> str: print("Obtaining foobar...") return foo*bar c = CachedClass('baz', 50000) foobar = c.foobar # prints "Obtaining foobar..." foobar = c.foobar # prints nothing, since it's returning cached value. hickle.dump(c, 'foobar.h5') d = hickle.load('foobar.h5') d_foobar = d.foobar # prints nothing! The value is cached in the hickle file.
One thing to note is that the cached properties are only saved in the hickle file if
they have already been evaluated. To force hickle
to write out all cached
properties, use the evaluate_cached_properties=True
parameter in the call to
hickleable()
.
While hickleable
will automatically render most classes hickle-able, there are bound
to be corner cases in which constituent attributes are not themselves hickleable, or
other concerns that you will want to customize. While all of this is of course
totally customizable by using the dumping/loading hooks from hickle
, the
hickleable
decorator also respects the magic methods __gethstate__
and
__sethstate__
, which act exactly like __getstate__
and __setstate__
do for
pickling. In fact, if the latter exist and the former don't, the latter will be used to
serialize the object in hickle
. For instance, let's say you have a class that keeps
track of the number of times it is called in its lifecycle:
@hickleable() class Counter: def __init__(self, a): self.a = a self._counts = 0 def __call__(self, b): self._counts += 1 self.a *= b
If we make an instance and call it a few times, the _counts
attribute is larger than
zero. If we save the object to a hickle file and load it back up somewhere else, it will
start with _counts > 0
. We can avoid this as follows:
def ignore_counts(self, state: dict): state['_counts'] = 0 self.__dict__.update(state) Counter.__setstate__ = ignore_counts
We could also have removed _counts
entirely from the hickle file:
def remove_counts(self) -> dict: return {k: v for k, v in self.__dict__.items() if k != '_counts'} Counter.__gethstate__ = remove_counts
Note that since we set ignore_counts
to be the __setstate__
method, it will be
respected both for hickle
and pickle
. We set remove_counts
as the
__gethstate__
method, which means it will only be respected for hickle
.