Add numerical stabilization for difference of exponentials

[Smit-create]

Closes #1303.

[Smit-create]

I'm unable to test the changes locally because of some lock issues in macOS. If I run the following command:

% pytest -v tests/tensor/test_math.py::test_log1mexp_grad_lim

Timeout Error

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <filelock._unix.UnixFileLock object at 0x1368913c0>, timeout = 120
poll_interval = 0.05

    def acquire(
        self,
        timeout: float | None = None,
        poll_interval: float = 0.05,
        *,
        poll_intervall: float | None = None,
        blocking: bool = True,
    ) -> AcquireReturnProxy:
        """
        Try to acquire the file lock.
    
        :param timeout: maximum wait time for acquiring the lock, ``None`` means use the default :attr:`~timeout` is and
         if ``timeout < 0``, there is no timeout and this method will block until the lock could be acquired
        :param poll_interval: interval of trying to acquire the lock file
        :param poll_intervall: deprecated, kept for backwards compatibility, use ``poll_interval`` instead
        :param blocking: defaults to True. If False, function will return immediately if it cannot obtain a lock on the
         first attempt. Otherwise this method will block until the timeout expires or the lock is acquired.
        :raises Timeout: if fails to acquire lock within the timeout period
        :return: a context object that will unlock the file when the context is exited
    
        .. code-block:: python
    
            # You can use this method in the context manager (recommended)
            with lock.acquire():
                pass
    
            # Or use an equivalent try-finally construct:
            lock.acquire()
            try:
                pass
            finally:
                lock.release()
    
        .. versionchanged:: 2.0.0
    
            This method returns now a *proxy* object instead of *self*,
            so that it can be used in a with statement without side effects.
    
        """
        # Use the default timeout, if no timeout is provided.
        if timeout is None:
            timeout = self.timeout
    
        if poll_intervall is not None:
            msg = "use poll_interval instead of poll_intervall"
            warnings.warn(msg, DeprecationWarning, stacklevel=2)
            poll_interval = poll_intervall
    
        # Increment the number right at the beginning. We can still undo it, if something fails.
        with self._thread_lock:
            self._lock_counter += 1
    
        lock_id = id(self)
        lock_filename = self._lock_file
        start_time = time.monotonic()
        try:
            while True:
                with self._thread_lock:
                    if not self.is_locked:
                        _LOGGER.debug("Attempting to acquire lock %s on %s", lock_id, lock_filename)
                        self._acquire()
    
                if self.is_locked:
                    _LOGGER.debug("Lock %s acquired on %s", lock_id, lock_filename)
                    break
                elif blocking is False:
                    _LOGGER.debug("Failed to immediately acquire lock %s on %s", lock_id, lock_filename)
                    raise Timeout(self._lock_file)
                elif 0 <= timeout < time.monotonic() - start_time:
                    _LOGGER.debug("Timeout on acquiring lock %s on %s", lock_id, lock_filename)
>                   raise Timeout(self._lock_file)
E                   filelock._error.Timeout: The file lock 'The file lock '/Users/thebigbool/.aesara/compiledir_macOS-13.0-x86_64-i386-64bit-i386-3.10.8-64/.lock' could not be acquired.
E                   Apply node that caused the error: Elemwise{Composite{Switch(IsInf(Composite{(i0 / expm1((-i1)))}(i0, i1)), i2, Composite{(i0 / expm1((-i1)))}(i0, i1))}}(TensorConstant{-1.0}, x, TensorConstant{-inf})
E                   Toposort index: 0
E                   Inputs types: [TensorType(float64, ()), TensorType(float64, ()), TensorType(float32, ())]
E                   
E                   HINT: Use a linker other than the C linker to print the inputs' shapes and strides.
E                   HINT: Re-running with most Aesara optimizations disabled could provide a back-trace showing when this node was created. This can be done by setting the Aesara flag 'optimizer=fast_compile'. If that does not work, Aesara optimizations can be disabled with 'optimizer=None'.
E                   HINT: Use the Aesara flag `exception_verbosity=high` for a debug print-out and storage map footprint of this Apply node.' could not be acquired.

../../opt/anaconda3/envs/aesara-dev/lib/python3.10/site-packages/filelock/_api.py:183: Timeout
============================= slowest 50 durations =============================
120.03s call     tests/tensor/test_math.py::test_log1mexp_grad_lim

(2 durations < 0.005s hidden.  Use -vv to show these durations.)
=========================== short test summary info ============================
FAILED tests/tensor/test_math.py::test_log1mexp_grad_lim - filelock._error.Timeout: The file lock 'The file lock '/Users/thebigbool/.a...
======================== 1 failed in 140.14s (0:02:20) =========================

[Smit-create]

Seems like some OS issue as the tests already run on CI. Any pointers on this @brandonwillard?

[brandonwillard]

Seems like some OS issue as the tests already run on CI. Any pointers on this @brandonwillard?

I would first check that ~/.aesara—or wherever your Aesara cache directory (i.e. base_compiledir) is located—isn't pointing to a remote directory or something similar that could trip up filelock.

Does this happen literally every time something is compiled in Aesara, or only when that particular test is run?

[Smit-create]

Does this happen literally every time something is compiled in Aesara, or only when that particular test is run?

This happens to me whenever I try to run the tests.

[brandonwillard]

Does this happen literally every time something is compiled in Aesara, or only when that particular test is run?

This happens to me whenever I try to run the tests.

Very interesting! Don't fix anything just yet; I would like to get on a call first and check it out. If it's an Aesara bug, this would be a great opportunity to squash it.

In the meantime, you can always test things using the Python backend by setting the Aesara cxx setting to an empty string.

[codecov]

Codecov Report

Merging #1399 (8cbed4d) into main (b64cb85) will increase coverage by 0.00%.
The diff coverage is 80.95%.

❗ Current head 8cbed4d differs from pull request most recent head 5fbbc8a. Consider uploading reports for the commit 5fbbc8a to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1399   +/-   ##
=======================================
  Coverage   74.69%   74.69%           
=======================================
  Files         194      194           
  Lines       49730    49751   +21     
  Branches    10527    10532    +5     
=======================================
+ Hits        37145    37162   +17     
- Misses      10262    10264    +2     
- Partials     2323     2325    +2     

Impacted Files	Coverage Δ
aesara/tensor/rewriting/math.py	85.95% <80.95%> (-0.07%)	⬇️

[brandonwillard]

Does this happen literally every time something is compiled in Aesara, or only when that particular test is run?

This happens to me whenever I try to run the tests.

Very interesting! Don't fix anything just yet; I would like to get on a call first and check it out. If it's an Aesara bug, this would be a great opportunity to squash it.

In the meantime, you can always test things using the Python backend by setting the Aesara cxx setting to an empty string.

This was solved by re-constructing the environment using the environment-arm.yml discussed in #909.

[brandonwillard]

Looks good; we just need to squash before/while merging.

[rlouf]

Looks good, merging!