Refactor algorithms observe/suggest to work on trial objects directly #686
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Why: The algorithms will be working with trial objects instead of tuple of values. This means the space needs to sample trials, and space transformations should be applied on trials as well instead of tuples of values. How: For simplicity, only interface of the space classes (TransformedSpace and ReshapedSpace) will be working with trials. The transformations per dimension will be applied using tuple of values so that, in particular, reshaping operations remain straightforward. To facilitate debugging, transformed trials are wrapped so that the original trial can still be accessible. This will prove handy in algorithms if we need access to original trial objects, and also because the ID of the transformed trial should always be based on the original parameters (otherwise the ID gets incoherent with the database).
The use of `__getattr__` to copy the TransformedTrial was causing an infinite recursion error.
Why: If the names of the dimensions have the same prefix, but some have a shape, a transformed space with have a different ordering of dimension. For example, the following dimensions will have their named sorted differently: dim (shape 2), dim1 (no shape) --> dim1, dim[0], dim[1] This is causing an issue when we try to restore the shape of the transformed dimension, with the names being swapped. How: When restoring shape, keep track of the original keys and their order, and reassign the restored dimensions to the correct index (correct dim name).
Reshaped transform on dim: Sometimes we need to apply a transformation on a dimension only. So far the transformation of View could only be applied on a full point, because the ReshapedDimension wrapping the view was expected to fetch the corresponding dimension at self.index. This commit pushes up the fetch at self.index at the level of the ReshapedSpace instead, the only place where we are expected to work on a full point, not only on a dimension. default value: Transformed dimensions should support default value like original dimensions.
Why:
The branching method will be used by algorithm generating new trials
based on existing one, such as Hyperband increasing the fidelity, and
later on Population Based Training which could also change the working
directory.
How:
When calling branch(), a new trial is created with same parameters as
current one, except for the parameter values passed to
(ex: ``branch(params={'some': 'values'})``).
bouthilx
changed the title
Why: The wrapper makes it cumbersome to keep a coherent interface with the base Trial object. Also it does not bring much value beside providing access to the original trial and its corresponding id. Algorithms can revert the trial to access the trial id. How: Return a trial object with params overwritten instead. Like it was already done when reverting a transformed trial anyway. Now base algo.get_id() method compute the trial hash based on the reverted trial.
Why: Deprecated and to be removed for v0.2.0. The newest versions regroup_dims and flatten_dims are also useless now due to the flattened space. They are also removed.
Why: When flattening the space, dims of shape (1, ) should be flattened as well otherwise the parameters will be a list of one element.
Why:
The deepcopy is failing on github-actions with error
`RuntimeError: dictionary changed size during iteration`. I have been
unable to reproduce the issue locally both with python 3.6 and 3.7. It
does fail on 3.7 on github-actions. Taking a copy of the dictionary to
do the deep copy should fix the issue only a dictionary inside some
trials is the source of the issue. The stack trace seams to hint towards
trials_info as the culprit however.
```
tests/unittests/benchmark/test_benchmark_client.py:345:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
src/orion/benchmark/__init__.py:90: in process
study.execute(n_workers)
src/orion/benchmark/__init__.py:341: in execute
experiment.workon(self.task, n_workers=n_workers, max_trials=max_trials)
src/orion/client/experiment.py:767: in workon
for _ in range(n_workers)
src/orion/executor/joblib_backend.py:32: in wait
return joblib.Parallel(n_jobs=self.n_workers)(futures)
.tox/py/lib/python3.7/site-packages/joblib/parallel.py:1056: in __call__
self.retrieve()
.tox/py/lib/python3.7/site-packages/joblib/parallel.py:935: in retrieve
self._output.extend(job.get(timeout=self.timeout))
/opt/hostedtoolcache/Python/3.7.12/x64/lib/python3.7/multiprocessing/pool.py:657: in get
raise self._value
/opt/hostedtoolcache/Python/3.7.12/x64/lib/python3.7/multiprocessing/pool.py:121: in worker
result = (True, func(*args, **kwds))
.tox/py/lib/python3.7/site-packages/joblib/_parallel_backends.py:595: in __call__
return self.func(*args, **kwargs)
.tox/py/lib/python3.7/site-packages/joblib/parallel.py:263: in __call__
for func, args, kwargs in self.items]
.tox/py/lib/python3.7/site-packages/joblib/parallel.py:263: in <listcomp>
for func, args, kwargs in self.items]
src/orion/client/experiment.py:781: in _optimize
with self.suggest(pool_size=pool_size) as trial:
src/orion/client/experiment.py:560: in suggest
trial = reserve_trial(self._experiment, self._producer, pool_size)
src/orion/client/experiment.py:54: in reserve_trial
producer.produce(pool_size)
src/orion/core/worker/producer.py:115: in produce
self.algorithm.set_state(self.naive_algorithm.state_dict)
src/orion/core/worker/primary_algo.py:47: in state_dict
return self.algorithm.state_dict
src/orion/algo/tpe.py:265: in state_dict
_state_dict = super(TPE, self).state_dict
src/orion/algo/base.py:132: in state_dict
return {"_trials_info": copy.deepcopy(self._trials_info)}
/opt/hostedtoolcache/Python/3.7.12/x64/lib/python3.7/copy.py:150: in deepcopy
y = copier(x, memo)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
x = {'0ef99dad51485ac9518b49c19b43f4ec': (Trial(experiment=2, status='completed', params=x[0]:0.5497,x[1]:2.397,x[2]:-4.29...params=x[0]:-0.8816,x[1]:2.087,x[2]:1.176), {'constraint': [], 'gradient': None, 'objective': 1187.240632192948}), ...}
memo = {140575410255504: datetime.datetime(2021, 11, 11, 22, 57, 32, 364584), 140575419437584: datetime.datetime(2021, 11, 11....datetime(2021, 11, 11, 22, 57, 32, 177829), 140575419438016: datetime.datetime(2021, 11, 11, 22, 57, 32, 241378), ...}
deepcopy = <function deepcopy at 0x7fdac6dfb680>
def _deepcopy_dict(x, memo, deepcopy=deepcopy):
y = {}
memo[id(x)] = y
> for key, value in x.items():
E RuntimeError: dictionary changed size during iteration
```
bouthilx
changed the title
Merged
bouthilx
added
the
enhancement
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
[Fixes #660]