[tuner] Regularized Evolution #2802
Conversation
|
This tuner has been added as builtin tuners. |
|
Tuner behavior should be adjusted to be consistent with #2695 . |
| n_chosen = val['_value']['n_chosen'] | ||
| if n_chosen == None: | ||
| raise RuntimeError('Key n_chosen must be set in InputChoice.') | ||
| idxs = [random.randint(0, len(candidates) - 1) for _ in range(n_chosen)] |
There was a problem hiding this comment.
This could generate duplicate models.
There was a problem hiding this comment.
What if all models are evaluated? Are random models generated or throwing a NO_MORE_TRIAL exception?
docs/en_US/NAS/classic_nas.rst
Outdated
| Classic NAS Algorithms | ||
| ======================= | ||
|
|
||
| Classic NAS algorithms use a controller to ensemble a model from search space. The classic tuner algorithms train one model each time and use the reward of the model to guide the latter models. NNI has supported many popular classic NAS algorithms as following. |
There was a problem hiding this comment.
This part may need refactoring.
tabVersion
force-pushed
the
evo-nas-tuner
branch
from
e109c3c
to
11f375e
Compare
src/sdk/pynni/nni/regularized_evolution_tuner/regularized_evolution_tuner.py
Outdated
Show resolved
Hide resolved
| self.history[parameter_id] = arch | ||
| return arch | ||
| elif self.population: | ||
| sample = [] |
There was a problem hiding this comment.
Looks like the logic here is:
- Find a candidate from a candidate pool, which is actually a queue with size
population_size. - Mutate the candidate, send it.
- Receive the result of candidate, push it into candidate pool.
It's not evolution that I remember, which is conducted in "batch", i.e., a candidate of models are sent and received, then produce a new batch of candidates.
There was a problem hiding this comment.
Actually, it is a little different here.
Every time to generate a new model after warmup:
- randomly select
sample_sizemodels frompopulationassample(they remain inpopulation) - choose the model with highest acc in sample
- mutate the model then send it & push to
populationafter getting the result
src/sdk/pynni/nni/regularized_evolution_tuner/regularized_evolution_tuner.py
Show resolved
Hide resolved
src/sdk/pynni/nni/regularized_evolution_tuner/regularized_evolution_tuner.py
Show resolved
Hide resolved
docs/en_US/NAS/classic_nas.rst
Outdated
| @@ -0,0 +1,11 @@ | |||
| Classic NAS Algorithms | |||
There was a problem hiding this comment.
remove this file and put the content in RegularizedEvolutionTuner.md to ClassicNas.md
| self.population = deque() | ||
| self.history = {} | ||
| self.search_space = None | ||
| self._from = {} |
There was a problem hiding this comment.
_from -> _from_initial_population
No description provided.