Add capability to provide custom seeds to GP #502
Conversation
tpot/base.py
Outdated
| seed_individuals = [creator.Individual.from_string(x, self._pset) for x in seeds] | ||
| self._pop = [] | ||
|
|
||
| # Add the same set of seeds to the population until we have population_size seeds |
There was a problem hiding this comment.
With a little discussions with @rhiever earlier today, we think the seeds (maybe need better name for this parameter) should only specify a small part of initial population instead of the initial population that is full of duplicated pipelines in the seeds (If seeds is a small list, then the pipeline diversity of GP will be limited in the beginning). Other pipelines in initial population rather than pipelines in seeds should be randomly generated.
There was a problem hiding this comment.
population_seeds should probably be the parameter name.
| <p>TPOT allows for the initial population of pipelines to be seeded. This can be done either through the <code>population_seeds</code> parameter in the TPOT constructor, or through a <code>population_seeds</code> attribute in a custom config file.</p> | ||
| <pre><code class="Python">population_seeds = [ | ||
| 'BernoulliNB(GaussianNB(input_matrix), BernoulliNB__alpha=0.1, BernoulliNB__fit_prior=False)', | ||
| 'BernoulliNB(input_matrix, BernoulliNB__alpha=0.01, BernoulliNB__fit_prior=True)' |
There was a problem hiding this comment.
How easy would it be to take actual sklearn pipelines as input instead of the string representations? I sense that users would find it easier to seed with sklearn pipelines rather than this (slightly) weird string representation we use.
There was a problem hiding this comment.
Also, we should use better example clfs in the docs: Maybe a RandomForestClassifier and a LogisticRegression?
There was a problem hiding this comment.
How easy would it be to take actual sklearn pipelines as input instead of the string representations?
We don't have any code to go from sklearn pipelines to deap pipelines, only the other direction. It would be non-trivial to write, but doable given Python's complete reflection of objects.
we should use better example clfs in the docs: Maybe a RandomForestClassifier and a LogisticRegression?
Sure.
There was a problem hiding this comment.
It would be nice to have a function to go from sklearn pipelines to DEAP pipelines. :-) That will be one step closer to having DEAP work directly on sklearn pipelines themselves.
There was a problem hiding this comment.
Alright. I'll start working on that.
There was a problem hiding this comment.
How do you propose doing that? We need to make sure that it only takes in the operators, parameters, and parameter values that are defined in the config.
There was a problem hiding this comment.
Pass in the pset that's being used and check that the pipeline is valid (within the context of the pset) as you walk though it.
I'd throw a TypeError if a parameter is missing, or a parameter is used that's not specified. If an operator is used that doesn't exist in the pset I'd throw a NameError.
There was a problem hiding this comment.
@teaearlgraycold I think the function to go from sklearn pipeline to DEAP pipelines was not added in this PR yet. But I will start merging some PRs to dev branch today in case of tons of conflicts between PRs. Please add this function later in another PR.
| @@ -1068,7 +1081,7 @@ def _operator_count(self, individual): | |||
| return operator_count | |||
|
|
|||
| def _update_pbar(self, val, resulting_score_list): | |||
| """Update self._pbar during pipeline evaluration | |||
| """Update self._pbar during pipeline evaluration. | |||
There was a problem hiding this comment.
Not your typo, but there's a typo here in "evaluation."
| tpot_obj = TPOTRegressor(config_dict='tests/test_config.py') | ||
| n_seeds = len(tpot_obj._read_config_file('tests/test_config.py').population_seeds) | ||
|
|
||
| assert len(tpot_obj._pop) == n_seeds |
There was a problem hiding this comment.
I believe the unit tests need to be updated.
There was a problem hiding this comment.
There seems to be just one that broke, and only in 2.7. I'll install conda for 2.7 so I can see what went wrong.
What does this PR do?
seedsparameter toTPOTBaseseedsin their custom configWhere should the reviewer start?
Customizing TPOT's starting population in the "Using" section of the docs.
In this function in
TPOTBase:How should this PR be tested?
There are tests added to the
tpot_tests.pytest file.What are the relevant issues?
#59
#296