aimclub / FEDOT

Automated modeling and machine learning framework FEDOT
https://fedot.readthedocs.io
BSD 3-Clause "New" or "Revised" License
627 stars 86 forks source link

Mutation error. Crash in random package. #342

Closed MAGLeb closed 2 years ago

MAGLeb commented 3 years ago

If build composer with initial chain contains only ONE node.

**initial_chain** = PrimaryNode('xgboost')
builder = GPComposerBuilder(task=task).
with_requirements(composer_requirements).
with_metrics(metric_function).
with_optimiser_parameters(optimiser_parameters).
with_initial_chain(**initial_chain**)
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-36-7c15fbfd9fe1> in <module>
     10 
     11 # the optimal chain generation by composition - the most time-consuming task
---> 12 chain_evo_composed = composer.compose_chain(data=input_data,
     13                                             is_visualise=True)

~/.local/lib/python3.8/site-packages/fedot/core/composer/gp_composer/gp_composer.py in compose_chain(self, data, is_visualise, is_tune, on_next_iteration_callback)
    138             self.cache = OperationsCache(self.cache_path, clear_exiting=not self.use_existing_cache)
    139 
--> 140         best_chain = self.optimiser.optimise(metric_function_for_nodes,
    141                                              on_next_iteration_callback=on_next_iteration_callback)
    142 

~/.local/lib/python3.8/site-packages/fedot/core/composer/optimisers/gp_comp/param_free_gp_optimiser.py in optimise(self, objective_function, offspring_rate, on_next_iteration_callback)
    118 
    119                     for parent_num in range(0, len(selected_individuals), 2):
--> 120                         new_population += self.reproduce(selected_individuals[parent_num],
    121                                                          selected_individuals[parent_num + 1])
    122 

~/.local/lib/python3.8/site-packages/fedot/core/composer/optimisers/gp_comp/gp_optimiser.py in reproduce(self, selected_individual_first, selected_individual_second)
    299             new_inds = [selected_individual_first]
    300 
--> 301         new_inds = tuple([mutation(types=self.parameters.mutation_types,
    302                                    chain_generation_params=self.chain_generation_params,
    303                                    ind=new_ind, requirements=self.requirements,

~/.local/lib/python3.8/site-packages/fedot/core/composer/optimisers/gp_comp/gp_optimiser.py in <listcomp>(.0)
    299             new_inds = [selected_individual_first]
    300 
--> 301         new_inds = tuple([mutation(types=self.parameters.mutation_types,
    302                                    chain_generation_params=self.chain_generation_params,
    303                                    ind=new_ind, requirements=self.requirements,

~/.local/lib/python3.8/site-packages/fedot/core/composer/optimisers/gp_comp/operators/mutation.py in mutation(types, chain_generation_params, ind, requirements, log, max_depth)
     63         if mutation_type in mutation_by_type:
     64             for _ in range(MAX_NUM_OF_ATTEMPTS):
---> 65                 new_chain = mutation_by_type[mutation_type](chain=deepcopy(ind.chain), requirements=requirements,
     66                                                             chain_generation_params=chain_generation_params,
     67                                                             max_depth=max_depth)

~/.local/lib/python3.8/site-packages/fedot/core/composer/optimisers/gp_comp/operators/mutation.py in reduce_mutation(chain, requirements, chain_generation_params, max_depth)
    143 
    144     nodes = [node for node in chain.nodes if node is not chain.root_node]
--> 145     node_to_del = choice(nodes)
    146     children = chain.operator.node_children(node_to_del)
    147     is_possible_to_delete = all([len(child.nodes_from) - 1 >= requirements.min_arity for child in children])

/usr/lib/python3.8/random.py in choice(self, seq)
    288             i = self._randbelow(len(seq))
    289         except ValueError:
--> 290             raise IndexError('Cannot choose from an empty sequence') from None
    291         return seq[i]
    292 

IndexError: Cannot choose from an empty sequence
nicl-nno commented 3 years ago

Да, есть такое. В качестве воркэраунда можно задавать scaling -> xgboost, все равно на практике в одной модели смысла мало.

Dreamlone commented 3 years ago

В качестве воркэраунда можно задавать scaling -> xgboost, все равно на практике в одной модели смысла мало.

С одной стороны такое себе "экспертное начальное приближение" из одной модели, а с другой: может такие цепочки искуственно расширять при попадании в композер: например "xgboost" превращать в "scaling -> xgboost" при вызове метода compose_chain, или уже на этапе инициализации билдера? Или ты это и имел в виду?

nicl-nno commented 3 years ago

Потом просто доработает мутацию.

aPovidlo commented 2 years ago

@nicl-nno не смог воспроизвести баг в новой версии FEDOT. Выполнял следующий код:

initial_pipeline = PrimaryNode('xgboost')
# initial_pipeline = Pipeline(PrimaryNode('xgboost'))

builder = GPComposerBuilder(task=task).with_requirements(composer_requirements)\
        .with_metrics(metric_function)\
        .with_optimiser_parameters(optimiser_parameters)\
        .with_initial_pipeline(initial_pipeline)
aPovidlo commented 2 years ago

Также @Dreamlone было предложено добавить в with_initial_pipeline() исключение, когда в функцию попадают объекты с типом не Pipeline, и выводить raise Exception(ValueError('...')).

nicl-nno commented 2 years ago

не смог воспроизвести баг в новой версии FEDOT Если кроссовер тоже не падает, то всё норм.

когда в функцию попадают объекты с типом не Pipeline

Да, можно. Только достаточно просто raise ValueError, он и так унаследован от Exception.