def predict_many(data_source,
saved_model_tag=None,
ensemble=None,
client=None,
serialize=None,
to_raster=True,
elm_predict_path=None):
'''See elm.pipeline.Pipeline.predict_many method
'''
env = parse_env_vars()
elm_predict_path = elm_predict_path or env.get('ELM_PREDICT_PATH')
if serialize and elm_predict_path and not os.path.exists(elm_predict_path):
os.mkdir(elm_predict_path)
pipe_example = ensemble[0][1]
ds = data_source.copy()
X = ds.pop('X', None)
y = ds.pop('y', None)
args_list = ds.pop('args_list', None)
sampler = ds.pop('sampler', None)
dsk = make_samples_dask(X, y, None, pipe_example, args_list, sampler, ds)
sample_keys = tuple(dsk)
args_list = tuple(itertools.product(sample_keys, ensemble))
keys = []
last_file_name = None
for idx, (sample_key, (estimator_tag, estimator)) in enumerate(args_list):
name = _next_name('predict_many')
predict_tag = '{}-{}'.format(estimator_tag, sample_key)
if saved_model_tag:
predict_tag += '-' + saved_model_tag
dsk[name] = (
_predict_one_sample_one_arg,
estimator,
serialize,
to_raster,
predict_tag,
elm_predict_path,
sample_key,
)
keys.append(name)
logger.info('Predict {} estimator(s) and {} sample(s) '
'({} combination[s])'.format(len(ensemble), len(sample_keys),
len(args_list)))
preds = []
if client is None:
new = dask.get(dsk, keys)
else:
new = client.get(dsk, keys)
return tuple(itertools.chain.from_iterable(new))
def evolve_train(pipe,
evo_params,
X=None,
y=None,
sample_weight=None,
sampler=None,
args_list=None,
client=None,
init_ensemble_size=1,
saved_ensemble_size=None,
ensemble_init_func=None,
scoring_kwargs=None,
method='fit',
partial_fit_batches=1,
classes=None,
method_kwargs=None,
**data_source):
'''evolve_train runs an evolutionary algorithm to
find the most fit elm.pipeline.Pipeline instances
Parameters:
pipe: elm.pipeline.Pipeline instance
evo_params: the EvoParams instance, typically from
from elm.model_selection import ea_setup
evo_params = ea_setup(param_grid=param_grid,
param_grid_name='param_grid_example',
score_weights=[-1]) # minimization
See also the help from (elm.pipeline.ensemble) where
most arguments are interpretted similary.
''' + ensemble.__doc__
models_share_sample = True
method_kwargs = method_kwargs or {}
scoring_kwargs = scoring_kwargs or {}
get_func = _find_get_func_for_client(client)
control = evo_params.deap_params['control']
required_args, _, _ = get_args_kwargs_defaults(ea_general)
evo_args = [
evo_params,
]
data_source = dict(X=X,
y=y,
sample_weight=sample_weight,
sampler=sampler,
args_list=args_list,
**data_source)
fit_one_generation = partial(_on_each_generation, pipe, data_source,
evo_params.deap_params, get_func,
partial_fit_batches, method, method_kwargs)
dsk = make_samples_dask(X, y, sample_weight, pipe, args_list, sampler,
data_source)
sample_keys = list(dsk)
if models_share_sample:
np.random.shuffle(sample_keys)
gen_to_sample_key = lambda gen: [sample_keys[gen]]
else:
gen_to_sample_key = lambda gen: sample_keys
sample_keys = tuple(sample_keys)
try:
param_history = []
for a in required_args[1:]:
if a not in control:
raise ValueError('Expected {} in {} (control kwargs '
'to evolutionary '
'algorithm)'.format(a, control))
evo_args.append(control[a])
ea_gen = ea_general(*evo_args)
pop, _, _ = next(ea_gen)
sample_keys_passed = gen_to_sample_key(0)
def log_once(len_models, sample_keys_passed, gen):
total_calls = len_models * len(
sample_keys_passed) * partial_fit_batches
msg = (len_models, len(sample_keys_passed), partial_fit_batches,
method, gen, total_calls)
fmt = 'Evolve generation {4}: {0} models x {1} samples x {2} {3} calls = {5} calls in total'
logger.info(fmt.format(*msg))
log_once(len(pop), sample_keys_passed, 0)
pop_names = [ind.name for ind in pop]
models, fitnesses = fit_one_generation(dsk, 0, sample_keys_passed, pop)
assign_check_fitness(pop, fitnesses, param_history,
evo_params.deap_params['choices'],
evo_params.score_weights)
invalid_ind = True
fitted_models = {n: m for n, (_, m) in zip(pop_names, models)}
ngen = evo_params.deap_params['control'].get('ngen') or None
if not ngen and not evo_params.early_stop:
raise ValueError('param_grids: pg_name: control: has neither '
'ngen or early_stop keys')
elif not ngen:
ngen = 1000000
for gen in range(ngen):
# on last generation invalid_ind becomes None
# and breaks this loop
if models_share_sample:
sample_keys_passed = (gen_to_sample_key(gen %
len(sample_keys)), )
else:
sample_keys_passed = sample_keys
if gen > 0:
log_once(len(invalid_ind), sample_keys_passed, gen)
names = [ind.name for ind in invalid_ind]
models, fitnesses = fit_one_generation(dsk, gen,
sample_keys_passed,
invalid_ind)
fitted_models.update(
{n: m
for n, (_, m) in zip(names, models)})
(pop, invalid_ind, param_history) = ea_gen.send(fitnesses)
pop_names = [ind.name for ind in pop]
fitted_models = {
k: v
for k, v in fitted_models.items() if k in pop_names
}
if not invalid_ind:
break # If there are no new solutions to try, break
pop = evo_params.toolbox.select(pop, saved_ensemble_size)
pop_names = [ind.name for ind in pop]
models = [(k, v) for k, v in fitted_models.items() if k in pop_names]
finally:
columns = list(evo_params.deap_params['param_order'])
columns += [
'objective_{}_{}'.format(idx, 'min' if sw == -1 else 'max')
for idx, sw in enumerate(evo_params.score_weights)
]
if param_history:
assert len(columns) == len(param_history[0])
param_history = pd.DataFrame(np.array(param_history),
columns=columns)
param_history.to_csv(evo_params.history_file,
index_label='parameter_set')
return models
def ensemble(pipe,
ngen,
X=None,
y=None,
sample_weight=None,
sampler=None,
args_list=None,
client=None,
init_ensemble_size=1,
saved_ensemble_size=None,
ensemble_init_func=None,
models_share_sample=True,
model_selection=None,
model_selection_kwargs=None,
scoring_kwargs=None,
method='fit',
partial_fit_batches=1,
classes=None,
method_kwargs=None,
**data_source):
'''Fit or partial_fit an ensemble of models to a series of samples
Call this function from an elm.pipeline.Pipeline instance's methods:
"fit_ensemble"
"fit_transform_ensemble"
"transform_ensemble"
Parameters:
pipe: instance of elm.pipeline.Pipeline
ngen: number of ensemble generations
X: earthio.ElmStore, if not using "sampler" and "args_list"
y: numpy array if not using "sampler" and "args_list", or None
if not needed by Pipeline
sample_weight: numpy array if not using "sampler" and "args_list", or None
if not needed by Pipeline
sampler: Callable - required if not giving X. Called at least once
on each element of args_list where each element is unpacked
with *one_element_of_args_list
args_list: List of args - required if not giving X. See sampler above
client: dask-distributed or ThreadPool client
init_ensemble_size: number of ensemble members, ignored if giving
ensemble_init_func
saved_ensemble_size: how many members to keep at final generation
ensemble_init_func: Callable to return list of elm.pipeline.Pipeline
instances that initialize ensemble
models_share_sample: If True, ensure that in each generation, every
member is fit to the same sample. If False, fit every model
to every sample
model_selection: Callable after each generation to take a list of
(tag, Pipeline) tuples and return a list of new such tuples, or
None or repeatedly train each model on each generation without
replacement of model parameters
model_selection_kwargs: kwargs passed to model_selection
scoring_kwargs: kwargs that are passed to score_one_model.
See also elm.model_selection.scoring
method: This is the method of Pipeline that called this ensemble
function, typically "fit"
classes: Unique sequence of class integers passed to supervised
classifiers that need the known y classes.
method_kwargs: any other arguments to pass to method
**data_source: keywords passed to "sampler" if given
Returns:
new_models: list of (tag, Pipeline instance) tuples on which
"predict_many" can be called
'''
get_func = _find_get_func_for_client(client)
fit_score_kwargs = method_kwargs or {}
if not 'classes' in fit_score_kwargs and classes is not None:
fit_score_kwargs['classes'] = classes
model_selection_kwargs = model_selection_kwargs or {}
ensemble_size = init_ensemble_size or 1
partial_fit_batches = partial_fit_batches or 1
if partial_fit_batches > 1:
method = 'partial_fit'
if not ensemble_init_func:
models = tuple(copy.deepcopy(pipe) for _ in range(ensemble_size))
else:
ensemble_init_func = import_callable(ensemble_init_func)
models = ensemble_init_func(pipe, ensemble_size=ensemble_size)
logger.info("Init ensemble: {} members".format(len(models)))
if model_selection:
model_selection = import_callable(model_selection)
final_names = []
dsk = make_samples_dask(X, y, sample_weight, pipe, args_list, sampler,
data_source)
models = tuple(
zip(('tag_{}'.format(idx) for idx in range(len(models))), models))
sample_keys = list(dsk)
if models_share_sample:
random.shuffle(sample_keys)
gen_to_sample_key = {
gen: s
for gen, s in enumerate(sample_keys[:ngen])
}
sample_keys = tuple(sample_keys)
for gen in range(ngen):
if models_share_sample:
sample_keys_passed = (gen_to_sample_key[gen % len(sample_keys)], )
else:
sample_keys_passed = sample_keys
logger.info('Ensemble generation {} of {} - ({} estimators) '.format(
gen + 1, ngen, len(models)))
msg = (len(models), len(sample_keys_passed), partial_fit_batches,
method,
len(models) * len(sample_keys_passed) * partial_fit_batches,
gen + 1, ngen)
logger.info(
'Ensemble Generation {5} of {6}: ({0} members x {1} samples x {2} calls) = {4} {3} calls this gen'
.format(*msg))
dsk, model_keys, new_models_name = _one_generation_dask_graph(
dsk, models, fit_score_kwargs, sample_keys_passed,
partial_fit_batches, gen, method)
if get_func is None:
new_models = tuple(dask.get(dsk, new_models_name))
else:
new_models = tuple(get_func(dsk, new_models_name))
models = tuple(zip(model_keys, new_models))
logger.info('Trained {} estimators'.format(len(models)))
if model_selection:
models = _run_model_selection(models, model_selection,
model_selection_kwargs or {}, ngen,
gen, scoring_kwargs)
else:
pass # Just training all ensemble members
# without replacing / re-ininializing / editing
# the model params
if saved_ensemble_size:
final_models = models[:saved_ensemble_size]
else:
final_models = models
return final_models
