def optimise_proportional_training(songs,
goal,
conf,
datasaver=None,
iday=None):
"""
Optimise gestures randomly from the song models
with a stochastic local search
The number of trainings for this day is proportional to
the total number of gestures in the day songs
"""
measure = conf['measure_obj']
comp = conf['comp_obj']
rng = conf['rng_obj']
train_multiplier = conf.get('train_multiplier', 3)
if datasaver is None:
datasaver = QuietDataSaver()
if rng is None:
rng = np.random.RandomState()
train_per_day = 0
for song in songs:
train_per_day += len(song.gestures)
train_per_day *= train_multiplier
train_per_day = round(
train_per_day) # avoid possible float value with train_multiplier < 1
datasaver.add(label='day',
cond='define_number_of_trainings',
train_per_day=train_per_day)
improvement_cpt = np.zeros(train_per_day)
for itrain in range(train_per_day):
isong = rng.randint(len(songs))
song = songs[isong]
ig = rng.randint(len(song.gestures))
s = song.gen_sound()
c = measure(s)
pre_score = comp(goal, c)
logger.info(
'{}/{}: fit gesture {} of song {} (length {}, score {})'.format(
itrain + 1, train_per_day, ig, isong, len(s), pre_score))
res, hill_score = fit_gesture_whole_local_search(goal, song, ig, conf)
songs[isong].gestures[ig][1] = deepcopy(res)
# datasaver.add(iday=iday, itrain=itrain, isong=isong, ig=ig,
# pre_score=pre_score, new_score=hill_score) # too much verbose
logger.info('new score {}'.format(hill_score))
assert pre_score >= hill_score, "{} >= {} est faux".format(
pre_score, hill_score)
if hill_score < pre_score:
improvement_cpt[itrain] += 1
datasaver.add(label='day',
cond='after_day_learning',
improvement_cpt=improvement_cpt)
return songs
def optimise_gesture_whole_local_search(songs,
goal,
conf,
datasaver=None,
iday=None):
"""
Optimise gestures randomly from the song models
with a stochastic local search
"""
measure = conf['measure_obj']
comp = conf['comp_obj']
train_per_day = conf['train_per_day']
rng = conf['rng_obj']
if datasaver is None:
datasaver = QuietDataSaver()
if rng is None:
rng = np.random.RandomState()
improvement_cpt = np.zeros(train_per_day)
for itrain in range(train_per_day):
isong = rng.randint(len(songs))
song = songs[isong]
ig = rng.randint(len(song.gestures))
s = song.gen_sound()
c = measure(s)
pre_score = comp(goal, c)
logger.info(
'{}/{}: fit gesture {} of song {} (length {}, score {})'.format(
itrain + 1, train_per_day, ig, isong, len(s), pre_score))
res, hill_score = fit_gesture_whole_local_search(goal, song, ig, conf)
songs[isong].gestures[ig][1] = deepcopy(res)
# datasaver.add(iday=iday, itrain=itrain, isong=isong, ig=ig,
# pre_score=pre_score, new_score=hill_score) # too much verbose
logger.info('new score {}'.format(hill_score))
assert pre_score >= hill_score, "{} >= {} est faux".format(
pre_score, hill_score)
if hill_score < pre_score:
improvement_cpt[itrain] += 1
datasaver.add(label='day',
cond='after_day_learning',
improvement_cpt=improvement_cpt)
return songs
def mutate_best_models_elite(songs, goal, conf,
datasaver=None):
"""Elite selection and mutation of the best models.
Keep the best mutations after each replay, parents are present in the
selection.
"""
if datasaver is None:
datasaver = QuietDataSaver()
measure = conf['measure_obj']
comp = conf['comp_obj']
nb_replay = conf['replay']
rng = conf['rng_obj']
nb_conc_song = len(songs)
pscore = get_scores(goal, songs, measure, comp)
score = pscore
nb_conc_night = nb_conc_song * 2
# make night_songs an array to do list indexes.
night_songs = np.array(songs)
for dummy_i in range(nb_replay):
fitness = len(night_songs) - rank(score)
night_songs = np.random.choice(night_songs, size=nb_conc_night,
p=fitness/np.sum(fitness))
night_songs = np.array([song.mutate() for song in night_songs])
score = get_scores(goal, night_songs, measure, comp)
fitness = len(night_songs) - rank(score)
isongs = rng.choice(len(night_songs),
size=nb_conc_song, replace=False,
p=fitness/np.sum(fitness))
night_songs = night_songs[isongs]
score = score[isongs]
nsongs = night_songs.tolist()
datasaver.add(prev_songs=songs, prev_scores=pscore, new_songs=nsongs,
new_scores=score)
return nsongs
def mutate_best_models_dummy(songs, goal, measure, comp, nb_replay,
datasaver=None, rng=None):
"""Dummy selection and mutation of the best models."""
if datasaver is None:
datasaver = QuietDataSaver()
if rng is None:
rng = np.random.RandomState()
nb_conc_song = len(songs)
night_songs = np.array(songs)
pscore = get_scores(goal, songs, measure, comp)
nb_conc_night = nb_conc_song * 2
fitness = len(night_songs) - rank(pscore)
night_songs = np.random.choice(night_songs, size=nb_conc_night,
p=fitness/np.sum(fitness))
night_songs = np.array([song.mutate(nb_replay) for song in night_songs])
score = get_scores(goal, night_songs, measure, comp)
fitness = len(night_songs) - rank(score)
isongs = rng.choice(len(night_songs),
size=nb_conc_song, replace=False,
p=fitness/np.sum(fitness))
nsongs = night_songs[isongs].tolist()
datasaver.add(prev_songs=songs, prev_scores=pscore, new_songs=nsongs,
new_scores=score[isongs])
return nsongs