def dump_model(file,
language,
method,
seed,
library,
sampling_rate,
normalization,
shape,
augmentation,
params=None,
extra={}):
from time import time
from deep_audio import JSON
dump_info = {
'method': method,
'language': language,
'normalization': normalization,
'seed': seed,
'augmentation': augmentation,
'library': library,
'sample_rate': sampling_rate,
'shape': shape,
'timestamp': time(),
'params': params,
**extra
}
JSON.create_json_file(file, dump_info, cls=JSON.NumpyEncoder)
return
def dump_grid(file,
model,
language,
method,
seed,
library,
sizes,
score_train,
score_test,
sampling_rate,
normalization,
shape,
augmentation,
score_valid=None,
model_file=None,
extra={}):
from time import time
from deep_audio import JSON
dump_info = {
'method': method,
'language': language,
'normalization': normalization,
'seed': seed,
'augmentation': augmentation,
'library': library,
'sample_rate': sampling_rate,
'shape': shape,
'sizes': sizes,
'score_train': score_train,
'score_test': score_test,
'timestamp': time(),
'params': model.best_params_,
'cv_results': model.cv_results_,
**extra
}
if score_valid:
dump_info['score_valid'] = score_valid
if model:
dump_info['model_file'] = model_file
JSON.create_json_file(file, dump_info, cls=JSON.NumpyEncoder)
return
def object_to_json(filename, attrs, files):
from deep_audio import JSON
data = {
'mapping': [file.replace('.wav', '') for _, file in enumerate(files)],
'classes': [],
'labels': [],
'attrs': []
}
for i in attrs:
data['attrs'].extend(i['attrs'])
data['labels'].extend(i['labels'])
data['classes'].extend(i['classes'])
JSON.create_json_file(filename, data, cls=JSON.NumpyEncoder)
del data
# %%
# CRIA O MODELO
learning_rate = 0.0001
model = build_model(learning_rate=learning_rate)
# %%
# SALVA A ESTRUTURA DO MODELO
timestamp = int(time.time())
Directory.create_directory(
f'{language}/models/{model_algo}/{library}/{filename_ps}{timestamp}')
JSON.create_json_file(
f'{language}/models/{model_algo}/{library}/{filename_ps}{timestamp}/model_structure.json',
model.to_json())
model_save_filename = f'{language}/models/{model_algo}/{library}/{filename_ps}{timestamp}/model_weight.h5'
# DECIDE QUANDO PARAR
earlystopping_cb = EarlyStopping(patience=300, restore_best_weights=True)
# SALVA OS PESOS
mdlcheckpoint_cb = ModelCheckpoint(model_save_filename,
monitor="val_accuracy",
save_best_only=True)
# %%
# TREINA O MODELO
history = model.fit(X_train,
score_test = model.score(X_test, y_test)
score_train = model.score(X_train, y_train)
y_hat = model.predict(X_test)
filename_model = Directory.model_filename(method,
language,
library,
normalization,
score_test,
augmentation=augment,
json=False) + 'model.json'
JSON.create_json_file(file=filename_model,
data=globals()['build_' + method]().to_json())
# SALVA ACURÁCIAS E PARAMETROS
Model.dump_grid(Directory.model_filename(method,
language,
library,
normalization,
score_test,
augmentation=augment),
model=model,
language=language,
method=method,
normalization=normalization,
sampling_rate=sampling_rate,
augmentation=augment,
shape=X_train.shape,
model.fit(X_train, y_train)
best_params = model.best_params_
score_test = model.score(X_test, y_test)
score_train = model.score(X_train, y_train)
y_hat = model.predict(X_test)
filename_model = Directory.model_filename(
'cnn', language, library, normalization, score_test, json=False)+'model.json'
JSON.create_json_file(
file=filename_model,
data=build_model().to_json()
)
# SALVA ACURÁCIAS E PARAMETROS
Model.dump_grid(
Directory.model_filename(
'cnn', language, library, normalization, score_test),
model=model,
language=language,
method='CNN',
normalization=normalization,
sampling_rate=sampling_rate,
seed=random_state,
library=library,
sizes=[len(X_train), len(X_valid), len(X_test)],
score_train=score_train,
augmentation=augment,
json=False,
models=True)
model_save_filename = filename_holder + 'weight.h5'
# SALVA OS PESOS
mdlcheckpoint_cb = keras.callbacks.ModelCheckpoint(model_save_filename,
monitor="accuracy",
save_best_only=True)
# history = model.fit(X_train, y_train, epochs=epochs,
# batch_size=batch_size, callbacks=[mdlcheckpoint_cb])
# SALVA ESTRUTURA DO MODELO
JSON.create_json_file(file=filename_holder + 'model.json',
data=globals()['build_' + method]().to_json())
dump(scaler, open(filename_holder + 'scaler.pkl', 'wb'))
# SALVA OS PARAMETROS
Model.dump_model(filename_holder + 'info.json',
params=param_grid,
language=language,
method=method,
normalization=normalization,
sampling_rate=sampling_rate,
augmentation=augment,
shape=X_train.shape,
seed=random_state,
library=library,
extra={