def loadTrainParameters(self):
params_file = os.path.join(self.RESULTS_PATH, self.analysis_name,
"trnparams.jbl")
if (os.path.exists(params_file)):
self.trn_params = TrainParameters.SAENoveltyDetectionTrnParams()
self.trn_params.load(params_file)
# Choose how many fold to be used in Cross Validation
self.CVO = TrainParameters.NoveltyDetectionFolds(
folder=self.RESULTS_PATH,
n_folds=self.n_folds,
trgt=self.all_trgt,
dev=self.development_flag,
verbose=True)
else:
self.setTrainParameters()
# Load train parameters
analysis_str = 'StackedAutoEncoder'
model_prefix_str = 'RawData'
trn_params_folder = '%s/%s/%s_trnparams.jbl' % (results_path, analysis_str,
analysis_name)
# if os.path.exists(trn_params_folder):
# os.remove(trn_params_folder)
if not os.path.exists(trn_params_folder):
trn_params = trnparams.NeuralClassificationTrnParams(
n_inits=1,
hidden_activation='tanh', # others tanh, relu, sigmoid, linear
output_activation='linear',
n_epochs=300, #500
patience=30, #30
batch_size=256, #128
verbose=False,
optmizerAlgorithm='Adam',
metrics=['accuracy'],
loss='mean_squared_error')
trn_params.save(trn_params_folder)
else:
trn_params = trnparams.NeuralClassificationTrnParams()
trn_params.load(trn_params_folder)
# Choose how many fold to be used in Cross Validation
n_folds = 10
CVO = trnparams.NoveltyDetectionFolds(folder=results_path,
n_folds=n_folds,
trgt=all_trgt,
def setTrainParameters(self,
n_inits=1,
hidden_activation='tanh',
output_activation='linear',
classifier_output_activation='softmax',
n_epochs=300,
n_folds=10,
patience=30,
batch_size=256,
verbose=False,
optmizerAlgorithm='Adam',
metrics=['accuracy'],
loss='mean_squared_error',
norm='mapstd',
train_verbose=False,
learning_rate=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
learning_decay=1e-6,
momentum=0.3,
nesterov=True):
self.trn_params = TrainParameters.SAENoveltyDetectionTrnParams(
n_inits=n_inits,
folds=n_folds,
hidden_activation=
hidden_activation, # others tanh, relu, sigmoid, linear
output_activation=output_activation,
classifier_output_activation=classifier_output_activation,
n_epochs=n_epochs,
patience=patience,
batch_size=batch_size,
verbose=verbose,
optmizerAlgorithm=optmizerAlgorithm,
metrics=metrics, #mean_squared_error
loss=loss, #kullback_leibler_divergence
norm=norm,
train_verbose=train_verbose,
learning_rate=learning_rate,
beta_1=beta_1,
beta_2=beta_2,
epsilon=epsilon,
learning_decay=learning_decay,
momentum=momentum,
nesterov=nesterov)
self.modelPath = self.trn_params.getModelPath()
self.baseResultsPath = self.getBaseResultsPath()
self.trn_params_file = os.path.join(self.baseResultsPath,
"trnparams.jbl")
if not os.path.exists(self.trn_params_file):
self.trn_params.save(self.trn_params_file)
self.trn_params.save(
os.path.join(self.RESULTS_PATH, self.analysis_name,
"trnparams.jbl"))
# Choose how many fold to be used in Cross Validation
self.CVO = TrainParameters.NoveltyDetectionFolds(
folder=self.RESULTS_PATH,
n_folds=n_folds,
trgt=self.all_trgt,
dev=self.development_flag,
verbose=True)
# turn targets in sparse mode
from keras.utils import np_utils
trgt_sparse = np_utils.to_categorical(all_trgt.astype(int))
# Load train parameters
analysis_str = 'StackedAutoEncoder'
model_prefix_str = 'RawData'
trn_params_folder = '%s/%s/%s_trnparams.jbl' % (results_path, analysis_str,
analysis_name)
os.remove(trn_params_folder)
if not os.path.exists(trn_params_folder):
trn_params = trnparams.NeuralClassificationTrnParams(
n_inits=1,
hidden_activation='tanh', # others tanh, relu, sigmoid, linear
output_activation='linear',
n_epochs=50, #500
patience=10, #30
batch_size=4, #256
verbose=False)
trn_params.save(trn_params_folder)
else:
trn_params = trnparams.NeuralClassificationTrnParams()
trn_params.load(trn_params_folder)
# Choose how many fold to be used in Cross Validation
n_folds = 2
CVO = trnparams.ClassificationFolds(folder=results_path,
n_folds=2,
trgt=all_trgt,
dev=development_flag,
verbose=True)