def load_best_lstm_mult_label_multi_class(exp, name_model, we_file_name): exp.pp_data.vocabulary_size = 5000 exp.pp_data.embedding_size = 300 exp.pp_data.max_posts = 1750 exp.pp_data.max_terms_by_post = 300 exp.pp_data.binary_classifier = True exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT exp.pp_data.remove_stopwords = False exp.pp_data.delete_low_tfid = False exp.pp_data.min_df = 0 exp.pp_data.min_tf = 0 exp.pp_data.random_posts = False exp.pp_data.random_users = False exp.pp_data.tokenizing_type = 'WE' exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL exp.use_custom_metrics = False exp.use_valid_set_for_train = True exp.valid_split_from_train_set = 0.0 exp.imbalanced_classes = False lstm = ModelClass(1) lstm.loss_function = 'binary_crossentropy' lstm.optmizer_function = 'adam' lstm.epochs = 15 lstm.batch_size = 32 lstm.patience_train = 10 lstm.use_embedding_pre_train = exp.pp_data.use_embedding lstm.embed_trainable = (lstm.use_embedding_pre_train == (dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC)) neuronios_by_layer = [16] epochs = [32] batch_sizes = [40] dropouts = [0.2] exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL np.random.seed(dn.SEED) time_ini_rep = datetime.datetime.now() x_test, y_test = exp.pp_data.load_dataset_generic( "/home/vanessa/PycharmProjects/RecurrentNetworks/dataset/anx_dep_multilabel/SMHD_multi_label_test_test_528.df", ['control', 'anxiety', 'depression']) exp.set_period_time_end(time_ini_rep, 'Load data') for neuronios in neuronios_by_layer: for batch_size in batch_sizes: for epoch in epochs: for dropout in dropouts: lstm.epochs = epoch lstm.batch_size = batch_size lstm.patience_train = epoch / 2 exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + \ str(batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_' + \ we_file_name lstm.model = exp.load_model(dn.PATH_PROJECT + exp.experiment_name + '.h5') exp.save_geral_configs() exp.save_summary_model(lstm.model) exp.predict_samples(lstm, x_test, y_test)
exp.pp_data.binary_classifier = True exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT exp.pp_data.remove_stopwords = False exp.pp_data.delete_low_tfid = False exp.pp_data.min_df = 0 exp.pp_data.min_tf = 0 exp.pp_data.random_posts = False exp.pp_data.random_users = False exp.pp_data.tokenizing_type = 'WE' exp.pp_data.use_embedding = dn.UseEmbedding.RAND exp.pp_data.embedding_type = dn.EmbeddingType.NONE lstm = ModelClass(1) lstm.loss_function = 'binary_crossentropy' lstm.optmizer_function = 'adam' lstm.epochs = 10 lstm.batch_size = 32 lstm.patience_train = 4 lstm.use_embedding_pre_train = exp.pp_data.use_embedding lstm.embed_trainable = True lstm.model = Sequential() lstm.model.add(Embedding(exp.pp_data.vocabulary_size, exp.pp_data.embedding_size, trainable=lstm.embed_trainable)) lstm.model.add(LSTM(64, activation='tanh', dropout=0.2, recurrent_dropout=0.2, return_sequences=True)) lstm.model.add(LSTM(32, activation='tanh', dropout=0.2, recurrent_dropout=0.2)) lstm.model.add(Dense(1, activation='sigmoid')) time_ini_exp = datetime.datetime.now() # exp.k_fold_cross_validation(lstm) exp.test_hypeparams(lstm) exp.set_period_time_end(time_ini_exp, 'Total experiment')
def generate_model(exp, name_model, kernel_name, set_params, function):
exp.pp_data.vocabulary_size = 5000
exp.pp_data.embedding_size = 300
exp.pp_data.max_posts = 1750
exp.pp_data.max_terms_by_post = 300
exp.pp_data.binary_classifier = True
exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT
exp.pp_data.remove_stopwords = False
exp.pp_data.delete_low_tfid = False
exp.pp_data.min_df = 0
exp.pp_data.min_tf = 0
exp.pp_data.random_posts = False
exp.pp_data.random_users = False
exp.pp_data.tokenizing_type = 'WE'
exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL
exp.use_custom_metrics = False
exp.use_valid_set_for_train = True
exp.valid_split_from_train_set = 0.0
exp.imbalanced_classes = False
cnn_lstm = ModelClass(1)
cnn_lstm.loss_function = 'binary_crossentropy'
cnn_lstm.optmizer_function = 'adadelta'
cnn_lstm.epochs = 15
cnn_lstm.batch_size = 32
cnn_lstm.patience_train = 10
filters_by_layer = set_params['filters_by_layer']
neuronios_by_lstm_layer = set_params['neuronios_by_lstm_layer']
dropouts = set_params['dropouts']
dropouts_lstm = set_params['dropouts_lstm']
kernels_size = set_params['kernels_size']
epochs = set_params['epochs']
batch_sizes = set_params['batch_sizes']
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
for embedding_type in set_params['embedding_types']:
for embedding_custom_file in set_params['embedding_custom_files']:
for use_embedding in set_params['use_embeddings']:
exp.pp_data.embedding_type = embedding_type
exp.pp_data.word_embedding_custom_file = embedding_custom_file
exp.pp_data.use_embedding = use_embedding
exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL
exp.set_period_time_end(time_ini_rep, 'Load data')
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data(
)
cnn_lstm.use_embedding_pre_train = exp.pp_data.use_embedding
cnn_lstm.embed_trainable = (
cnn_lstm.use_embedding_pre_train == (
dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
emb_name = function
if embedding_custom_file != '':
emb_name = exp.pp_data.word_embedding_custom_file.split(
'.')[0]
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + \
str(exp.pp_data.use_embedding.value) + '_EF_' + emb_name + kernel_name
for filter in filters_by_layer:
for kernel_size in kernels_size:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for dropout_lstm in dropouts_lstm:
for neuronios in neuronios_by_lstm_layer:
cnn_lstm.epochs = epoch
cnn_lstm.batch_size = batch_size
cnn_lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_cnn_lstm' + '_F' + str(filter) + '_K' +\
str(kernel_size) + '_P' + 'None' + '_B' + str(batch_size) +\
'_E' + str(epoch) + '_D' + str(dropout) + '_HLN' + str(filter) +\
'_LSTM_N' + str(neuronios) + '_D' + str(dropout_lstm) + '_' + we_file_name
cnn_lstm.model = Sequential()
cnn_lstm.model.add(
Embedding(
exp.pp_data.
vocabulary_size,
exp.pp_data.embedding_size,
trainable=cnn_lstm.
embed_trainable,
name='emb_' + name_model))
cnn_lstm.model.add(
Dropout(dropout,
name='dropout_1_' +
name_model))
cnn_lstm.model.add(
Conv1D(
filters=filter,
kernel_size=kernel_size,
kernel_initializer=
'glorot_uniform',
# kernel_regularizer=regularizers.l2(0.03),
padding='valid',
activation='relu',
name='conv_1_' +
name_model))
cnn_lstm.model.add(
MaxPooling1D(
name='max_pool_1_' +
name_model))
cnn_lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout_lstm,
recurrent_dropout=
dropout_lstm,
return_sequences=True,
name='lstm_1_' +
name_model))
cnn_lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout_lstm,
recurrent_dropout=
dropout_lstm,
return_sequences=True,
name='lstm_2_' +
name_model))
cnn_lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout_lstm,
recurrent_dropout=
dropout_lstm,
name='lstm_3_' +
name_model))
cnn_lstm.model.add(
Dense(3,
activation='sigmoid',
name='dense_1_' +
name_model))
time_ini_exp = datetime.datetime.now(
)
exp.generate_model_hypeparams(
cnn_lstm, x_train, y_train,
x_valid, y_valid,
embedding_matrix)
exp.set_period_time_end(
time_ini_exp,
'Total experiment')
del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix
# Test
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
for embedding_type in set_params['embedding_types']:
for embedding_custom_file in set_params['embedding_custom_files']:
for use_embedding in set_params['use_embeddings']:
exp.pp_data.embedding_type = embedding_type
exp.pp_data.word_embedding_custom_file = embedding_custom_file
exp.pp_data.use_embedding = use_embedding
exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL
exp.set_period_time_end(time_ini_rep, 'Load data')
x_test, y_test = exp.pp_data.load_data()
cnn_lstm.use_embedding_pre_train = exp.pp_data.use_embedding
cnn_lstm.embed_trainable = (
cnn_lstm.use_embedding_pre_train == (
dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
emb_name = function
if embedding_custom_file != '':
emb_name = exp.pp_data.word_embedding_custom_file.split(
'.')[0]
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + \
str(exp.pp_data.use_embedding.value) + '_EF_' + emb_name + kernel_name
for filter in filters_by_layer:
for kernel_size in kernels_size:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for dropout_lstm in dropouts_lstm:
for neuronios in neuronios_by_lstm_layer:
cnn_lstm.epochs = epoch
cnn_lstm.batch_size = batch_size
cnn_lstm.patience_train = epoch
exp.experiment_name = name_model + '_cnn_lstm' + '_F' + str(filter) + '_K' +\
str(kernel_size) + '_P' + 'None' + '_B' + str(batch_size) +\
'_E' + str(epoch) + '_D' + str(dropout) + '_HLN' + str(filter) +\
'_LSTM_N' + str(neuronios) + '_D' + str(dropout_lstm) + '_' + we_file_name
cnn_lstm.model = exp.load_model(
dn.PATH_PROJECT +
exp.experiment_name + '.h5')
exp.save_geral_configs(
'Experiment Specific Configuration: '
+ exp.experiment_name)
exp.save_summary_model(
cnn_lstm.model)
exp.predict_samples(
cnn_lstm, x_test, y_test)
del x_test, y_test
del cnn_lstm, exp
def generate_model(exp, name_model, kernel_function, set_params, function):
exp.pp_data.vocabulary_size = 5000
exp.pp_data.embedding_size = 300
exp.pp_data.max_posts = 1750
exp.pp_data.max_terms_by_post = 300
exp.pp_data.binary_classifier = True
exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT
exp.pp_data.remove_stopwords = False
exp.pp_data.delete_low_tfid = False
exp.pp_data.min_df = 0
exp.pp_data.min_tf = 0
exp.pp_data.random_posts = False
exp.pp_data.random_users = False
exp.pp_data.tokenizing_type = 'WE'
exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL
exp.use_custom_metrics = False
exp.use_valid_set_for_train = True
exp.valid_split_from_train_set = 0.0
exp.imbalanced_classes = False
lstm = ModelClass(1)
lstm.loss_function = 'binary_crossentropy'
optimizer_functions = set_params['optimizer_function']
neuronios_by_layer = set_params['neuronios_by_layer']
epochs = set_params['epochs']
batch_sizes = set_params['batch_sizes']
hidden_layers = set_params['hidden_layers']
dropouts = [0.2]
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
for embedding_type in set_params['embedding_types']:
for embedding_custom_file in set_params['embedding_custom_files']:
for use_embedding in set_params['use_embeddings']:
exp.pp_data.embedding_type = embedding_type
exp.pp_data.word_embedding_custom_file = embedding_custom_file
exp.pp_data.use_embedding = use_embedding
exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL
exp.set_period_time_end(time_ini_rep, 'Load data')
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data(
)
lstm.use_embedding_pre_train = exp.pp_data.use_embedding
lstm.embed_trainable = (lstm.use_embedding_pre_train == (
dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
emb_name = function
if embedding_custom_file != '':
emb_name = exp.pp_data.word_embedding_custom_file.split(
'.')[0]
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + \
str(exp.pp_data.use_embedding.value) + '_EF_' + emb_name + kernel_function
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for optmizer_function in optimizer_functions:
for hidden_layer in hidden_layers:
lstm.optmizer_function = dn.OPTIMIZER_FUNCTIONS[
optmizer_function]
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + \
str(batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_OF'+ \
lstm.optmizer_function + '_HL' + str(hidden_layer) + '_' + we_file_name
lstm.model = Sequential()
lstm.model.add(
Embedding(
exp.pp_data.vocabulary_size,
exp.pp_data.embedding_size,
trainable=lstm.embed_trainable,
name='emb_' + name_model))
for id_hl in range(hidden_layer):
lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout, recurrent_dropout=dropout,
return_sequences=True, name='dense_' + str(id_hl) + '_' + \
name_model))
lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout,
recurrent_dropout=dropout,
name='dense_' +
str(id_hl + 1) + '_' +
name_model))
lstm.model.add(
Dense(3,
activation='sigmoid',
name='dense_' +
str(id_hl + 2) + '_' +
name_model))
time_ini_exp = datetime.datetime.now()
exp.generate_model_hypeparams(
lstm, x_train, y_train, x_valid,
y_valid, embedding_matrix)
exp.set_period_time_end(
time_ini_exp, 'Total experiment')
del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix
# Test
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
for embedding_type in set_params['embedding_types']:
for embedding_custom_file in set_params['embedding_custom_files']:
for use_embedding in set_params['use_embeddings']:
exp.pp_data.embedding_type = embedding_type
exp.pp_data.word_embedding_custom_file = embedding_custom_file
exp.pp_data.use_embedding = use_embedding
exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL
exp.set_period_time_end(time_ini_rep, 'Load data')
x_test, y_test = exp.pp_data.load_data()
lstm.use_embedding_pre_train = exp.pp_data.use_embedding
lstm.embed_trainable = (lstm.use_embedding_pre_train == (
dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
emb_name = function
if embedding_custom_file != '':
emb_name = exp.pp_data.word_embedding_custom_file.split(
'.')[0]
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + \
str(exp.pp_data.use_embedding.value) + '_EF_' + emb_name + kernel_function
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for optmizer_function in optimizer_functions:
for hidden_layer in hidden_layers:
lstm.optmizer_function = dn.OPTIMIZER_FUNCTIONS[
optmizer_function]
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + \
str(batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_OF' + \
lstm.optmizer_function + '_HL' + str(hidden_layer) + '_' + we_file_name
lstm.model = exp.load_model(
dn.PATH_PROJECT +
exp.experiment_name + '.h5')
exp.save_geral_configs(
'Experiment Specific Configuration: '
+ exp.experiment_name)
exp.save_summary_model(lstm.model)
exp.predict_samples(
lstm, x_test, y_test)
del x_test, y_test
del lstm, exp
def generate_model_5(exp, name_model):
exp.pp_data.vocabulary_size = 5000
exp.pp_data.embedding_size = 300
exp.pp_data.max_posts = 1750
exp.pp_data.max_terms_by_post = 300
exp.pp_data.binary_classifier = True
exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT
exp.pp_data.remove_stopwords = False
exp.pp_data.delete_low_tfid = False
exp.pp_data.min_df = 0
exp.pp_data.min_tf = 0
exp.pp_data.random_posts = False
exp.pp_data.random_users = False
exp.pp_data.tokenizing_type = 'WE'
exp.pp_data.word_embedding_custom_file = ''
exp.pp_data.embedding_type = dn.EmbeddingType.WORD2VEC_CUSTOM
exp.pp_data.use_embedding = dn.UseEmbedding.NON_STATIC
exp.pp_data.word_embedding_custom_file = 'SMHD-CBOW-AllUsers-300.bin'
exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + str(exp.pp_data.use_embedding.value) + \
'_EF_' + exp.pp_data.word_embedding_custom_file
lstm = ModelClass(1)
lstm.loss_function = 'binary_crossentropy'
lstm.optmizer_function = 'adam'
lstm.epochs = 15
lstm.batch_size = 32
lstm.patience_train = 10
lstm.use_embedding_pre_train = exp.pp_data.use_embedding
lstm.embed_trainable = (lstm.use_embedding_pre_train == (
dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
neuronios_by_layer = [16]
epochs = [32]
batch_sizes = [40]
dropouts = [0.2]
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data(
)
exp.set_period_time_end(time_ini_rep, 'Load data')
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(
neuronios) + '_B' + str(batch_size) + '_E' + str(
epoch) + '_D' + str(dropout) + '_' + we_file_name
lstm.model = Sequential()
lstm.model.add(
Embedding(exp.pp_data.vocabulary_size,
exp.pp_data.embedding_size,
trainable=lstm.embed_trainable,
name='emb_' + name_model))
lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout,
recurrent_dropout=dropout,
return_sequences=True,
name='dense_1_' + name_model))
lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout,
recurrent_dropout=dropout,
return_sequences=True,
name='dense_2_' + name_model))
lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=dropout,
recurrent_dropout=dropout,
name='dense_3_' + name_model))
lstm.model.add(
Dense(1,
activation='sigmoid',
name='dense_4_' + name_model))
time_ini_exp = datetime.datetime.now()
exp.generate_model_hypeparams(lstm, x_train, y_train,
x_valid, y_valid,
embedding_matrix)
exp.set_period_time_end(time_ini_exp, 'Total experiment')
del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix
# Test
exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_test, y_test = exp.pp_data.load_data()
exp.save_data_format_train(x_test, name_model + '_x')
exp.save_data_format_train(y_test, name_model + '_y')
exp.set_period_time_end(time_ini_rep, 'Load data')
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(
neuronios) + '_B' + str(batch_size) + '_E' + str(
epoch) + '_D' + str(dropout) + '_' + we_file_name
lstm.model = exp.load_model(dn.PATH_PROJECT +
exp.experiment_name + '.h5')
exp.save_geral_configs()
exp.save_summary_model(lstm.model)
exp.predict_samples(lstm, x_test, y_test)
del x_test, y_test, lstm, exp
def generate_model_ml_le(exp, name_model, set_params): exp.pp_data.vocabulary_size = 5000 exp.pp_data.embedding_size = 300 exp.pp_data.max_posts = 1750 exp.pp_data.max_terms_by_post = 300 exp.pp_data.binary_classifier = True exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT exp.pp_data.remove_stopwords = False exp.pp_data.delete_low_tfid = False exp.pp_data.min_df = 0 exp.pp_data.min_tf = 0 exp.pp_data.random_posts = False # False = chronological order exp.pp_data.random_users = False exp.pp_data.tokenizing_type = 'WE' exp.pp_data.word_embedding_custom_file = '' exp.pp_data.embedding_type = dn.EmbeddingType.GLOVE_6B exp.pp_data.use_embedding = dn.UseEmbedding.STATIC exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL exp.use_custom_metrics = False exp.use_valid_set_for_train = True exp.valid_split_from_train_set = 0.0 exp.imbalanced_classes = False we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + str(exp.pp_data.use_embedding.value) + \ '_EF_' + 'glove6B300d' ## Load model according configuration lstm = ModelClass(1) lstm.loss_function = 'binary_crossentropy' lstm.optmizer_function = 'adam' lstm.epochs = 15 lstm.batch_size = 32 lstm.patience_train = 10 lstm.use_embedding_pre_train = exp.pp_data.use_embedding lstm.embed_trainable = (lstm.use_embedding_pre_train == (dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC)) # neuronios_by_layer = [16] # epochs = [32] # batch_sizes = [40] # dropouts = [0.2] neuronios_by_layer = set_params['neuronios_by_layer'] epochs = set_params['epochs'] batch_sizes = set_params['batch_sizes'] dropouts = set_params['dropouts'] np.random.seed(dn.SEED) time_ini_rep = datetime.datetime.now() x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data() exp.set_period_time_end(time_ini_rep, 'Load data') for neuronios in neuronios_by_layer: for batch_size in batch_sizes: for epoch in epochs: for dropout in dropouts: lstm.epochs = epoch lstm.batch_size = batch_size lstm.patience_train = epoch / 2 exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + str( batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_' + we_file_name lstm.model = Sequential() lstm.model.add(Embedding(exp.pp_data.vocabulary_size, exp.pp_data.embedding_size, trainable=lstm.embed_trainable, name='emb_' + name_model)) lstm.model.add(LSTM(neuronios, kernel_initializer='lecun_uniform', recurrent_initializer='orthogonal', activation='tanh', dropout=dropout, recurrent_dropout=dropout, return_sequences=True, name='dense_1_' + name_model)) lstm.model.add(LSTM(neuronios, kernel_initializer='lecun_uniform', recurrent_initializer='orthogonal', activation='tanh', dropout=dropout, recurrent_dropout=dropout, return_sequences=True, name='dense_2_' + name_model)) lstm.model.add(LSTM(neuronios, kernel_initializer='lecun_uniform', recurrent_initializer='orthogonal', activation='tanh', dropout=dropout, recurrent_dropout=dropout, name='dense_3_' + name_model)) lstm.model.add(Dense(3, kernel_initializer='lecun_uniform', activation='sigmoid', name='dense_4_' + name_model)) time_ini_exp = datetime.datetime.now() exp.generate_model_hypeparams(lstm, x_train, y_train, x_valid, y_valid, embedding_matrix) exp.set_period_time_end(time_ini_exp, 'Total experiment') del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix # Test exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL np.random.seed(dn.SEED) time_ini_rep = datetime.datetime.now() x_test, y_test = exp.pp_data.load_data() exp.set_period_time_end(time_ini_rep, 'Load data') for neuronios in neuronios_by_layer: for batch_size in batch_sizes: for epoch in epochs: for dropout in dropouts: lstm.epochs = epoch lstm.batch_size = batch_size lstm.patience_train = epoch / 2 exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + str( batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_' + we_file_name lstm.model = exp.load_model(dn.PATH_PROJECT + exp.experiment_name + '.h5') exp.save_geral_configs() exp.save_summary_model(lstm.model) exp.predict_samples(lstm, x_test, y_test) del x_test, y_test, lstm, exp
def load_submodel_anx(exp, name_model, kernel_name, set_params):
exp.pp_data.vocabulary_size = 5000
exp.pp_data.embedding_size = 300
exp.pp_data.max_posts = 1750
exp.pp_data.max_terms_by_post = 300
exp.pp_data.binary_classifier = True
exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT
exp.pp_data.remove_stopwords = False
exp.pp_data.delete_low_tfid = False
exp.pp_data.min_df = 0
exp.pp_data.min_tf = 0
exp.pp_data.random_posts = False
exp.pp_data.random_users = False
exp.pp_data.tokenizing_type = 'WE'
exp.pp_data.embedding_type = dn.EmbeddingType.GLOVE_CUSTOM
exp.pp_data.use_embedding = dn.UseEmbedding.STATIC
exp.pp_data.word_embedding_custom_file = 'SMHD-glove-A-D-ADUsers-300.pkl'
exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL
exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + str(exp.pp_data.use_embedding.value) + \
'_EF_' + exp.pp_data.word_embedding_custom_file.split('.')[0] + kernel_name
lstm = ModelClass(1)
lstm.loss_function = 'binary_crossentropy'
lstm.optmizer_function = 'adam'
lstm.epochs = 15
lstm.batch_size = 32
lstm.patience_train = 10
lstm.use_embedding_pre_train = exp.pp_data.use_embedding
lstm.embed_trainable = (lstm.use_embedding_pre_train == (dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
# neuronios_by_layer = [16]
# epochs = [96]
# batch_sizes = [20]
# dropouts = [0.1]
neuronios_by_layer = set_params['neuronios_by_layer']
epochs = set_params['epochs']
batch_sizes = set_params['batch_sizes']
dropouts = set_params['dropouts']
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data()
exp.set_period_time_end(time_ini_rep, 'Load data')
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + str(
batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_' + we_file_name
lstm.model = Sequential()
# Embedding(tamanho_vocabulario, embedding_size, input_length=max_len), onde max_len
# representa o corte nos textos após max_len tokens.
lstm.model.add(Embedding(exp.pp_data.vocabulary_size, exp.pp_data.embedding_size,
trainable=lstm.embed_trainable, name='emb_' + name_model))
lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout, recurrent_dropout=dropout,
return_sequences=True, name='dense_1_' + name_model))
lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout, recurrent_dropout=dropout,
return_sequences=True, name='dense_2_' + name_model))
lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout, recurrent_dropout=dropout,
name='dense_3_' + name_model))
lstm.model.add(Dense(3,
activation='sigmoid',
name='dense_4_' + name_model))
time_ini_exp = datetime.datetime.now()
exp.generate_model_hypeparams(lstm, x_train, y_train, x_valid, y_valid, embedding_matrix)
exp.set_period_time_end(time_ini_exp, 'Total experiment')
del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix
# Test
exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_test, y_test = exp.pp_data.load_data()
exp.set_period_time_end(time_ini_rep, 'Load data')
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
exp.experiment_name = name_model + '_lstm_exp9_var_L3' + '_N' + str(neuronios) + '_B' + str(
batch_size) + '_E' + str(epoch) + '_D' + str(dropout) + '_' + we_file_name
lstm.model = exp.load_model(dn.PATH_PROJECT + exp.experiment_name + '.h5')
exp.save_geral_configs()
exp.save_summary_model(lstm.model)
exp.predict_samples(lstm, x_test, y_test)
del x_test, y_test, lstm, exp
exp.pp_data.binary_classifier = True exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT exp.pp_data.remove_stopwords = False exp.pp_data.delete_low_tfid = False exp.pp_data.min_df = 0 exp.pp_data.min_tf = 0 exp.pp_data.random_posts = False exp.pp_data.random_users = False exp.pp_data.tokenizing_type = 'WE' exp.pp_data.use_embedding = dn.UseEmbedding.STATIC exp.pp_data.embedding_type = dn.EmbeddingType.GLOVE_6B lstm = ModelClass(1) lstm.loss_function = 'binary_crossentropy' lstm.optmizer_function = 'adam' lstm.epochs = 15 lstm.batch_size = 32 lstm.patience_train = 10 lstm.use_embedding_pre_train = exp.pp_data.use_embedding lstm.embed_trainable = False neuronios_by_layer = [16, 32, 64, 100, 128] epochs = [16, 32, 64, 100] batch_sizes = [20, 40, 80, 160] dropouts = [0.2, 0.3, 0.5] for neuronios in neuronios_by_layer: for batch_size in batch_sizes: for epoch in epochs: for dropout in dropouts: lstm.epochs = epoch
epochs = [16, 32, 64, 96, 128]
batch_sizes = [20, 40, 80]
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data(
)
exp.set_period_time_end(time_ini_rep, 'Load data')
for neuronios in neuronios_by_layer:
for batch_size in batch_sizes:
for epoch in epochs:
exp.experiment_name = 'lstm_exp14_L3' + '_N' + str(
neuronios) + '_B' + str(batch_size) + '_E' + str(epoch)
lstm.epochs = epoch
lstm.batch_size = batch_size
lstm.patience_train = epoch / 2
data_dim = exp.pp_data.max_terms_by_post
timesteps = exp.pp_data.max_posts
lstm.model = Sequential()
lstm.model.add(
LSTM(neuronios,
activation='tanh',
dropout=0.2,
recurrent_dropout=0.2,
return_sequences=True,
stateful=True,
batch_input_shape=(batch_size, timesteps, data_dim)))
def generate_model(exp, name_model, kernel_function, set_params):
# Configura pre-processamento dos dados para importação
exp.pp_data.vocabulary_size = 5000
exp.pp_data.embedding_size = 300 # 300 obrigatório se for usar word_embedding word2vec google neg300
exp.pp_data.max_posts = 1750
exp.pp_data.max_terms_by_post = 300
exp.pp_data.binary_classifier = True
exp.pp_data.format_input_data = dn.InputData.POSTS_ONLY_TEXT
exp.pp_data.remove_stopwords = False
exp.pp_data.delete_low_tfid = False
exp.pp_data.min_df = 0
exp.pp_data.min_tf = 0
exp.pp_data.random_posts = False # False = ordem cronológica
exp.pp_data.random_users = False # Não usada, as amostras são sempre random no validation k-fold
exp.pp_data.tokenizing_type = 'WE'
exp.pp_data.word_embedding_custom_file = ''
exp.pp_data.embedding_type = dn.EmbeddingType.GLOVE_6B
exp.pp_data.use_embedding = dn.UseEmbedding.STATIC
exp.pp_data.word_embedding_custom_file = ''
exp.pp_data.load_dataset_type = dn.LoadDataset.TRAIN_DATA_MODEL
exp.pp_data.type_prediction_label = dn.TypePredictionLabel.MULTI_LABEL_CATEGORICAL
exp.use_custom_metrics = False
exp.use_valid_set_for_train = True
exp.valid_split_from_train_set = 0.0
exp.imbalanced_classes = False
we_file_name = 'ET_' + str(exp.pp_data.embedding_type.value) + '_UE_' + str(exp.pp_data.use_embedding.value) + \
'_EF_' + 'glove6B300d' + kernel_function
## Gera dados conforme configuração
cnn_lstm = ModelClass(1)
cnn_lstm.loss_function = 'binary_crossentropy'
cnn_lstm.optmizer_function = 'adadelta'
cnn_lstm.epochs = 15
cnn_lstm.batch_size = 32
cnn_lstm.patience_train = 10
cnn_lstm.use_embedding_pre_train = exp.pp_data.use_embedding
cnn_lstm.embed_trainable = (cnn_lstm.use_embedding_pre_train == (dn.UseEmbedding.RAND or dn.UseEmbedding.NON_STATIC))
# set_params is empty
if not bool(set_params):
filters_by_layer = [32, 64, 128]
neuronios_by_lstm_layer = [64, 128, 256]
dropouts = [0.2, 0.5]
dropouts_lstm = [0.2, 0.5]
else:
filters_by_layer = set_params['filters_by_layer']
neuronios_by_lstm_layer = set_params['neuronios_by_lstm_layer']
dropouts = set_params['dropouts']
dropouts_lstm = set_params['dropouts_lstm']
kernels_size = [5]
epochs = [10]
batch_sizes = [20]
# Expected input batch shape: (batch_size, timesteps, data_dim)
# Note that we have to provide the full batch_input_shape since the network is stateful.
# the sample of index i in batch k is the follow-up for the sample i in batch k-1.
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_train, y_train, x_valid, y_valid, num_words, embedding_matrix = exp.pp_data.load_data()
exp.set_period_time_end(time_ini_rep, 'Load data')
for filter in filters_by_layer:
for kernel_size in kernels_size:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for dropout_lstm in dropouts_lstm:
for neuronios in neuronios_by_lstm_layer:
cnn_lstm.epochs = epoch
cnn_lstm.batch_size = batch_size
cnn_lstm.patience_train = epoch
exp.experiment_name = name_model + '_cnn_lstm' + '_F' + str(filter) + '_K' + \
str(kernel_size) + '_P' + 'None' + '_B' + str(batch_size) + \
'_E' + str(epoch) + '_D' + str(dropout) + '_HLN' + \
str(filter) + '_LSTM_N' + str(neuronios) + \
'_D'+ str(dropout_lstm) + '_' + we_file_name
cnn_lstm.model = Sequential()
cnn_lstm.model.add(Embedding(exp.pp_data.vocabulary_size, exp.pp_data.embedding_size,
trainable=cnn_lstm.embed_trainable, name='emb_' + name_model))
cnn_lstm.model.add(Dropout(dropout, name='dropout_1_' + name_model))
cnn_lstm.model.add(Conv1D(filters=filter, kernel_size=kernel_size,
kernel_initializer='glorot_uniform',
# kernel_regularizer=regularizers.l2(0.03),
padding='valid', activation='relu',
name='conv_1_' + name_model))
cnn_lstm.model.add(MaxPooling1D(name='max_pool_1_' + name_model))
cnn_lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout_lstm,
recurrent_dropout=dropout_lstm,
return_sequences=True, name='lstm_1_' + name_model))
cnn_lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout_lstm,
recurrent_dropout=dropout_lstm,
return_sequences=True, name='lstm_2_' + name_model))
cnn_lstm.model.add(LSTM(neuronios,
activation='tanh', dropout=dropout_lstm,
recurrent_dropout=dropout_lstm,
name='lstm_3_' + name_model))
cnn_lstm.model.add(Dense(3, activation='sigmoid', name='dense_1_' + name_model))
time_ini_exp = datetime.datetime.now()
exp.generate_model_hypeparams(cnn_lstm, x_train, y_train, x_valid, y_valid, embedding_matrix)
exp.set_period_time_end(time_ini_exp, 'Total experiment')
del x_train, y_train, x_valid, y_valid, num_words, embedding_matrix
# Test
exp.pp_data.load_dataset_type = dn.LoadDataset.TEST_DATA_MODEL
np.random.seed(dn.SEED)
time_ini_rep = datetime.datetime.now()
x_test, y_test = exp.pp_data.load_data()
exp.set_period_time_end(time_ini_rep, 'Load data')
for filter in filters_by_layer:
for kernel_size in kernels_size:
for batch_size in batch_sizes:
for epoch in epochs:
for dropout in dropouts:
for dropout_lstm in dropouts_lstm:
for neuronios in neuronios_by_lstm_layer:
cnn_lstm.epochs = epoch
cnn_lstm.batch_size = batch_size
cnn_lstm.patience_train = epoch
exp.experiment_name = name_model + '_cnn_lstm' + '_F' + str(filter) + '_K' + \
str(kernel_size) + '_P' + 'None' + '_B' + str(batch_size) + \
'_E' + str(epoch) + '_D' + str(dropout) + '_HLN' + \
str(filter) + '_LSTM_N' + str(neuronios) + \
'_D'+ str(dropout_lstm) + '_' + we_file_name
cnn_lstm.model = exp.load_model(dn.PATH_PROJECT + exp.experiment_name + '.h5')
exp.save_geral_configs('Experiment Specific Configuration: ' + exp.experiment_name)
exp.save_summary_model(cnn_lstm.model)
exp.predict_samples(cnn_lstm, x_test, y_test)
del x_test, y_test, cnn_lstm, exp