def get_model(feature_encoder, num_filter, region_size, num_labels,
**kwargs):
onehot_cnn = OnehotBowCNN(
rand_seed=1377,
verbose=kwargs.get('verbose', 0),
feature_encoder=feature_encoder,
# optimizers='adadelta',
optimizers='sgd',
num_labels=num_labels,
l1_conv_filter_type=[
[
num_filter, region_size, -1, 'valid', (-1, 1), 0., 'none',
'none'
],
],
l2_conv_filter_type=[],
full_connected_layer_units=[[0, 0, 'relu', 'batch_normalization']],
embedding_dropout_rate=0.,
nb_epoch=30,
nb_batch=32,
earlyStoping_patience=30,
lr=1e-2,
)
return onehot_cnn
def cross_validation(
train_data=None,
test_data=None,
cv_data=None,
input_length=None,
feature_type='word',
num_filter_list=None,
region_size_list=None,
word2vec_to_solve_oov=False,
word2vec_model_file_path=None,
verbose=0,
):
from data_processing_util.cross_validation_util import transform_cv_data, get_k_fold_data, get_val_score
# 1. 获取交叉验证的数据
if cv_data is None:
assert train_data is not None, 'cv_data和train_data必须至少提供一个！'
cv_data = get_k_fold_data(
k=3,
train_data=train_data,
test_data=test_data,
include_train_data=True,
)
# 2. 将数据进行特征编码转换
feature_encoder = OnehotBowCNN.get_feature_encoder(
input_length=input_length,
verbose=verbose,
feature_type=feature_type,
word2vec_to_solve_oov=word2vec_to_solve_oov,
word2vec_model_file_path=word2vec_model_file_path,
)
cv_data = transform_cv_data(feature_encoder, cv_data, verbose=0)
# 交叉验证
for num_filter in num_filter_list:
for region_size in region_size_list:
print('=' * 40)
print('num_filter and region_size is %d,%d.' %
(num_filter, region_size))
get_val_score(OnehotCNNWithOneConv,
cv_data=cv_data,
verbose=verbose,
region_size=region_size,
num_filter=num_filter,
num_labels=24)
def cross_validation(
train_data=None,
test_data=None,
cv_data=None,
input_length =None,
feature_type = 'word',
num_filter_list=None,
region_size_list=None,
word2vec_to_solve_oov = False,
word2vec_model_file_path = None,
verbose = 0,
):
from data_processing_util.cross_validation_util import transform_cv_data, get_k_fold_data, get_val_score
# 1. 获取交叉验证的数据
if cv_data is None:
assert train_data is not None, 'cv_data和train_data必须至少提供一个！'
cv_data = get_k_fold_data(
k=3,
train_data=train_data,
test_data=test_data,
include_train_data=True,
)
# 2. 将数据进行特征编码转换
feature_encoder = OnehotBowCNN.get_feature_encoder(
input_length=input_length,
verbose=verbose,
feature_type=feature_type,
word2vec_to_solve_oov = word2vec_to_solve_oov,
word2vec_model_file_path=word2vec_model_file_path,
)
cv_data = transform_cv_data(feature_encoder, cv_data,verbose=0)
# 交叉验证
for num_filter in num_filter_list:
for region_size in region_size_list:
print('=' * 40)
print('num_filter and region_size is %d,%d.'%(num_filter,region_size))
get_val_score(OnehotCNNWithOneConv,
cv_data=cv_data,
verbose=verbose,
region_size = region_size,
num_filter=num_filter,
num_labels=24
)
print('bow length:%d'%bow_length)
print('='*80)
onehot_cnn = OnehotBowCNN(
rand_seed=rand_seed,
verbose=0,
feature_encoder=feature_encoder,
# optimizers='adadelta',
optimizers='sgd',
input_length=sentence_padding_length,
input_dim=feature_encoder.vocabulary_size,
num_labels=len(index_to_label),
l1_conv_filter_type=[
[layer1, 2, -1, 'bow', (2, 1), 0., 'relu', 'batch_normalization'],
# [layer1, 3, -1, 'bow', (2, 1), 0., 'relu', 'batch_normalization'],
# [1000, 4, -1, 'valid', (2, 1), 0., 'relu', 'batch_normalization'],
],
l2_conv_filter_type=[
# [layer2, 2, -1, 'valid', (2, 1), 0.25, 'relu', 'none']
],
full_connected_layer_units=[
# (hidden1, 0.5, 'relu', 'batch_normalization'),
# (hidden2, 0.5, 'relu', 'none'),
],
embedding_dropout_rate=0.,
nb_epoch=30,
nb_batch=5,
earlyStoping_patience=20,
lr=1e-2,
)
onehot_cnn.print_model_descibe()
# +++++++++++++ region start : 3、构建onehot编码 +++++++++++++
# ****************************************************************
logging.debug('=' * 20)
logging.debug('3、构建onehot编码')
onehot_cnn = OnehotBowCNN(
rand_seed=1337,
verbose=config['verbose'],
feature_encoder=feature_encoder,
# optimizers='adadelta',
optimizers=config['optimizers'],
input_length=sentence_padding_length,
num_labels=len(index_to_label),
conv1_filter_type=config['l1_conv_filter_type'],
conv2_filter_type=config['l2_conv_filter_type'],
full_connected_layer_units=config['full_connected_layer_units'],
output_dropout_rate=config['output_dropout_rate'],
nb_epoch=config['nb_epoch'],
nb_batch=config['nb_batch'],
earlyStoping_patience=config['earlyStoping_patience'],
lr=config['lr'],
)
onehot_cnn.print_model_descibe()
onehot_cnn.fit((train_X_feature, train_y),
(test_X_feature, test_y))
onehot_cnn.accuracy((train_X_feature, train_y), transform_input=False)
logging.debug('=' * 20)
# ****************************************************************
logging.debug('=' * 20)
logging.debug('3、构建onehot编码')
from deep_learning.cnn.wordEmbedding_cnn.onehot_cnn_model import OnehotBowCNN
onehot_cnn = OnehotBowCNN(
rand_seed=config['rand_seed'],
verbose=config['verbose'],
feature_encoder=feature_encoder,
# optimizers='adadelta',
optimizers=config['optimizers'],
input_length=sentence_padding_length,
num_labels=len(index_to_label),
conv1_filter_type=config['l1_conv_filter_type'],
conv2_filter_type=config['l2_conv_filter_type'],
full_connected_layer_units=config['full_connected_layer_units'],
output_dropout_rate=config['output_dropout_rate'],
nb_epoch=config['nb_epoch'],
batch_size=config['nb_batch'],
earlyStoping_patience=config['earlyStoping_patience'],
lr=config['lr'],
)
# np.random.seed(0)
print(np.random.randint(0,100))
# quit()
onehot_cnn.print_model_descibe()
onehot_cnn.fit((train_X_feature, train_y),