def build_(method=1, beta=None):
tf.reset_default_graph()
# Training
if method == 1:
classifier = DBN(
hidden_act_func=['tanh', 'gauss'],
output_act_func='affine',
loss_func='mse', # gauss 激活函数会自动转换为 mse 损失函数
struct=[
x_dim, x_dim * 40, x_dim * 20, x_dim * 10, x_dim * 2, y_dim
],
lr=1e-4,
use_for='prediction',
bp_algorithm='rmsp',
epochs=240,
batch_size=32,
dropout=0.08,
units_type=['gauss', 'gauss'],
rbm_lr=1e-4,
rbm_epochs=60,
cd_k=1,
pre_train=True)
elif method == 2:
classifier = supervised_sAE(
output_func='gauss',
hidden_func='affine',
loss_func='mse',
struct=[
x_dim, x_dim * 40, x_dim * 20, x_dim * 10, x_dim * 2, y_dim
],
lr=1e-4,
use_for='prediction',
epochs=180,
batch_size=32,
dropout=0.15,
ae_type='sae', # ae | dae | sae
act_type=[
'gauss', 'affine'
], # decoder：[sigmoid] with ‘cross_entropy’ | [affine] with ‘mse’
noise_type='mn', # Gaussian noise (gs) | Masking noise (mn)
beta=beta, # DAE：噪声损失系数 | SAE：稀疏损失系数 | YAE：Y系数比重
p=0.1, # DAE：样本该维作为噪声的概率 | SAE稀疏性参数：期望的隐层平均活跃度（在训练批次上取平均）
ae_lr=1e-4,
ae_epochs=60,
pre_train=True)
return classifier
def run_(method=1, beta=None):
tf.reset_default_graph()
# Training
if method == 1:
classifier = DBN(
hidden_act_func='gauss',
output_act_func='gauss',
loss_func='mse', # gauss 激活函数会自动转换为 mse 损失函数
struct=[x_dim, y_dim * 20, y_dim * 10, y_dim],
# struct=[x_dim, int(x_dim/2), int(x_dim/4), int(x_dim/8), y_dim],
lr=1e-4,
use_for='classification',
bp_algorithm='rmsp',
epochs=240,
batch_size=16,
dropout=0.05,
units_type=['gauss', 'gauss'],
rbm_lr=1e-4,
rbm_epochs=45,
cd_k=1,
pre_train=True)
elif method == 2:
classifier = supervised_sAE(
output_func='gauss',
hidden_func='gauss',
loss_func='mse',
struct=[x_dim, y_dim * 60, y_dim * 30, y_dim * 10, y_dim],
lr=1e-4,
use_for='classification',
epochs=240,
batch_size=32,
dropout=0.34,
ae_type='yae', # ae | dae | sae
act_type=[
'gauss', 'affine'
], # decoder：[sigmoid] with ‘cross_entropy’ | [affine] with ‘mse’
noise_type='mn', # Gaussian noise (gs) | Masking noise (mn)
beta=beta, # DAE：噪声损失系数 | SAE：稀疏损失系数 | YAE：Y系数比重
p=0.3, # DAE：样本该维作为噪声的概率 | SAE稀疏性参数：期望的隐层平均活跃度（在训练批次上取平均）
ae_lr=1e-4,
ae_epochs=30,
pre_train=True)
run_sess(classifier, datasets, filename, load_saver='')
return classifier
epochs=20,
img_shape=[p_dim, p_dim],
channels=[1, 6, 6, 64, y_dim], # 前几维给 ‘Conv’ ，后几维给 ‘Full connect’
layer_tp=['C', 'P', 'C', 'P'],
fsize=[[4, 4], [3, 3]],
ksize=[[2, 2], [2, 2]],
batch_size=32,
dropout=0.2)
if select_case == 3:
classifier = supervised_sAE(
output_func='softmax',
hidden_func='sigmoid',
loss_func='cross_entropy',
struct=[x_dim, 400, 200, 100, y_dim],
lr=1e-3,
use_for='classification',
epochs=10,
batch_size=32,
dropout=0.12,
ae_type='ae', # ae | dae | sae
act_type=[
'sigmoid', 'sigmoid'
], # decoder：[sigmoid] with ‘cross_entropy’ | [affine] with ‘mse’
noise_type='mn', # Gaussian noise (gs) | Masking noise (mn)
beta=0.5, # DAE：噪声损失系数 | SAE：稀疏损失系数
p=0.3, # DAE：样本该维作为噪声的概率 / SAE稀疏性参数：期望的隐层平均活跃度（在训练批次上取平均）
ae_lr=1e-3,
ae_epochs=5,
pre_train=True)
run_sess(classifier, datasets, filename, load_saver='')
label_distribution = classifier.label_distribution
img_shape=[p_dim, p_dim],
channels=[1, 6, 6, 64, y_dim], # 前几维给 ‘Conv’ ，后几维给 ‘Full connect’
layer_tp=['C', 'P', 'C', 'P'],
fsize=[[4, 4], [3, 3]],
ksize=[[2, 2], [2, 2]],
batch_size=32,
dropout=1)
if select_case == 3:
classifier = supervised_sAE(
out_func='softmax',
en_func='affine', # encoder：[sigmoid] | [affine]
use_for='classification',
loss_func=
'mse', # decoder：[sigmoid] with ‘cross_entropy’ | [affine] with ‘mse’
ae_type='ae', # ae | dae | sae
noise_type='gs', # Gaussian noise (gs) | Masking noise (mn)
beta=0.6, # 惩罚因子权重（KL项 | 非噪声样本项）
p=0.01, # DAE：样本该维作为噪声的概率 / SAE稀疏性参数：期望的隐层平均活跃度（在训练批次上取平均）
sup_ae_struct=[x_dim, 200, 50, y_dim],
sup_ae_epochs=100,
ae_epochs=30,
batch_size=32,
ae_lr=1e-3,
dropout=1)
Initializer.sess_init_all(sess) # 初始化变量
summ = Summaries(os.path.basename(__file__), sess=sess)
classifier.train_model(X_train, Y_train, sess, summ)
# Test
print("[Test data...]")
Y_pred = classifier.test_model(X_test, Y_test, sess)
p_dim = int(np.sqrt(x_dim))
tf.reset_default_graph()
# Training
select_case = 3
if select_case == 3:
classifier = supervised_sAE(
output_func='softmax',
hidden_func='sigmoid', # encoder：[sigmoid] | [affine]
use_for='classification',
loss_func=
'cross_entropy', # decoder：[sigmoid] with ‘cross_entropy’ | [affine] with ‘mse’
struct=[196, 2000, 2000, 2],
lr=1e-3,
epochs=1645,
batch_size=1024,
dropout=0.01,
ae_type='ae', # ae | dae | sae
act_type=['sigmoid', 'affine'],
noise_type='mn', # Gaussian n oise (gs) | Masking noise (mn)
beta=0.25, # 惩罚因子权重（KL项 | 非噪声样本项）
p=0.5, # DAE：样本该维作为噪声的概率 / SAE稀疏性参数：期望的隐层平均活跃度（在训练批次上取平均）
ae_lr=1e-3,
ae_epochs=2000,
pre_train=True)
# run_sess(classifier,datasets,filename,load_saver='')#验证
# print("***********************")
run_sess(classifier, datasets1, filename, load_saver='f') # 测试
# label_distribution = classifier.label_distribution
# data = pd.read_csv(r"D:\gao\jinzhou\code1\saver\label_distribution.csv")
# data.to_csv(r"D:\gao\jinzhou\code1\saver\label_distribution_2_"+str(units[i])+"_"+str(random[j])+".csv",sep=",",index=False)