markilue
/
self_example
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

小论文

This commit is contained in:
kevinding1125 2023-11-09 19:45:21 +08:00
parent d8746d192f
commit 564ba3f669
14 changed files with 1050 additions and 64 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

229
TensorFlow_eaxmple/Model_train_test/RUL/good_model/dctLSTM/LSTMTest.py Normal file
View File

@ -0,0 +1,229 @@
# -*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/6/14 14:56
@Usage :
@Desc : 测试所实现的LSTM
'''
import tensorflow as tf
import numpy as np
from model.LSTM.DCTAttention_embed_LSTM import AttentionEmbedLSTMLayer as LSTMLayer
# from model.LSTM.LSTM import LSTMLayer as LSTMLayer
import matplotlib.pyplot as plt
from keras.callbacks import EarlyStopping
from model.LossFunction.FTMSE import FTMSE
import math
from sklearn.metrics import mean_absolute_error, mean_squared_error
from pylab import *
'''
超参数设置:
'''
hidden_num = 10 # LSTM细胞个数
feature = 10 # 一个点的维度
batch_size = 32
EPOCH = 1000
unit = 512 # LSTM的维度
predict_num = 50 # 预测个数
model_name = "dctLSTM"
save_name = r"self_{0}_hidden{1}_unit{2}_feature{3}_predict{4}.h5".format(model_name, hidden_num, unit, feature,
predict_num)
def getData(filter_num, dims):
# 数据读入
HI_merge_data_origin = np.load("../../2012轴承数据集预测挑战/HI_create/HI_merge_data.npy")
# plt.plot(HI_merge_data[0:1250, 1])
# 去除掉退化特征不明显前面的点
HI_merge_data = HI_merge_data_origin[0:1250, 1]
# plt.plot(HI_merge_data)
# plt.show()
(total_dims,) = HI_merge_data.shape
# # 将其分成重叠采样状态-滑动窗口函数
predict_data = np.empty(shape=[total_dims - filter_num, filter_num])
# 重叠采样获取时间部和训练次数
for dim in range(total_dims - filter_num):
predict_data[dim] = HI_merge_data[dim:dim + filter_num]
train_label = predict_data[dims:, :]
train_label_single = HI_merge_data[dims + filter_num - 1:-1]
# 再重叠采样获取一个点的维度
'''train_data.shape:(sample,filter_num) -> (sample,filter_num,dims)'''
# # 将其分成重叠采样状态-滑动窗口函数
train_data = np.empty(shape=[dims, total_dims - filter_num - dims, filter_num])
for dim in range(dims):
train_data[dim] = predict_data[dim:total_dims - filter_num - dims + dim, :]
# 转置变成想要的数据 (dims,sample,filter_num) -> (sample,filter_num,dims)
train_data = tf.transpose(train_data, [1, 2, 0])
# todo 解决模型保存时,query无法序列化的问题
total_data = tf.cast(HI_merge_data, dtype=tf.float32)
train_data = tf.cast(train_data, dtype=tf.float32)
train_label = tf.cast(train_label, dtype=tf.float32)
train_label_single = tf.cast(train_label_single, dtype=tf.float32)
print("total_data.shape:", total_data.shape)
print("train_data.shape:", train_data.shape) # (20, 1200, 30)
print("train_label.shape:", train_label.shape) # (20, 1200)
print("train_label_single.shape:", train_label_single.shape)
# 所有的原始数据;所有的训练数据;所有的训练标签(预测一个序列);所有的训练标签(预测一个点)
return total_data, train_data, train_label, train_label_single
'''
train_data.shape: (total_dims - filter_num - 1, filter_num,dims) :(570,600,30)
predict_data.shape: (total_dims - filter_num, filter_num) :(571,600,30)
train_label.shape: (total_dims - filter_num - 1, filter_num) :(570,600)
'''
def remove(train_data, train_label, batch_size):
epoch, _, _ = train_data.shape
size = int(epoch / batch_size)
return train_data[:size * batch_size], train_label[:size * batch_size]
'''
train_data.shape: (1230, 10, 10)
train_label.shape: (1230, 10)
train_label_single.shape: (1230,)
'''
def splitValData(data, label, label_single, predict_num=50):
sample, hidden, feature = data.shape
train_data = data[:sample - predict_num, :, :]
val_data = data[sample - predict_num:, :, :]
train_label = label[:sample - predict_num, :]
val_label = label[sample - predict_num:, :]
train_label_single = label_single[:sample - predict_num, ]
val_label_single = label_single[sample - predict_num:, ]
return train_data, val_data, train_label, val_label, train_label_single, val_label_single
def predict_model(filter_num, dims):
input = tf.keras.Input(shape=[filter_num, dims])
input = tf.cast(input, tf.float32)
#### 官方
# LSTM = tf.keras.layers.LSTM(units=512, return_sequences=True)(input)
# LSTM = tf.keras.layers.LSTM(units=256, return_sequences=False)(LSTM)
#### 自己
# LSTM = tf.keras.layers.Conv1D(512, kernel_size=8, padding='same')(input)
LSTM = LSTMLayer(units=512, return_sequences=True)(input)
LSTM = LSTMLayer(units=256, return_sequences=False)(LSTM)
x = tf.keras.layers.Dense(128, activation="relu")(LSTM)
x = tf.keras.layers.Dense(64, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dense(32, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dense(16, activation="relu")(x)
output = tf.keras.layers.Dense(1, activation="relu", name='output')(x)
model = tf.keras.Model(inputs=input, outputs=output)
return model
def split_data(train_data, train_label):
return train_data[:1150, :, :], train_label[:1150, :], train_data[-70:, :, :], train_label[-70:, :]
# 仅使用预测出来的最新的一个点预测以后
def predictOneByOne(newModel, train_data, predict_num=50):
# 取出训练数据的最后一条
each_predict_data = np.expand_dims(train_data[-1, :, :], axis=0)
predicted_list = np.empty(shape=(predict_num, 1)) # (5,filter_num,30)
# all_data = total_data # (1201,)
for each_predict in range(predict_num):
# predicted_data.shape : (1,1)
predicted_data = newModel.predict(each_predict_data) # (batch_size,filer_num,1)
predicted_list[each_predict] = predicted_data
# (1,1) => (10,1)
temp1 = np.transpose(np.concatenate([each_predict_data[:, 1:, -1], predicted_data], axis=1), [1, 0])
each_predict_data = np.expand_dims(
np.concatenate([np.squeeze(each_predict_data[:, :, 1:], axis=0), temp1], axis=1), axis=0)
return predicted_list
# 不使用预测的数据，直接使用已知的数据持续预测
def predictByEveryData(trained_model: tf.keras.Model, predict_data):
predicted_data = trained_model.predict(predict_data)
predicted_data = np.concatenate([np.expand_dims(total_data[:hidden_num + feature, ], axis=1), predicted_data],
axis=0)
data = predictOneByOne(trained_model, predict_data)
predicted_data = np.concatenate([predicted_data, data], axis=0)
return predicted_data
pass
if __name__ == '__main__':
# 数据读取
# 数据读入 --> 所有的原始数据;所有的训练数据;所有的训练标签(预测一个序列);所有的训练标签(预测一个点)
total_data, train_data, train_label, train_label_single = getData(hidden_num, feature)
# 根据预测的点数划分训练集和测试集(验证集)
train_data, val_data, train_label, val_label, train_label_single, val_label_single = splitValData(train_data,
train_label,
train_label_single,
predict_num=predict_num)
# # # #### TODO 训练
model = predict_model(hidden_num, feature)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=save_name,
monitor='val_loss',
verbose=2,
save_best_only=True,
mode='min')
lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=10, min_lr=0.001)
model.compile(optimizer=tf.optimizers.SGD(), loss=tf.losses.mse)
model.summary()
early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=20, mode='min', verbose=1)
history = model.fit(train_data, train_label_single, epochs=EPOCH, validation_data=(val_data, val_label_single),
shuffle=True, verbose=1,
callbacks=[checkpoint, lr_scheduler, early_stop])
#### TODO 测试
trained_model = tf.keras.models.load_model(save_name, custom_objects={'AttentionEmbedLSTMLayer': LSTMLayer})
# 使用已知的点进行预测
print("开始预测")
predicted_data = predictByEveryData(trained_model, train_data)
# 使用预测的点持续预测
# predicted_data = predictOneByOne(trained_model, total_data, train_data)
print("predicted_data:", predicted_data)
print("predicted_data.shape:", predicted_data.shape)
plt.figure(1)
plt.subplot(2, 1, 1)
plt.plot(total_data)
# plt.subplot(2, 1, 2)
plt.plot(predicted_data)
# plt.scatter()
plt.show()

12
TensorFlow_eaxmple/Model_train_test/RUL/otherIdea/lru/lru.py
View File

@ -2,7 +2,11 @@
# 线性循环单元（Linear Recurrent Unit）
# tensorflow 1.15 + bert4keras 0.11.4 测试通过
from bert4keras.layers import *
from tensorflow.keras.layers import Layer,Dense
import numpy as np
import tensorflow as tf
import tensorflow.keras.backend as K
class LRU(Layer):
@ -26,9 +30,9 @@ class LRU(Layer):
self.unroll = unroll
self.kernel_initializer = initializers.get(kernel_initializer)
@integerize_shape
def build(self, input_shape):
super(LRU, self).build(input_shape)
hidden_size = input_shape[-1]
self.i_dense = Dense(
units=self.units * 2,
@ -57,7 +61,7 @@ class LRU(Layer):
name='params_log', shape=(3, self.units), initializer=initializer
)
@recompute_grad
def call(self, inputs, mask=None):
u = self.i_dense(inputs)
params = K.exp(self.params_log)

239
TensorFlow_eaxmple/Model_train_test/RUL/test/AdamRNNTest.py Normal file
View File

@ -0,0 +1,239 @@
# -*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/6/14 14:56
@Usage :
@Desc : 测试所实现的LSTM
'''
import tensorflow as tf
import numpy as np
from model.AdamRNN.AdamRNN import AdaRNN
# from model.LSTM.LSTM import LSTMLayer as LSTMLayer
import matplotlib.pyplot as plt
from keras.callbacks import EarlyStopping
from model.LossFunction.FTMSE import FTMSE
import math
from sklearn.metrics import mean_absolute_error, mean_squared_error
from pylab import *
'''
超参数设置:
'''
hidden_num = 40 # LSTM细胞个数
feature = 10 # 一个点的维度
batch_size = 32
EPOCH = 1000
unit = 512 # LSTM的维度
predict_num = 50 # 预测个数
model_name = "adaRNN"
save_name = r"self_{0}_hidden{1}_unit{2}_feature{3}_predict{4}.h5".format(model_name, hidden_num, unit, feature,
predict_num)
def standardization(data):
mu = np.mean(data, axis=0)
sigma = np.std(data, axis=0)
return (data - mu) / sigma
def normalization(data):
_range = np.max(data) - np.min(data)
return (data - np.min(data)) / _range
# LSTM_cell的数目,维度,是否正则化
def getData(filter_num, dims, if_norm: bool = False):
# 数据读入
HI_merge_data_origin = np.load("../../2012轴承数据集预测挑战/HI_create/HI_merge_data.npy")
# plt.plot(HI_merge_data[0:1250, 1])
# 去除掉退化特征不明显前面的点
HI_merge_data = HI_merge_data_origin[0:1250, 1]
# 是否正则化
if if_norm:
HI_merge_data = normalization(HI_merge_data)
# plt.plot(HI_merge_data)
# plt.show()
(total_dims,) = HI_merge_data.shape
# # 将其分成重叠采样状态-滑动窗口函数
predict_data = np.empty(shape=[total_dims - filter_num, filter_num])
# 重叠采样获取时间部和训练次数
for dim in range(total_dims - filter_num):
predict_data[dim] = HI_merge_data[dim:dim + filter_num]
train_label = predict_data[dims:, :]
train_label_single = HI_merge_data[dims + filter_num - 1:-1]
# 再重叠采样获取一个点的维度
'''train_data.shape:(sample,filter_num) -> (sample,filter_num,dims)'''
# # 将其分成重叠采样状态-滑动窗口函数
train_data = np.empty(shape=[dims, total_dims - filter_num - dims, filter_num])
for dim in range(dims):
train_data[dim] = predict_data[dim:total_dims - filter_num - dims + dim, :]
# 转置变成想要的数据 (dims,sample,filter_num) -> (sample,filter_num,dims)
train_data = tf.transpose(train_data, [1, 2, 0])
# todo 解决模型保存时,query无法序列化的问题
total_data = tf.cast(HI_merge_data, dtype=tf.float32)
train_data = tf.cast(train_data, dtype=tf.float32)
train_label = tf.cast(train_label, dtype=tf.float32)
train_label_single = tf.cast(train_label_single, dtype=tf.float32)
print("total_data.shape:", total_data.shape)
print("train_data.shape:", train_data.shape) # (20, 1200, 30)
print("train_label.shape:", train_label.shape) # (20, 1200)
print("train_label_single.shape:", train_label_single.shape)
# 所有的原始数据;所有的训练数据;所有的训练标签(预测一个序列);所有的训练标签(预测一个点)
return total_data, train_data, train_label, train_label_single
'''
train_data.shape: (total_dims - filter_num - 1, filter_num,dims) :(570,600,30)
predict_data.shape: (total_dims - filter_num, filter_num) :(571,600,30)
train_label.shape: (total_dims - filter_num - 1, filter_num) :(570,600)
'''
def remove(train_data, train_label, batch_size):
epoch, _, _ = train_data.shape
size = int(epoch / batch_size)
return train_data[:size * batch_size], train_label[:size * batch_size]
'''
train_data.shape: (1230, 10, 10)
train_label.shape: (1230, 10)
train_label_single.shape: (1230,)
'''
def splitValData(data, label, label_single, predict_num=50):
sample, hidden, feature = data.shape
train_data = data[:sample - predict_num, :, :]
val_data = data[sample - predict_num:, :, :]
train_label = label[:sample - predict_num, :]
val_label = label[sample - predict_num:, :]
train_label_single = label_single[:sample - predict_num, ]
val_label_single = label_single[sample - predict_num:, ]
return train_data, val_data, train_label, val_label, train_label_single, val_label_single
def predict_model(filter_num, dims):
input = tf.keras.Input(shape=[filter_num, dims])
input = tf.cast(input, tf.float32)
x = tf.keras.layers.Dense(128, activation="relu")(LSTM)
x = tf.keras.layers.Dense(64, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dense(32, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dense(16, activation="relu")(x)
output = tf.keras.layers.Dense(1, activation="relu", name='output')(x)
model = tf.keras.Model(inputs=input, outputs=output)
return model
def split_data(train_data, train_label):
return train_data[:1150, :, :], train_label[:1150, :], train_data[-70:, :, :], train_label[-70:, :]
# 仅使用预测出来的最新的一个点预测以后
def predictOneByOne(newModel, train_data, predict_num=50):
# 取出训练数据的最后一条
each_predict_data = np.expand_dims(train_data[-1, :, :], axis=0)
predicted_list = np.empty(shape=(predict_num, 1)) # (5,filter_num,30)
# all_data = total_data # (1201,)
for each_predict in range(predict_num):
# predicted_data.shape : (1,1)
predicted_data = newModel.predict(each_predict_data) # (batch_size,filer_num,1)
predicted_list[each_predict] = predicted_data
# (1,1) => (10,1)
temp1 = np.transpose(np.concatenate([each_predict_data[:, 1:, -1], predicted_data], axis=1), [1, 0])
each_predict_data = np.expand_dims(
np.concatenate([np.squeeze(each_predict_data[:, :, 1:], axis=0), temp1], axis=1), axis=0)
return predicted_list
# 不使用预测的数据，直接使用已知的数据持续预测
def predictByEveryData(trained_model: tf.keras.Model, predict_data):
predicted_data = trained_model.predict(predict_data)
predicted_data = np.concatenate([np.expand_dims(total_data[:hidden_num + feature, ], axis=1), predicted_data],
axis=0)
data = predictOneByOne(trained_model, predict_data)
predicted_data = np.concatenate([predicted_data, data], axis=0)
return predicted_data
pass
if __name__ == '__main__':
# 数据读取
# 数据读入 --> 所有的原始数据;所有的训练数据;所有的训练标签(预测一个序列);所有的训练标签(预测一个点)
total_data, train_data, train_label, train_label_single = getData(hidden_num, feature)
# 根据预测的点数划分训练集和测试集(验证集)
train_data, val_data, train_label, val_label, train_label_single, val_label_single = splitValData(train_data,
train_label,
train_label_single,
predict_num=predict_num)
# # # #### TODO 训练
model = predict_model(hidden_num, feature)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=save_name,
monitor='val_loss',
verbose=2,
save_best_only=True,
mode='min')
lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=10, min_lr=0.0001)
model.compile(optimizer=tf.optimizers.SGD(), loss=tf.losses.mse)
model.summary()
early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=20, mode='min', verbose=1)
history = model.fit(train_data, train_label_single, epochs=EPOCH, validation_data=(val_data, val_label_single),
shuffle=True, verbose=1,
callbacks=[checkpoint, lr_scheduler, early_stop])
#### TODO 测试
trained_model = tf.keras.models.load_model(save_name, custom_objects={'AttentionEmbedLSTMLayer': LSTMLayer})
# 使用已知的点进行预测
print("开始预测")
predicted_data = predictByEveryData(trained_model, train_data)
# 使用预测的点持续预测
# predicted_data = predictOneByOne(trained_model, total_data, train_data)
print("predicted_data:", predicted_data)
print("predicted_data.shape:", predicted_data.shape)
plt.figure(1)
plt.subplot(2, 1, 1)
plt.plot(total_data)
# plt.subplot(2, 1, 2)
plt.plot(predicted_data)
# plt.scatter()
plt.show()

68
TensorFlow_eaxmple/Model_train_test/RUL/test/LSTMContinueTest.py
View File

@ -9,13 +9,13 @@
import tensorflow as tf
import numpy as np
from model.LSTM.LSTMByDense import LSTMLayer as LSTMLayer
import matplotlib.pyplot as plt
from keras.callbacks import EarlyStopping
from model.LossFunction.FTMSE import FTMSE
from model.LSTM.DCTAttention_embed_LSTM import AttentionEmbedLSTMLayer as LSTMLayer
# from model.LSTM.LSTM import LSTMLayer as LSTMLayer
from model.ChannelAttention.DCT_channelAttention import DCTChannelAttention
from model.ChannelAttention.Light_channelAttention import LightChannelAttention1 as LightChannelAttention
import math
from sklearn.metrics import mean_absolute_error, mean_squared_error
from pylab import *
@ -29,7 +29,7 @@ batch_size = 8
EPOCH = 1000
unit = 512 # LSTM的维度
predict_num = 50 # 预测个数
model_name = "FC_FTLSTM"
model_name = "dctLSTM"
save_name = r"selfMulti_{0}_hidden{1}_unit{2}_feature{3}_predict{4}.h5".format(model_name, hidden_num, unit,
feature,
predict_num)
@ -100,22 +100,13 @@ def getData(filter_num, dims, if_norm: bool = False):
return total_data, train_data, train_label, train_label_single
'''
train_data.shape: (total_dims - filter_num - 1, filter_num,dims) :(570,600,30)
predict_data.shape: (total_dims - filter_num, filter_num) :(571,600,30)
train_label.shape: (total_dims - filter_num - 1, filter_num) :(570,600)
'''
def remove(train_data, train_label, batch_size):
epoch, _, _ = train_data.shape
size = int(epoch / batch_size)
return train_data[:size * batch_size], train_label[:size * batch_size]
'''
train_data.shape: (1230, 10, 10)
train_label.shape: (1230, 10)
train_label_single.shape: (1230,)
'''
def splitValData(data, label, label_single, predict_num=50):
sample, hidden, feature = data.shape
@ -145,19 +136,16 @@ def predict_model_multi(filter_num, dims):
LSTM = LSTMLayer(units=512, return_sequences=True)(input)
# LSTM = LightChannelAttention()(LSTM)
LSTM = LSTMLayer(units=256, return_sequences=True)(LSTM)
LSTM = LightChannelAttention()(LSTM)
### flatten
x = tf.keras.layers.Flatten()(LSTM)
x = tf.keras.layers.Dense(128, activation="relu")(x)
x = tf.keras.layers.Dense(128, activation="relu")(LSTM)
x = tf.keras.layers.Dense(64, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dense(32, activation="relu")(x)
x = tf.keras.layers.Dropout(0.2)(x)
x = tf.keras.layers.BatchNormalization()(x)
# x = tf.keras.layers.Dense(16, activation="relu")(x)
output = tf.keras.layers.Dense(10, activation="relu", name='output')(x)
x = tf.keras.layers.Dense(16, activation="relu")(x)
output = tf.keras.layers.Dense(1, activation="relu", name='output')(x)
model = tf.keras.Model(inputs=input, outputs=output)
return model
@ -207,31 +195,31 @@ if __name__ == '__main__':
train_label_single,
predict_num=predict_num)
# # #### TODO 训练
# model = predict_model_multi(hidden_num, feature)
# checkpoint = tf.keras.callbacks.ModelCheckpoint(
# filepath=save_name,
# monitor='val_loss',
# verbose=2,
# save_best_only=True,
# mode='min')
# lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=20, min_lr=0.001)
#
# model.compile(optimizer=tf.optimizers.SGD(), loss=tf.losses.mse)
# # model.compile(optimizer=tf.optimizers.SGD(learning_rate=0.001), loss=FTMSE())
# model.summary()
# early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=100, mode='min', verbose=1)
#
# history = model.fit(train_data, train_label, epochs=EPOCH,
# batch_size=batch_size, validation_data=(val_data, val_label_single), shuffle=True, verbose=2,
# callbacks=[checkpoint, lr_scheduler, early_stop])
model = predict_model_multi(hidden_num, feature)
checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=save_name,
monitor='val_loss',
verbose=2,
save_best_only=True,
mode='min')
lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=10, min_lr=0.001)
model.compile(optimizer=tf.optimizers.SGD(), loss=tf.losses.mse)
# model.compile(optimizer=tf.optimizers.SGD(learning_rate=0.001), loss=FTMSE())
model.summary()
early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=20, mode='min', verbose=1)
history = model.fit(train_data, train_label, epochs=EPOCH, validation_data=(val_data, val_label),
shuffle=True, verbose=1,
callbacks=[checkpoint, lr_scheduler, early_stop])
#### TODO 测试
# trained_model = tf.keras.models.load_model(save_name, custom_objects={'LSTMLayer': LSTMLayer, 'FTMSE': FTMSE})
# todo 解决自定义loss无法导入的问题
trained_model = tf.keras.models.load_model(save_name, compile=False, custom_objects={'LSTMLayer': LSTMLayer,'LightChannelAttention1':LightChannelAttention})
trained_model.compile(optimizer=tf.optimizers.SGD(), loss=FTMSE())
trained_model = tf.keras.models.load_model(save_name, compile=False, custom_objects={'AttentionEmbedLSTMLayer': LSTMLayer})
# trained_model.compile(optimizer=tf.optimizers.SGD(), loss=FTMSE())
# 使用已知的点进行预测
predicted_data = predictByEveryData(trained_model, train_data)

42
TensorFlow_eaxmple/Model_train_test/RUL/test/LSTMTest.py
View File

@ -9,7 +9,8 @@
import tensorflow as tf
import numpy as np
from model.LSTM.LSTM import LSTMLayer as LSTMLayer
from model.LSTM.DCTAttention_embed_LSTM import AttentionEmbedLSTMLayer as LSTMLayer
# from model.LSTM.LSTM import LSTMLayer as LSTMLayer
import matplotlib.pyplot as plt
from keras.callbacks import EarlyStopping
@ -21,18 +22,30 @@ from pylab import *
'''
超参数设置:
'''
hidden_num = 10 # LSTM细胞个数
hidden_num = 40 # LSTM细胞个数
feature = 10 # 一个点的维度
batch_size = 32
EPOCH = 1000
unit = 512 # LSTM的维度
predict_num = 50 # 预测个数
model_name = "cnn_LSTM"
model_name = "dctLSTM"
save_name = r"self_{0}_hidden{1}_unit{2}_feature{3}_predict{4}.h5".format(model_name, hidden_num, unit, feature,
predict_num)
def getData(filter_num, dims):
def standardization(data):
mu = np.mean(data, axis=0)
sigma = np.std(data, axis=0)
return (data - mu) / sigma
def normalization(data):
_range = np.max(data) - np.min(data)
return (data - np.min(data)) / _range
# LSTM_cell的数目,维度,是否正则化
def getData(filter_num, dims, if_norm: bool = False):
# 数据读入
HI_merge_data_origin = np.load("../../2012轴承数据集预测挑战/HI_create/HI_merge_data.npy")
@ -40,6 +53,10 @@ def getData(filter_num, dims):
# 去除掉退化特征不明显前面的点
HI_merge_data = HI_merge_data_origin[0:1250, 1]
# 是否正则化
if if_norm:
HI_merge_data = normalization(HI_merge_data)
# plt.plot(HI_merge_data)
# plt.show()
(total_dims,) = HI_merge_data.shape
@ -125,8 +142,8 @@ def predict_model(filter_num, dims):
# LSTM = tf.keras.layers.LSTM(units=256, return_sequences=False)(LSTM)
#### 自己
LSTM = tf.keras.layers.Conv1D(512, kernel_size=8, padding='same')(input)
LSTM = LSTMLayer(units=512, return_sequences=True)(LSTM)
# LSTM = tf.keras.layers.Conv1D(512, kernel_size=8, padding='same')(input)
LSTM = LSTMLayer(units=512, return_sequences=True)(input)
LSTM = LSTMLayer(units=256, return_sequences=False)(LSTM)
x = tf.keras.layers.Dense(128, activation="relu")(LSTM)
@ -194,20 +211,23 @@ if __name__ == '__main__':
# verbose=2,
# save_best_only=True,
# mode='min')
# lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=20, min_lr=0.001)
# lr_scheduler = tf.keras.callbacks.ReduceLROnPlateau(monitor='val_loss', factor=0.2, patience=10, min_lr=0.0001)
#
# model.compile(optimizer=tf.optimizers.SGD(), loss=tf.losses.mse)
# model.summary()
# early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=100, mode='min', verbose=1)
#
# history = model.fit(train_data, train_label_single, epochs=EPOCH, validation_data=(val_data, val_label_single), shuffle=True, verbose=1,
# model.summary()
# early_stop = EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=20, mode='min', verbose=1)
#
# history = model.fit(train_data, train_label_single, epochs=EPOCH, validation_data=(val_data, val_label_single),
# shuffle=True, verbose=1,
# callbacks=[checkpoint, lr_scheduler, early_stop])
#### TODO 测试
trained_model = tf.keras.models.load_model(save_name, custom_objects={'LSTMLayer': LSTMLayer})
trained_model = tf.keras.models.load_model(save_name, custom_objects={'AttentionEmbedLSTMLayer': LSTMLayer})
# 使用已知的点进行预测
print("开始预测")
predicted_data = predictByEveryData(trained_model, train_data)
# 使用预测的点持续预测
# predicted_data = predictOneByOne(trained_model, total_data, train_data)

18
TensorFlow_eaxmple/Model_train_test/RUL/test/test.py Normal file
View File

@ -0,0 +1,18 @@
#-*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/9 10:26
@Usage :
@Desc :
'''
import tensorflow as tf
# 假设有两个形状为 (3,) 的张量
tensor1 = tf.constant([1, 2, 3])
tensor2 = tf.constant([4, 5, 6])
# 在新的轴上堆叠这两个张量
stacked_tensor = tf.stack([tensor1, tensor2],axis=-1)
print(stacked_tensor)

149
TensorFlow_eaxmple/Model_train_test/model/AdamRNN/AdamRNN.py Normal file
View File

@ -0,0 +1,149 @@
# -*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/9 16:42
@Usage :
@Desc :
'''
import tensorflow as tf
from model.LossFunction.TransferLoss import TransferLoss
class AdaRNN(tf.keras.Model):
def __init__(self, n_input=128, n_hiddens=[64, 64], n_output=6, len_seq=9, trans_loss='mmd'):
super(AdaRNN, self).__init__()
self.n_input = n_input
self.num_layers = len(n_hiddens)
self.hiddens = n_hiddens
self.n_output = n_output
self.trans_loss = trans_loss
self.len_seq = len_seq
self.features = tf.keras.Sequential()
for hidden in n_hiddens:
rnn = tf.keras.layers.GRU(
units=hidden,
return_sequences=True
)
self.features.add(rnn)
self.fc_out = tf.keras.layers.Dense(n_output, activation=None)
self.gate = []
for _ in range(len(n_hiddens)):
gate_weight = tf.keras.layers.Dense(len_seq, activation=None)
self.gate.append(gate_weight)
self.bn_lst = [tf.keras.layers.BatchNormalization() for _ in range(len(n_hiddens))]
self.softmax = tf.keras.layers.Softmax(axis=0)
# def init_layers(self):
# for gate_layer in self.gate:
# gate_layer.build((None, self.len_seq * self.hiddens[i] * 2))
def forward_pre_train(self, x, len_win=0):
# 两层GRU之后的结果,每层GRU之后的结果,每层GRU前后权重归一化之后的结果
out, out_list_all, out_weight_list = self.gru_features(x)
fea = out
fc_out = self.fc_out(fea[:, -1, :]).squeeze()
out_list_s, out_list_t = self.get_features(out_list_all)
loss_transfer = tf.zeros((1,))
for i in range(len(out_list_s)):
criterion_transder = TransferLoss(
loss_type=self.trans_loss)
h_start = 0
for j in range(h_start, self.len_seq, 1):
i_start = max(j - len_win, 0)
i_end = j + len_win if j + len_win < self.len_seq else self.len_seq - 1
for k in range(i_start, i_end + 1):
weight = out_weight_list[i][j]
loss_transfer = loss_transfer + weight * criterion_transder(
out_list_s[i][:, j, :], out_list_t[i][:, k, :])
return fc_out, loss_transfer, out_weight_list
def call(self, x, len_win=0, training=False):
# 两层GRU之后的结果,每层GRU之后的结果,每层GRU前后权重归一化之后的结果
out, out_list_all, out_weight_list = self.gru_features(x, training=training)
fea = out
fc_out = self.fc_out(fea[:, -1, :])
loss_transfer = tf.zeros((1,))
for i in range(len(out_list_all)):
criterion_transder = TransferLoss(
loss_type=self.trans_loss)
h_start = 0
for j in range(h_start, self.len_seq, 1):
i_start = max(j - len_win, 0)
i_end = j + len_win if j + len_win < self.len_seq else self.len_seq - 1
for k in range(i_start, i_end + 1):
weight = out_weight_list[i][j]
loss_transfer = loss_transfer + weight * criterion_transder.compute(
out_list_all[i][:, j, :], out_list_all[i][:, k, :])
return fc_out, loss_transfer, out_weight_list
def gru_features(self, x, training=False):
x_input = x
out = None
out_lis = []
out_weight_list = [] if (
self.model_type == 'AdaRNN') else None
for i in range(self.num_layers):
out = self.features[i](x_input, training=training)
x_input = out
out_lis.append(out)
if self.model_type == 'AdaRNN':
out_gate = self.process_gate_weight(x_input, i, training=training)
out_weight_list.append(out_gate)
return out, out_lis, out_weight_list
def process_gate_weight(self, out, index, training=False):
x_s = out[:, :out.shape[1] // 2] # 可以理解为LSTM的前半段
x_t = out[:, out.shape[1] // 2:] # 可以理解为LSTM的后半段
x_all = tf.concat((x_s, x_t), 2)
x_all = tf.reshape(x_all, (x_all.shape[0], -1))
weight = tf.sigmoid(self.bn_lst[index](self.gate[index](x_all)), training=training)
weight = tf.reduce_mean(weight, axis=0)
res = self.softmax(weight)
return res
def get_features(self, output_list):
fea_list_src, fea_list_tar = [], []
for fea in output_list:
fea_list_src.append(fea[:, :fea.shape[1] // 2])
fea_list_tar.append(fea[:, fea.shape[1] // 2:])
return fea_list_src, fea_list_tar
def forward_Boosting(self, x, weight_mat=None):
out, out_list_all, _ = self.gru_features(x, training=False)
fea = out
fc_out = self.fc_out(fea[:, -1, :])
out_list_all = out_list_all
out_list_s, out_list_t = self.get_features(out_list_all)
loss_transfer = tf.zeros((1,))
if weight_mat is None:
weight = (1.0 / self.len_seq *
tf.ones((self.num_layers, self.len_seq), dtype=tf.float32))
else:
weight = weight_mat
dist_mat = tf.zeros((self.num_layers, self.len_seq), dtype=tf.float32)
for i in range(len(out_list_s)):
criterion_transder = TransferLoss(
loss_type=self.trans_loss)
for j in range(self.len_seq):
loss_trans = criterion_transder(out_list_s[i][:, j, :], out_list_t[i][:, j, :])
loss_transfer = loss_transfer + weight[i, j] * loss_trans
dist_mat[i, j] = loss_trans
return fc_out, loss_transfer, dist_mat, weight
def update_weight_Boosting(self, weight_mat, dist_old, dist_new):
epsilon = 1e-12
dist_old = tf.stop_gradient(dist_old)
dist_new = tf.stop_gradient(dist_new)
ind = dist_new

8
TensorFlow_eaxmple/Model_train_test/model/AdamRNN/__init__.py Normal file
View File

@ -0,0 +1,8 @@
#-*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/9 16:42
@Usage :
@Desc :
'''

23
TensorFlow_eaxmple/Model_train_test/model/ChannelAttention/DCT_channelAttention.py
View File

@ -16,7 +16,6 @@ import matplotlib.pyplot as plt
from tensorflow.keras.layers import Dense, Dropout, ReLU, BatchNormalization
from scipy.fftpack import dct
# def dct(x, norm=None):
# """
# Discrete Cosine Transform, Type II (a.k.a. the DCT)
@ -55,16 +54,15 @@ from scipy.fftpack import dct
# return V
import tensorflow as tf
'''
参考：
[1] https://github.com/Zero-coder/FECAM/blob/main/layers/dctnet.py
[2] https://arxiv.org/pdf/2212.01209v1.pdf
'''
def sdct_tf(signals, frame_length, frame_step, window_fn=tf.signal.hamming_window):
"""Compute Short-Time Discrete Cosine Transform of `signals`.
@ -128,25 +126,34 @@ def isdct_tf(dcts, *, frame_step, frame_length=None, window_fn=tf.signal.hamming
signals = signals / window_signal
return signals
class DCTChannelAttention(layers.Layer):
def build(self, input_shape):
_, hidden, channel = input_shape
self.l1 = Dense(channel * 2, use_bias=False)
self.drop1 = Dropout(0.1)
self.relu = ReLU(0.1)
self.relu = ReLU()
self.l2 = Dense(channel, use_bias=False)
self.bn = BatchNormalization()
self.bn = BatchNormalization(axis=-1, epsilon=1e-6)
def call(self, inputs, **kwargs):
batch_size, hidden, channel = inputs.shape
list = []
stack_dct = tf.signal.dct(inputs, norm="ortho",axis=-1)
change = tf.transpose(inputs, [0, 2, 1])
stack_dct = tf.signal.dct(change, norm="ortho", axis=-1)
stack_dct = tf.transpose(stack_dct, [0, 2, 1])
# for i in range(channel):
# freq = tf.signal.dct(inputs[:, i, :], norm="ortho", axis=-1)
# freq = tf.signal.dct(inputs[:, :, i], norm="ortho", axis=-1)
# # print("freq-shape:",freq.shape)
# freq = tf.expand_dims(freq, axis=2)
# if i == 0:
# stack_dct = freq
# else:
# stack_dct = tf.concat([stack_dct, freq], axis=2)
# list.append(freq)
# stack_dct = tf.stack(list, dim=1)
# stack_dct = tf.stack(list, axis=-1)
lr_weight = self.bn(stack_dct)
lr_weight = self.l1(stack_dct)

80
TensorFlow_eaxmple/Model_train_test/model/ChannelAttention/DCT_channelAttentionV2.py Normal file
View File

@ -0,0 +1,80 @@
#-*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/8 19:39
@Usage :
@Desc :
'''
import tensorflow as tf
import numpy as np
# 定义DCT函数
def dct(x, norm=None):
"""
Discrete Cosine Transform, Type II (a.k.a. the DCT)
"""
x_shape = x.shape
N = x_shape[-1]
x = tf.reshape(x, [-1, N])
v = tf.concat([x[:, ::2], tf.reverse(x[:, 1::2], [1])], axis=1)
Vc = tf.signal.rfft(v, 1)
k = - tf.range(N, dtype=x.dtype) * np.pi / (2 * N)
W_r = tf.cos(k)
W_i = tf.sin(k)
V = Vc[:, :, 0] * W_r - Vc[:, :, 1] * W_i
if norm == 'ortho':
V[:, 0] /= np.sqrt(N) * 2
V[:, 1:] /= np.sqrt(N / 2) * 2
V = 2 * tf.reshape(V, x_shape)
return V
# 定义tf.keras版本的dct_channel_block
class DctChannelBlock(tf.keras.layers.Layer):
def __init__(self, channel):
super(DctChannelBlock, self).__init__()
self.fc = tf.keras.Sequential([
tf.keras.layers.Dense(channel * 2, use_bias=False),
tf.keras.layers.Dropout(0.1),
tf.keras.layers.ReLU(),
tf.keras.layers.Dense(channel, use_bias=False),
tf.keras.layers.Activation('sigmoid')
])
self.dct_norm = tf.keras.layers.LayerNormalization(axis=-1, epsilon=1e-6)
# def get_config(self):
# # 自定义层里面的属性
# config = (
# {
# 'units': self.units,
# 'return_sequences': self.return_sequences
# }
# )
# base_config = super(DctChannelBlock, self).get_config()
# return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, **kwargs):
x = inputs
b, c, l = x.shape
dct_list = []
for i in range(c):
freq = dct(x[:, i, :])
dct_list.append(freq)
stack_dct = tf.stack(dct_list, axis=1)
lr_weight = self.dct_norm(stack_dct)
lr_weight = self.fc(stack_dct)
lr_weight = self.dct_norm(lr_weight)
return x * lr_weight

120
TensorFlow_eaxmple/Model_train_test/model/LSTM/DCTAttention_embed_LSTM.py Normal file
View File

@ -0,0 +1,120 @@
# -*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/6/14 13:49
@Usage :
@Desc : 标准版LSTM
'''
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.keras.layers import Dense, Conv2D, Conv1D
from model.ChannelAttention.DCT_channelAttention import DCTChannelAttention as DctChannelBlock
from tensorflow.keras import *
import tensorflow.keras.layers as layers
class AttentionEmbedLSTMLayer(layers.Layer):
# 定义两个权重初始化方法，方便后续调用
k_ini = initializers.GlorotUniform()
b_ini = initializers.Zeros()
def __init__(self, units=30, return_sequences: bool = False, **kwargs):
super(AttentionEmbedLSTMLayer, self).__init__()
self.units = units
self.return_sequences = return_sequences
def get_params(self, num_inputs, num_outputs):
def _one(shape, name):
# return tf.Variable(tf.random.normal(shape=shape, stddev=0.01, mean=0, dtype=tf.float32))
return self.add_weight(shape=shape, name=name, initializer=tf.random_normal_initializer)
def _three(name1, name2):
return (_one(shape=(num_inputs + num_outputs, num_outputs), name=name1),
self.add_weight(shape=(num_outputs,), name=name2,
initializer=tf.zeros_initializer))
W_i, b_i = _three("W_i", "b_i") # 输入门参数
W_f, b_f = _three("W_f", "b_f") # 遗忘门参数
W_o, b_o = _three("W_o", "b_o") # 输出门参数
W_c, b_c = _three("W_c", "b_c") # 候选记忆细胞参数
# 输出层参数
return W_i, b_i, W_f, b_f, W_o, b_o, W_c, b_c
def get_config(self):
# 自定义层里面的属性
config = (
{
'units': self.units,
'return_sequences': self.return_sequences
}
)
base_config = super(AttentionEmbedLSTMLayer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def build(self, input_shape):
num_inputs, num_outputs = input_shape[-1], self.units
self.W_i, self.b_i, self.W_f, self.b_f, self.W_o, self.b_o, self.W_c, self.b_c = self.get_params(
num_inputs=num_inputs, num_outputs=num_outputs)
self.dctAttention = DctChannelBlock(num_inputs + num_outputs)
pass
def call(self, inputs, **kwargs):
epoch, hiddens, dims = inputs.shape
# print(filter_num, dims)
for hidden in range(hiddens):
new_input = inputs[:, hidden, :]
new_input = tf.expand_dims(new_input, axis=1)
if hidden != 0:
new_input = tf.concat([new_input, ht_1], axis=-1)
else:
new_input = tf.pad(new_input, [[0, 0], [0, 0], [0, self.units]])
new_input = self.dctAttention(new_input)
Wi = tf.matmul(new_input, self.W_i) + self.b_i
Wf = tf.matmul(new_input, self.W_f) + self.b_f
Wc = tf.matmul(new_input, self.W_c) + self.b_c
Wo = tf.matmul(new_input, self.W_o) + self.b_o
ft = tf.nn.sigmoid(Wf)
it = tf.nn.sigmoid(Wi)
ct_ = tf.nn.tanh(Wc)
ot = tf.nn.sigmoid(Wo)
if hidden != 0:
ct = tf.add(tf.multiply(ft, ct_1), tf.multiply(it, ct_))
else:
ct = tf.multiply(it, ct_)
ht = tf.multiply(tf.nn.tanh(ct), ot)
if self.return_sequences:
if hidden == 0:
output = ht
else:
output = tf.concat([output, ht], axis=1)
else:
if hidden == hiddens - 1:
output = tf.squeeze(ht, axis=1)
ht_1 = ht
ct_1 = ct
# output = tf.reshape(output, [-1, filter_num, units])
# print(output.shape)
return output
if __name__ == '__main__':
pass
# tf.keras.layers.LSTM(return_sequences=)

41
TensorFlow_eaxmple/Model_train_test/model/LossFunction/TransferLoss.py Normal file
View File

@ -0,0 +1,41 @@
# -*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/9 16:47
@Usage :
@Desc :
'''
from model.LossFunction.transfer.mmd import MMDLoss
import tensorflow as tf
import tensorflow.losses
class TransferLoss(tf.keras.losses.Loss):
def __init__(self, loss_type='cosine'):
"""
Supported loss_type: mmd(mmd_lin), mmd_rbf, coral, cosine, kl, js, mine, adv
"""
self.loss_type = loss_type
def call(self, X, Y):
"""Compute adaptation loss
Arguments:
X {tensor} -- source matrix
Y {tensor} -- target matrix
Returns:
[tensor] -- transfer loss
"""
if self.loss_type == 'mmd_lin' or self.loss_type == 'mmd':
mmdloss = MMDLoss(kernel_type='linear')
loss = mmdloss(X, Y)
elif self.loss_type == 'cosine' or self.loss_type == 'cos':
loss = 1 - tf.losses.cosine_similarity(X, Y)
elif self.loss_type == 'mmd_rbf':
mmdloss = MMDLoss(kernel_type='rbf')
loss = mmdloss(X, Y)
return loss

8
TensorFlow_eaxmple/Model_train_test/model/LossFunction/transfer/__init__.py Normal file
View File

@ -0,0 +1,8 @@
#-*- encoding:utf-8 -*-
'''
@Author : dingjiawen
@Date : 2023/11/9 16:54
@Usage :
@Desc :
'''

75
TensorFlow_eaxmple/Model_train_test/model/LossFunction/transfer/mmd.py Normal file
View File

@ -0,0 +1,75 @@
import tensorflow as tf
import numpy as np
class MMDLoss(tf.keras.losses.Loss):
def __init__(self, kernel_type='linear', kernel_mul=2.0, kernel_num=5):
super(MMDLoss, self).__init__()
self.kernel_type = kernel_type
self.kernel_mul = kernel_mul
self.kernel_num = kernel_num
def get_config(self):
# 自定义层里面的属性
config = (
{
'kernel_type': self.kernel_type,
'kernel_mul': self.kernel_mul,
'kernel_num': self.kernel_num
}
)
base_config = super(MMDLoss, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def guassian_kernel(self, source, target, kernel_mul=2.0, kernel_num=5, fix_sigma=None):
n_samples = int(source.shape[0]) + int(target.shape[0])
total = tf.concat([source, target], axis=0)
total0 = tf.expand_dims(total, 0)
total0 = tf.tile(total0, [total.shape[0], 1, 1])
total1 = tf.expand_dims(total, 1)
total1 = tf.tile(total1, [1, total.shape[0], 1])
L2_distance = tf.reduce_sum((total0 - total1) ** 2, axis=2)
if fix_sigma:
bandwidth = fix_sigma
else:
bandwidth = tf.reduce_sum(L2_distance) / (n_samples ** 2 - n_samples)
bandwidth /= kernel_mul ** (kernel_num // 2)
bandwidth_list = [bandwidth * (kernel_mul ** i)
for i in range(kernel_num)]
kernel_val = [tf.exp(-L2_distance / bandwidth_temp)
for bandwidth_temp in bandwidth_list]
return sum(kernel_val)
def linear_mmd(self, X, Y):
delta = tf.reduce_mean(X, axis=0) - tf.reduce_mean(Y, axis=0)
loss = tf.linalg.matmul(delta, delta, transpose_b=True)
return loss
def call(self, source, target):
if self.kernel_type == 'linear':
return self.linear_mmd(source, target)
elif self.kernel_type == 'rbf':
batch_size = int(source.shape[0])
kernels = self.guassian_kernel(
source, target, kernel_mul=self.kernel_mul, kernel_num=self.kernel_num, fix_sigma=None)
with tf.GradientTape(persistent=True) as tape:
tape.watch(kernels)
XX = tf.reduce_mean(kernels[:batch_size, :batch_size])
YY = tf.reduce_mean(kernels[batch_size:, batch_size:])
XY = tf.reduce_mean(kernels[:batch_size, batch_size:])
YX = tf.reduce_mean(kernels[batch_size:, :batch_size])
loss = XX + YY - XY - YX
return loss
if __name__ == '__main__':
# 示例用法
source = np.random.randn(100, 128)
target = np.random.randn(100, 128)
source_tf = tf.convert_to_tensor(source, dtype=tf.float32)
target_tf = tf.convert_to_tensor(target, dtype=tf.float32)
mmd_loss = MMDLoss(kernel_type='rbf', kernel_mul=2.0, kernel_num=5)
loss = mmd_loss(source_tf, target_tf)
print("MMD Loss:", loss.numpy())