在本实验中构造了一系列正弦曲线的变形：
X=np.sin(T, dtype=np.float32)0.01np.random.rand()(T+np.random.rand()10)+np.random.rand()*10
样子的话，大概是：

image.png

1. Raged tensor的使用。这里采用了from_tensor方法构造出了一个Raged tensor来存储一系列长度不一致的序列。
2. 利用karas处理变长序列。由于karas中的LSTM是更高级的API，因此可以直接对Raged tensor进行处理。
3. Seq2seq框架。这里采用的是tensorflow_addons中的seq2seq框架，由于此框架本身对文本处理的支持比较多，在应对我们的（相对简单的）实数序列的案例中，反而需要进行一些定制。主要是对sampler的定制。
4. Earlystop机制。直接采用karas的earlystop即可。

预测效果：

image.png

完整代码

import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

T=np.arange(2000)*0.05
X=np.sin(T, dtype=np.float32)*0.01*np.random.rand()*(T+np.random.rand()*10)+np.random.rand()*10
plt.plot(X)

batch_size = 32
max_time = 200
hidden_size = 128

YLEN=max_time

def generateTrain(records=32):
    trainXs=[]
    trainYs=[]
    XLens=[]
    YLens=[]
    for i in range(records):
        X=(np.sin(T)*0.01*np.random.rand()*(T+np.random.rand()*10)+np.random.rand()*10).astype(np.float32)
        xi=X[:1800].copy()
        sepre=np.random.randint(400,1600)
        xi[sepre::]=0
        yi=X[sepre:sepre+YLEN].copy()
        trainXs.append(np.expand_dims(xi,axis=0))
        trainYs.append(np.expand_dims(yi,axis=0))
        XLens.append(sepre)
        YLens.append(YLEN)
    trainX=np.concatenate(trainXs,axis=0)
    trainY=np.concatenate(trainYs,axis=0)
    seqLen=np.array(XLens)
    outLen=np.array(YLens)
    return (trainX,trainY,seqLen,outLen)

def generateTest():
    X=(np.sin(T)*0.01*np.random.rand()*(T+np.random.rand()*10)+np.random.rand()*10).astype(np.float32)
    xi=X[:1800].copy()
    yi=X[1800:1800+YLEN].copy()
    testX=np.expand_dims(xi, axis=0)
    testY=np.expand_dims(yi, axis=0)
    seqLen_test=np.array([1800])
    outLen_test=np.array([YLEN])
    return (testX,testY,seqLen_test,outLen_test)

def gen_raged_train():
    while True:
        trainX,trainY,seqLen,_ = generateTrain()
        raged_train_x = tf.RaggedTensor.from_tensor(trainX, lengths=seqLen)
        raged_train_x = tf.expand_dims(raged_train_x, -1)
        tensor_train_y = tf.convert_to_tensor(trainY)
        yield raged_train_x,tensor_train_y
    
dataset = tf.data.Dataset.from_generator(
     gen_raged_train,
     output_signature=(
         tf.RaggedTensorSpec(shape=(batch_size, None, 1), dtype=tf.float32, ragged_rank=1),
         tf.TensorSpec(shape=(batch_size, 200,), dtype=tf.float32))
)

import tensorflow_addons as tfa
import tensorflow as tf

inputs = tf.keras.layers.Input(shape=[None, 1], ragged=True)

encoding, state_h, state_c = tf.keras.layers.LSTM(hidden_size, return_state=True)(inputs)

encoder_state = [state_h, state_c]

decoder_cell = tf.keras.layers.LSTMCell(hidden_size)

sample_fn = lambda x: x

end_fn = lambda x:False

sampler = tfa.seq2seq.InferenceSampler(sample_fn = sample_fn, sample_shape=[hidden_size],sample_dtype=tf.float32,end_fn=end_fn)

decoder = tfa.seq2seq.BasicDecoder(decoder_cell, sampler, maximum_iterations=200)

input_lengths = tf.fill([batch_size], max_time)

initial_state = decoder_cell.get_initial_state(encoding)

output, state, lengths = decoder(
    tf.convert_to_tensor(encoding), initial_state=initial_state)

logits = output.rnn_output

output_layer = tf.keras.layers.Dense(1)

out_seq = tf.squeeze(output_layer(logits))

print(out_seq.shape)

model = tf.keras.Model(inputs=inputs, outputs=out_seq)

model.compile(optimizer="Adam", loss="mse", metrics=["mse", "mae", "mape"])

early_stop = keras.callbacks.EarlyStopping(monitor='val_loss', patience=100)

model.fit(dataset.take(128),validation_data=val_data, epochs=1024,callbacks=[early_stop])

model.save('saved_model/my_model')

plt.figure()
plt.plot(X)
plt.plot([np.nan]*2000+model.predict(np.expand_dims(X,0)).tolist())