无法减少此张量流 lstm 模型的损失

Mehd 发表于 Dev

马赫迪——

我正在尝试使用 tensorflow 的 LSTM 模块来预测序列，但无济于事。我无法弄清楚问题所在，我希望有人能帮我一把。这是我的代码：

首先我主要创建合成数据，并准备数据加载器

x = np.linspace(0,30.,500)
y = x*np.sin(x) + 2*np.sin(5*x)
nb_steps = 20

def load_batch(batch_size = 32): 

    x_b = np.zeros((nb_steps,batch_size,1))
    y_b = np.zeros((nb_steps*batch_size,1))

    inds = np.random.randint(0, 479, (batch_size))
    for i,ind in enumerate(inds): 
        x_b[:,i,0] = x[ind:ind+nb_steps]
        y_b[i*nb_steps:(i+1)*nb_steps,0] = y[ind+1:ind+nb_steps+1]

    return x_b, y_b

一些捷径

adam = tf.train.AdamOptimizer
layers = tf.layers
dense = layers.dense
lstm = tf.contrib.rnn.LSTMCell
batch_size = 64

然后，是我创建模型的部分

with tf.variable_scope('data'): 

    x_p = tf.placeholder(tf.float32, shape = [nb_steps, None, 1], name = 'x') # batch, steps, features 
    y_p = tf.placeholder(tf.float32, shape = [None, 1], name = 'labels')

with tf.variable_scope('network'): 

    cell = lstm(num_units = 100)
    outputs, states = tf.nn.dynamic_rnn(cell, x_p, dtype = tf.float32, time_major = True)


    reshaped_outputs = tf.reshape(outputs, [-1,100])
    projection = dense(reshaped_outputs, 1, activation = None, name = 'projection')

以上是我最不确定的部分。我为每个时间步重塑 lstm 的输出并将它们堆叠在第一个轴上（或者我？）。然后我将整个矩阵发送到一个线性层中。

with tf.variable_scope('training'): 

    loss = tf.reduce_mean(tf.square(projection - y_p))
    train_lstm = adam(1e-3).minimize(loss)


epochs = 1000
batch_size = 64
f, ax = plt.subplots(2,1)
with tf.Session() as sess:

    sess.run(tf.global_variables_initializer())

    mean_loss = 0. 
    for epoch in range(1,epochs+1): 


        x_b,y_b = load_batch(batch_size)

        batch_loss,_ = sess.run([loss, train_lstm], feed_dict = {x_p:x_b, y_p:y_b})

        mean_loss += batch_loss

        if epoch%100 == 0: 
            print('Epoch: {} | Loss: {:.6f}'.format(epoch, mean_loss/100.))
            mean_loss = 0. 

    while True : 

        x_b, y_b = load_batch(1)
        pred = sess.run(projection, feed_dict = {x_p:x_b}).reshape(-1)

        ax[0].plot(x,y, label= 'Real')
        ax[0].plot(x_b.reshape(-1),y_b.reshape(-1), label= 'Real batch')
        ax[0].plot(x_b.reshape(-1), pred, label = 'Pred')

        ax[1].scatter(x_b.reshape(-1),y_b.reshape(-1), label= 'Real')
        ax[1].scatter(x_b.reshape(-1), pred, label = 'Pred')

        for a in ax: a.legend()

        plt.pause(0.1)
        input()

        for a in ax: 
            a.clear()

非常感谢！

卡斯帕

每个 LSTM 单元产生 100 个输出，因此在执行 tf.nn.dynamic_rnn 后，您需要展平该输出。我宁愿用

reshaped_outputs = tf.layers.Flatten()(outputs)

在这一行之后：

 outputs, states = tf.nn.dynamic_rnn(cell, x_p, dtype = tf.float32, time_major = True)

而不是这一行：

reshaped_outputs = tf.reshape(outputs, [-1,100])

希望能帮助到你：）

编辑：我没有注意到您使用了 time_major = True。我对您的代码进行了一些更改，time_major = False，因为它更易于使用。

我假设您想预测 nb_steps 输出。

代码：

import tensorflow as tf
import numpy as np

import matplotlib.pyplot as plt

x = np.linspace(0,30.,500)
y = x*np.sin(x) + 2*np.sin(5*x)
nb_steps = 20

def load_batch(batch_size = 32):

    x_b = np.zeros((batch_size, nb_steps))
    y_b = np.zeros((batch_size, nb_steps))

    inds = np.random.randint(0, 479, (batch_size))
    for i,ind in enumerate(inds):
        x_b[i] = x[ind:ind+nb_steps]
        y_b[i] = y[ind+1:ind+nb_steps+1]

    return x_b, y_b

adam = tf.train.AdamOptimizer
layers = tf.layers
dense = layers.dense
lstm = tf.contrib.rnn.LSTMCell
batch_size = 64

with tf.variable_scope('data'):
    x_p = tf.placeholder(tf.float32, shape = [None, nb_steps], name = 'x')
    x_rnn = tf.expand_dims(x_p, 2) # batch, steps, features
    y_p = tf.placeholder(tf.float32, shape = [None, nb_steps], name = 'labels')

with tf.variable_scope('network'):
    cell = lstm(num_units = 100)
    outputs, states = tf.nn.dynamic_rnn(cell, x_rnn, dtype = tf.float32, time_major = False)
    reshaped_outputs = tf.layers.Flatten()(outputs)
    projection = dense(reshaped_outputs, nb_steps, activation = None, name = 'projection')

with tf.variable_scope('training'):
    loss = tf.reduce_mean(tf.square(projection - y_p))
    train_lstm = adam(1e-3).minimize(loss)


epochs = 1000
batch_size = 64
f, ax = plt.subplots(2,1)
with tf.Session() as sess:

    sess.run(tf.global_variables_initializer())

    mean_loss = 0.
    for epoch in range(1,epochs+1):

        x_b,y_b = load_batch(batch_size)

        batch_loss,_ = sess.run([loss, train_lstm], feed_dict = {x_p:x_b, y_p:y_b})

        mean_loss += batch_loss

        if epoch%100 == 0:
            print('Epoch: {} | Loss: {:.6f}'.format(epoch, mean_loss/100.))
            mean_loss = 0.

本文收集自互联网，转载请注明来源。

如有侵权，请联系 [email protected] 删除。