Question

我有1212个样本的数据集，每个值都是股票价格。我创建了RNN模型来预测它们，但是当我生成预测和原始图时，我看到该预测图看起来很奇怪，所以我在这里。而且我不知道如何做出更像原始情节的预测

配置：

n_inputs=1 
n_outputs=1
n_neurons=48,
n_iterations=1212
n_time_steps=12
learning_rate=.03

LSTM：

lstm_cell = tf.contrib.rnn.OutputProjectionWrapper(tf.contrib.rnn.BasicLSTMCell(num_units=self.n_neurons, activation=tf.nn.sigmoid), output_size=self.n_outputs)
outputs, sates = tf.nn.dynamic_rnn(lstm_cell, self.x, dtype=tf.float32)

损失和优化器：

loss = tf.reduce_mean(tf.square(outputs - self.y))
optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(loss)

批处理系统：

def next_batch(self, data):
    rand_start = np.random.randint(len(data) - self.n_time_steps)
    y_batch = np.array(data[rand_start:rand_start + self.n_time_steps + 1]).reshape(1, self.n_time_steps + 1)

    return y_batch[: ,:-1].reshape(-1, self.n_time_steps, 1), y_batch[:, 1:].reshape(-1, self.n_time_steps, 1)

培训：

for i in range(self.n_iterations):    
    x_batch, y_batch = self.next_batch(data)

    sess.run(optimizer, feed_dict={
        self.x:x_batch,
        self.y:y_batch
    })

这是我预测值的方式-恢复保存的模型

graph = tf.get_default_graph()

x = graph.get_tensor_by_name('x_input:0')
outputs_tensor = graph.get_tensor_by_name('outputs:0')
n_time_steps = graph.get_tensor_by_name('n_time_steps:0')

n_time_steps = sess.run(n_time_steps)
data = data.reshape(-1, n_time_steps, 1)

outputs_values = sess.run(outputs_tensor, feed_dict={
    x: data
})

在其他文件中：

selected = list(data[-48:])
n_time_steps = 48
for _ in range(48):
    part = selected[-n_time_steps:]
    val = predictor.predict(part)

selected.append(val[0, -1, 0])

Tensorflow RNN预测股价-怪异的预测图

0 个答案: