问题在于如何使用大约70万个样本的数据集学习RNN。通过tf.train.Saver（）保存模型，然后在其他程序执行中恢复模型和会话，从而有可能基于文件中保存的模型对新样本进行预测。
def prepareNeuralNet(self):
self.X = tf.placeholder(tf.float32, [None, self.n_steps, self.n_inputs], name="X")
self.y = tf.placeholder(tf.float32, [None, self.n_steps, self.n_outputs], name="y")
layers = [
tf.contrib.rnn.LSTMCell(
num_units=self.n_neurons,
activation=tf.nn.tanh,
use_peepholes=True,
initializer=tf.contrib.layers.xavier_initializer())
for layer in range(2)]
cell = tf.contrib.rnn.OutputProjectionWrapper(
tf.contrib.rnn.MultiRNNCell(layers),
output_size=self.n_outputs)
outputs, states = tf.nn.dynamic_rnn(cell, self.X, dtype=tf.float32)
learning_rate = 0.001
self.loss = tf.reduce_mean(tf.square(outputs[:][-1][:] - self.y[:][-1][:]))
tf.add_to_collection('outpouts', outputs)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
self.training_op = optimizer.minimize(self.loss)
self.init = tf.global_variables_initializer()
self.saver = tf.train.Saver()
self.mse_summary = tf.summary.scalar('MSE', self.loss)
def learnNeuralNet(self):
self.prepareNeuralNet()
with tf.Session() as sess:
totalRes = []
totalSal = []
self.init.run()
for iteration, index in enumerate(self.shuffledIndexes):
X_batch, y_batch, batchScaler = self.next_batch()
if iteration % 1000:
mse = self.loss.eval(feed_dict={self.X: X_batch, self.y: y_batch})
summary_str = self.mse_summary.eval(feed_dict={self.X: X_batch, self.y: y_batch})
self.fileWriter.add_summary(summary_str, iteration * (self.n_steps + self.batch_size))
self.saver.save(sess, "model_files/model.ckpt")
self.fileWriter.close()
def makePrediction(self, sess):
try:
d = DataManager.data.iloc[-self.n_steps:].copy().values
X_batch = np.array(d).reshape(1, self.n_steps, self.n_inputs)
y_batch = X_batch[0][:, self.price_close_idx].reshape(1, self.n_steps, self.n_outputs)
X = tf.get_default_graph().get_tensor_by_name("X:0")
y = tf.get_default_graph().get_tensor_by_name("y:0")
outputs = tf.get_collection('outputs')[0]
result = sess.run(outputs, feed_dict={X: X_batch, y: y_batch})
except Exception as ex:
print(ex.__repr__())
if __name__ == '__main__':
saver = tf.train.import_meta_graph('model_files/model.ckpt.meta')
with tf.Session() as sess:
saver.restore(sess, "model_files/model.ckpt")
p = Predictor()
p.makePrediction(sess)
我以以下方式执行此操作，我正在保存X和y张量，以便以后为这些占位符提供feed_dict数据，但是我想从makePrediction（sess）获得预测值。所以我想以某种方式查看dynamic_rnn的“输出”。如何实现这一点，我尝试使用add_collection，然后在还原get_collection时使用。恢复LSTM网络以便将其用于预测是否正确？我想指出的是，在以前的程序执行中。通过执行p.learnNeuralNetwork（）学习了神经网络。