Question

我一直在尝试通过不仅包含过去的功能，还包括目标输出功能旁边的功能来使多功能RNN正常工作。

我遇到的问题是不知道如何构造输入和输出，以使输入也包含与输出相同实例的信息。

到目前为止，我有这个：

mp = pd.read_csv('monthly-milk-production.csv')
mp.index = pd.to_datetime(mp.Month)
mp.drop('Month', axis = 1, inplace = True)

#Making an irrelevant 2nd feature
import random as rand
ran_arr = []
for i in range(len(mp)):
    ran_arr.append(rand.randint(0,9))
mp['Noise'] = ran_arr

#Train test split
mp_train = mp.head(len(mp)-12)

mp_train DataFrame看起来像这样：

   Milk Production     Noise
0         0.086538  0.333333
1         0.019231  0.000000
2         0.209135  0.444444
3         0.247596  0.777778
4         0.418269  0.444444
...

_

mp_test = mp.tail(12)

#Normalizing
from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler()
scaler.fit_transform(mp_train)
mp_train[['Milk Production', 'Noise']] = scaler.transform(mp_train)
mp_test[['Milk Production', 'Noise']] = scaler.transform(mp_test)

#Next_batch Function
def next_batch(training_data,batch_size,steps):
    rand_start = np.random.randint(0, len(training_data)-steps)
    x = np.array(training_data[rand_start: (rand_start + steps + 1)][['Milk Production', 'Noise']]).reshap       e(steps + 1, 2)
    y = np.array(training_data[rand_start: (rand_start + steps + 1)]['Milk Production']).reshape(1, steps + 1)
    lower = x[:-1,:].reshape(-1, steps, 2)
#Looking for lower[0][-1][0] so i stuck in a None in its place
    lower[0][-1][0] = None
    higher = y[:, :-1].reshape(-1, steps, 1)
    return lower, higher

_

a, b = next_batch(mp_train, 1, 12)
print(a)
print(b)
#Output
[[[0.25240385 0.33333333]
  [0.16586538 0.22222222]
  [0.375      0.44444444]
  [0.40625    0.66666667]
  [0.55048077 0.11111111]
  [0.48798077 0.44444444]
  [0.35817308 0.        ]
  [0.24038462 0.88888889]
  [0.14903846 0.77777778]
  [0.16346154 1.        ]
  [0.11778846 0.22222222]
  [       nan 0.33333333]]]
[[[0.25240385]
  [0.16586538]
  [0.375     ]
  [0.40625   ]
  [0.55048077]
  [0.48798077]
  [0.35817308]
  [0.24038462]
  [0.14903846]
  [0.16346154]
  [0.11778846]
  [0.19711538]]]

然后我打算将其传递给Tensorflow，如下所示：

import tensorflow as tf
num_inputs = 2
num_neurons = 100
num_outputs = 1
learning_rate = 0.03
num_itr = 1000
batch_size = 1
time_steps = 12
X = tf.placeholder(tf.float32, [None, time_steps, num_inputs])
y = tf.placeholder(tf.float32, [None, time_steps, num_outputs])
cell = tf.contrib.rnn.GRUCell(num_neurons, activation = tf.nn.relu)
cell = tf.contrib.rnn.OutputProjectionWrapper(cell, num_outputs)
outputs, states = tf.nn.dynamic_rnn(cell, X, dtype = tf.float32)
loss = tf.reduce_mean(tf.square(outputs - y))
optimizer = tf.train.AdamOptimizer(learning_rate)
train = optimizer.minimize(loss)
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session(config = tf.ConfigProto(gpu_options = gpu_option)) as sess:
    sess.run(init)
    for itr in range(num_itr):
        X_batch, y_batch = next_batch(mp_train, batch_size, time_steps)
        sess.run(train, feed_dict = {X: X_batch, y: y_batch})
        if itr % 100 == 0:
            mse = loss.eval(feed_dict = {X: X_batch, y: y_batch})
            print(str(itr) + '\t' + str(mse))
    # Save Model for Later 
    saver.save(sess, "./SavedModel2_test")

如何构造多特征RNN的x /输入和y /输出

0 个答案: