不相容的形状：[1020，1，1] 与 [1019，1，1] - 张量流

Question

当我将神经网络的输出设置为1时出现问题，它在数据数组中出现了问题，指责哪个形状更大，如果我使用，一切都可以正常工作！periods = len(valuesAnalisys) - 1时期：periods = 1返回：Imcompatible shapes: [1020,1,1] vs. [1019,1,1]神经网络：datecollect = [x[0] for x in dataSet]servers = [x[1] for x in dataSet]valuesAnalisys = [float(x[2]) for x in dataSet]base = np.array(valuesAnalisys)periods = 1future_forecast = 1X = base[0:(len(base) - (len(base) % periods))]X_batches = X.reshape(-1, periods, 1)y = base[1:(len(base) - (len(base) % periods)) + future_forecast]y_batches = y.reshape(-1, periods, 1)X_test = base[-(periods + future_forecast):]X_test = X_test[:periods]X_test = X_test.reshape(-1, periods, 1)y_test = base[-(periods):]y_test = y_test.reshape(-1, periods, 1)tf.reset_default_graph()appetizer = 1hidden_neurons = 100exit_neurons = 1xph = tf.placeholder(tf.float32, [None, periods, appetizer])yph = tf.placeholder(tf.float32, [None, periods, exit_neurons])cell = tf.contrib.rnn.BasicRNNCell(num_units = hidden_neurons, activation = tf.nn.relu)cell = tf.contrib.rnn.OutputProjectionWrapper(cell, output_size = 1)exit_rnn, _ = tf.nn.dynamic_rnn(cell, xph, dtype = tf.float32)calculateError = tf.losses.mean_squared_error(labels = yph, predictions = exit_rnn)otimizador = tf.train.AdamOptimizer(learning_rate = 0.001)training = otimizador.minimize(calculateError)with tf.Session() as sess:    sess.run(tf.global_variables_initializer())    for epoch in range(2000):        _, cost = sess.run([training, calculateError], feed_dict = {xph: X_batches, yph: y_batches})        if epoch % 100 == 0:            print("[INFO] Epoch: {} - Level Error: {}".format(epoch,cost))    forecast = sess.run(exit_rnn, feed_dict = {xph: X_test})y_test.shapey_test2 = np.ravel(y_test)final_forecast = np.ravel(forecast)mae = mean_absolute_error(y_test2, final_forecast)for (host, forecast, date) in list(zip(servers, final_forecast, datecollect)):    send.postForecastMemory(host, forecast, cost, date)

Answer 1

罪魁祸首似乎是RNN细胞中固定的时间维度。

xph = tf.placeholder(tf.float32, [None, periods, appetizer])

yph = tf.placeholder(tf.float32, [None, periods, exit_neurons])

cell = tf.contrib.rnn.BasicRNNCell(num_units = hidden_neurons, activation = tf.nn.relu)

在这里，在 xph 和 yph 中，您都已将时间维度指定为周期。因此，如果您有更长或更短的信号，则会出现错误。

我无法推断模型层的确切尺寸，因为您没有指定输入形状或模型摘要。因此，使用占位符数字。

有两种可能的修复方法。

不要使用固定的时间维度 = 周期，而应使用 None。

xph = tf.placeholder(tf.float32, [None, None, appetizer])

yph = tf.placeholder(tf.float32, [None, None, exit_neurons])

但是，缺点是每个批次中必须具有相同长度的信号，或者您可以简单地使用批量大小 = 1 进行训练，而不必担心时间长度。

使用截断/填充来解决长度问题。只需将信号传递到预处理函数即可添加/删除额外的时间点。

import numpy as np

def pre_process(x, fixed_len = 1000): # x.shape -> (100, 1000, 1)

    if x.shape[1] >= fixed_len:

       return x[:,:fixed_len,:]

    else:

       z_ph = np.zeros((x.shape[0], fixed_len, x.shape[2]))

       z_ph[:,:x.shape[1],:] = x

       return z_ph

X_batches = pre_process(X_batches, YOU_CHOOSE_THIS_LENGTH) # based on the length of your data

X_test = pre_process(X_test, YOU_CHOOSE_THIS_LENGTH)