输入包含 NaN、无穷大或对于 Tensorflow 中的 dtype('float64') 而言太大的值

Question

我正在尝试训练 LSTM，在我的模型中，我有一个指数学习率衰减和一个 dropout 层。为了在测试和验证时停用 dropout 层，我为 dropout 率放置了一个占位符，并给它一个默认值 1.0，在训练时我将它设置为 0.5。 dropou_rate 占位符值被传递给 tf.layers.dropout()。当我在验证期间运行此程序时，我收到以下错误。

ValueError: 输入包含 NaN、无穷大或一个太大的值 dtype('float64').

下面显示的是堆栈跟踪：

Traceback（最近一次调用最后一次）：文件 "/home/suleka/Documents/sales_prediction/SalesPrediction_LSTM_mv.py", 第 329 行，在 train_test() 文件“/home/suleka/Documents/sales_prediction/SalesPrediction_LSTM_mv.py”， 第 270 行，在 train_test 中 meanSquaredError = mean_squared_error(nonescaled_y, pred_vals) 文件 “/home/suleka/anaconda3/lib/python3.6/site-packages/sklearn/metrics/regression.py”， 第 238 行，在 mean_squared_error 中 y_true，y_pred，多输出）文件“/home/suleka/anaconda3/lib/python3.6/site-packages/sklearn/metrics/regression.py”， 第 77 行，在 _check_reg_targets y_pred = check_array(y_pred, ensure_2d=False) 文件 "/home/suleka/anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py", 第 453 行，在 check_array 中 _assert_all_finite（数组）文件“/home/suleka/anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py”， 第 44 行，在 _assert_all_finite " 或对于 %r 来说太大的值。" % X.dtype) ValueError: Input contains NaN, infinity or a value too large for dtype('float64').

当我将学习率作为 tf.layers.dropout 中的一个值时：

dropout = tf.layers.dropout(last, rate=0.5, training=True)

代码运行良好。我不确定代码中发生了什么。

下面是我的完整代码：

import tensorflow as tf
import matplotlib as mplt
mplt.use('agg')  # Must be before importing matplotlib.pyplot or pylab!
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from sklearn.metrics import mean_squared_error
from sklearn.metrics import mean_absolute_error
from math import sqrt
import csv
np.random.seed(1)
tf.set_random_seed(1)


class RNNConfig():
    input_size = 1
    num_steps = 7#5
    lstm_size = 64 #16
    num_layers = 1
    keep_prob = 0.8
    batch_size = 16 #64
    init_epoch = 15  # 5
    max_epoch = 20 # 100 or 50
    # test_ratio = 0.2
    fileName = 'store2_1.csv'
    graph = tf.Graph()
    column_min_max = [[0,11000], [1,7]]
    columns = ['Sales', 'DayOfWeek','SchoolHoliday', 'Promo']
    features = len(columns)
    hidden1_nodes = 64
    hidden2_nodes = 8



config = RNNConfig()

def segmentation(data):

    seq = [price for tup in data[config.columns].values for price in tup]

    seq = np.array(seq)

    # split into items of features
    seq = [np.array(seq[i * config.features: (i + 1) * config.features])
           for i in range(len(seq) // config.features)]

    # split into groups of num_steps
    X = np.array([seq[i: i + config.num_steps] for i in range(len(seq) -  config.num_steps)])

    y = np.array([seq[i +  config.num_steps] for i in range(len(seq) -  config.num_steps)])

    # get only sales value
    y = [[y[i][0]] for i in range(len(y))]

    y = np.asarray(y)

    return X, y

def scale(data):

    for i in range (len(config.column_min_max)):
        data[config.columns[i]] = (data[config.columns[i]] - config.column_min_max[i][0]) / ((config.column_min_max[i][1]) - (config.column_min_max[i][0]))

    return data

def rescle(test_pred):

    prediction = [(pred * (config.column_min_max[0][1] - config.column_min_max[0][0])) + config.column_min_max[0][0] for pred in test_pred]

    return prediction


def pre_process():
    store_data = pd.read_csv(config.fileName)

    store_data = store_data.drop(store_data[(store_data.Open == 0) & (store_data.Sales == 0)].index)
    #
    # store_data = store_data.drop(store_data[(store_data.Open != 0) & (store_data.Sales == 0)].index)

    # ---for segmenting original data --------------------------------
    # original_data = store_data.copy()

    ## train_size = int(len(store_data) * (1.0 - test_ratio))

    validation_len = len(store_data[(store_data.Month == 6) & (store_data.Year == 2015)].index)
    test_len = len(store_data[(store_data.Month == 7) & (store_data.Year == 2015)].index)
    train_size = int(len(store_data) - (validation_len + test_len))

    train_data = store_data[:train_size]
    validation_data = store_data[(train_size - config.num_steps): validation_len + train_size]
    test_data = store_data[((validation_len + train_size) - config.num_steps):]
    original_val_data = validation_data.copy()
    original_test_data = test_data.copy()

    # -------------- processing train data---------------------------------------
    scaled_train_data = scale(train_data)
    train_X, train_y = segmentation(scaled_train_data)

    # -------------- processing validation data---------------------------------------
    scaled_validation_data = scale(validation_data)
    val_X, val_y = segmentation(scaled_validation_data)

    # -------------- processing test data---------------------------------------
    scaled_test_data = scale(test_data)
    test_X, test_y = segmentation(scaled_test_data)

    # ----segmenting original validation data-----------------------------------------------
    nonescaled_val_X, nonescaled_val_y = segmentation(original_val_data)

    # ----segmenting original test data---------------------------------------------
    nonescaled_test_X, nonescaled_test_y = segmentation(original_test_data)

    return train_X, train_y, test_X, test_y, val_X, val_y, nonescaled_test_y, nonescaled_val_y


def generate_batches(train_X, train_y, batch_size):
    num_batches = int(len(train_X)) // batch_size
    if batch_size * num_batches < len(train_X):
        num_batches += 1

    batch_indices = range(num_batches)
    for j in batch_indices:
        batch_X = train_X[j * batch_size: (j + 1) * batch_size]
        batch_y = train_y[j * batch_size: (j + 1) * batch_size]
        assert set(map(len, batch_X)) == {config.num_steps}
        yield batch_X, batch_y

def mean_absolute_percentage_error(y_true, y_pred):
    y_true, y_pred = np.array(y_true), np.array(y_pred)
    itemindex = np.where(y_true == 0)
    y_true = np.delete(y_true, itemindex)
    y_pred = np.delete(y_pred, itemindex)
    return np.mean(np.abs((y_true - y_pred) / y_true)) * 100

def RMSPE(y_true, y_pred):
    y_true, y_pred = np.array(y_true), np.array(y_pred)
    return np.sqrt(np.mean(np.square(((y_true - y_pred) / y_pred)), axis=0))

def plot(true_vals,pred_vals,name):
    fig = plt.figure()
    fig = plt.figure(dpi=100, figsize=(20, 7))
    days = range(len(true_vals))
    plt.plot(days, pred_vals, label='pred sales')
    plt.plot(days, true_vals, label='truth sales')
    plt.legend(loc='upper left', frameon=False)
    plt.xlabel("day")
    plt.ylabel("sales")
    plt.grid(ls='--')
    plt.savefig(name, format='png', bbox_inches='tight', transparent=False)
    plt.close()

def write_results(true_vals,pred_vals,name):

    with open(name, "w") as f:
        writer = csv.writer(f)
        writer.writerows(zip(true_vals, pred_vals))


def train_test():
    train_X, train_y, test_X, test_y, val_X, val_y, nonescaled_test_y, nonescaled_val_y = pre_process()


    # Add nodes to the graph
    with config.graph.as_default():

        tf.set_random_seed(1)

        learning_rate = tf.placeholder(tf.float32, None, name="learning_rate")
        inputs = tf.placeholder(tf.float32, [None, config.num_steps, config.features], name="inputs")
        targets = tf.placeholder(tf.float32, [None, config.input_size], name="targets")
        global_step = tf.Variable(0, trainable=False)
        dropout_rate = tf.placeholder_with_default(1.0, shape=())

        learning_rate = tf.train.exponential_decay(learning_rate=learning_rate, global_step=global_step, decay_rate=0.96,  decay_steps=5, staircase=False)

        cell = tf.contrib.rnn.LSTMCell(config.lstm_size, state_is_tuple=True, activation=tf.nn.relu)

        val1, _ = tf.nn.dynamic_rnn(cell, inputs, dtype=tf.float32)

        val = tf.transpose(val1, [1, 0, 2])

        last = tf.gather(val, int(val.get_shape()[0]) - 1, name="last_lstm_output")


        # hidden layer
        last = tf.layers.dense(last, units=config.hidden1_nodes, activation=tf.nn.relu)
        last = tf.layers.dense(last, units=config.hidden2_nodes, activation=tf.nn.relu)



        weight = tf.Variable(tf.truncated_normal([config.hidden2_nodes, config.input_size]))
        bias = tf.Variable(tf.constant(0.1, shape=[config.input_size]))

        dropout = tf.layers.dropout(last, rate=dropout_rate, training=True)

        prediction = tf.matmul(dropout, weight) + bias

        loss = tf.losses.mean_squared_error(targets,prediction)
        optimizer = tf.train.AdamOptimizer(learning_rate)
        minimize = optimizer.minimize(loss, global_step=global_step)

        # correct_prediction = tf.sqrt(tf.losses.mean_squared_error(prediction, targets))


    # --------------------training------------------------------------------------------

    with tf.Session(graph=config.graph) as sess:
        tf.set_random_seed(1)

        tf.global_variables_initializer().run()

        iteration = 1


        for epoch_step in range(config.max_epoch):


            for batch_X, batch_y in generate_batches(train_X, train_y, config.batch_size):
                train_data_feed = {
                    inputs: batch_X,
                    targets: batch_y,
                    learning_rate: 0.01,
                    dropout_rate: 0.5
                }

                train_loss, _, value,gs = sess.run([loss, minimize, val1,global_step], train_data_feed)

                if iteration % 5 == 0:
                    print("Epoch: {}/{}".format(epoch_step, config.max_epoch),
                          "Iteration: {}".format(iteration),
                          "Train loss: {:.6f}".format(train_loss))
                iteration += 1

        saver = tf.train.Saver()
        saver.save(sess, "checkpoints_sales/sales_pred.ckpt")

        # --------------------validation------------------------------------------------------

        with tf.Session(graph=config.graph) as sess:
            tf.set_random_seed(1)

            saver.restore(sess, tf.train.latest_checkpoint('checkpoints_sales'))

            test_data_feed = {
                inputs: val_X,
                dropout_rate: 1.0
            }

            test_pred = sess.run(prediction, test_data_feed)

            # rmsse = sess.run(correct_prediction, test_data_feed)

            pred_vals = rescle(test_pred)

            pred_vals = np.array(pred_vals)

            pred_vals = pred_vals.flatten()

            pred_vals = pred_vals.tolist()

            nonescaled_y = nonescaled_val_y.flatten()

            nonescaled_y = nonescaled_y.tolist()

            plot(nonescaled_y, pred_vals, "Sales Prediction VS Truth mv testSet.png")
            write_results(nonescaled_y, pred_vals, "Sales Prediction batch mv results_all validationSet.csv")

            meanSquaredError = mean_squared_error(nonescaled_y, pred_vals)
            rootMeanSquaredError = sqrt(meanSquaredError)
            print("RMSE:", rootMeanSquaredError)
            mae = mean_absolute_error(nonescaled_y, pred_vals)
            print("MAE:", mae)
            mape = mean_absolute_percentage_error(nonescaled_y, pred_vals)
            print("MAPE:", mape)
            rmse_val = RMSPE(nonescaled_y, pred_vals)
            print("RMSPE:", rmse_val)

    # --------------------testing------------------------------------------------------

    with tf.Session(graph=config.graph) as sess:
        tf.set_random_seed(1)

        saver.restore(sess, tf.train.latest_checkpoint('checkpoints_sales'))

        test_data_feed = {
            inputs: test_X,
            dropout_rate: 1.0
        }

        test_pred = sess.run(prediction, test_data_feed)

        # rmsse = sess.run(correct_prediction, test_data_feed)


        pred_vals = rescle(test_pred)

        pred_vals = np.array(pred_vals)

        pred_vals = (np.round(pred_vals, 0)).astype(np.int32)

        pred_vals = pred_vals.flatten()


        pred_vals = pred_vals.tolist()

        nonescaled_y = nonescaled_test_y.flatten()

        nonescaled_y = nonescaled_y.tolist()

        plot(nonescaled_y, pred_vals, "Sales Prediction VS Truth mv testSet.png")
        write_results(nonescaled_y, pred_vals, "Sales Prediction batch mv results_all validationSet.csv")

        meanSquaredError = mean_squared_error(nonescaled_y, pred_vals)
        rootMeanSquaredError = sqrt(meanSquaredError)
        print("RMSE:", rootMeanSquaredError)
        mae = mean_absolute_error(nonescaled_y, pred_vals)
        print("MAE:", mae)
        mape = mean_absolute_percentage_error(nonescaled_y, pred_vals)
        print("MAPE:", mape)
        rmse_val = RMSPE(nonescaled_y, pred_vals)
        print("RMSPE:", rmse_val)




if __name__ == '__main__':
    train_test()

Answer 1

当使用 tf.layers.dropout 时，rate 参数告诉你在给 1.0 时要丢弃多少数据，所有输出都消失了，将 1.0 替换为 0.0，它应该可以工作。 TensorFlow 文档：https://www.tensorflow.org/api_docs/python/tf/layers/dropout

Answer 2

我之所以这么说是因为即使@Almog 的回答是正确的，它也没有我想要的解释。所以对于像我这样困惑的人：

如果你使用：

'tf.nn.dropout()'

要停用你应该放置的 dropout 层

keep_prob=1.0 不是 keep_prob=0.0

因为 keep_prob 的意思是“保留每个元素的概率。”因此，将其保持为 1.0 以停用它是有意义的。

如果你正在使用

'tf.layers.dropout()'

你应该输入：

rate=0.0 不是 rate=1.0

这里的 rate 表示“辍学率”（应该在 0 到 1 之间）。例如。 “rate=0.1”会丢弃 10% 的输入单元。因此，如果我输入 rate=0.0 ，则意味着不会删除任何输入单元。