我无法在 keras 中计算“损失”[关闭]

Question

这是我的代码：

# Library Imports
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
from sklearn.preprocessing import MinMaxScaler
plt.style.use("ggplot")

from keras.models import Sequential
from keras.layers import LSTM, Dense, Dropout

# Loading/Reading in the Data
df = pd.read_csv("BTC-USD.csv")


# Data Preprocessing
### Setting the datetime index as the date, only selecting the 'Close' column, then only the last 1000 closing prices.
df = df.set_index("Date")[['Close']].tail(3000)
df = df.set_index(pd.to_datetime(df.index))

# Normalizing/Scaling the Data
scaler = MinMaxScaler()
df = pd.DataFrame(scaler.fit_transform(df), columns=df.columns, index=df.index)


def visualize_training_results(results):
    """
    Plots the loss and accuracy for the training and testing data
    """
    history = results.history
    plt.figure(figsize=(12,4))
    plt.plot(history['val_loss'])
    plt.plot(history['loss'])
    plt.legend(['val_loss', 'loss'])
    plt.title('Loss')
    plt.xlabel('Epochs')
    plt.ylabel('Loss')
    plt.show()
    
    plt.figure(figsize=(12,4))
    plt.plot(history['val_accuracy'])
    plt.plot(history['accuracy'])
    plt.legend(['val_accuracy', 'accuracy'])
    plt.title('Accuracy')
    plt.xlabel('Epochs')
    plt.ylabel('Accuracy')
    plt.show()
    
    
def split_sequence(seq, n_steps_in, n_steps_out):
    """
    Splits the univariate time sequence
    """
    X, y = [], []
    
    for i in range(len(seq)):
        end = i + n_steps_in
        out_end = end + n_steps_out
        
        if out_end > len(seq):
            break
        
        seq_x, seq_y = seq[i:end], seq[end:out_end]
        
        X.append(seq_x)
        y.append(seq_y)
    
    return np.array(X), np.array(y)


def layer_maker(n_layers, n_nodes, activation, drop=None, d_rate=.5):
    """
    Create a specified number of hidden layers for an RNN
    Optional: Adds regularization option, dropout layer to prevent potential overfitting if necessary
    """
    
    # Creating the specified number of hidden layers with the specified number of nodes
    for x in range(1,n_layers+1):
        model.add(LSTM(n_nodes, activation=activation, return_sequences=True))

        # Adds a Dropout layer after every Nth hidden layer (the 'drop' variable)
        try:
            if x % drop == 0:
                model.add(Dropout(d_rate))
        except:
            pass

# How many periods looking back to train
n_per_in  = 30

# How many periods ahead to predict
n_per_out = 10

# Features (in this case it's 1 because there is only one feature: price)
n_features = 1

# Splitting the data into appropriate sequences
X, y = split_sequence(list(df.Close), n_per_in, n_per_out)

# Reshaping the X variable from 2D to 3D
X = X.reshape((X.shape[0], X.shape[1], n_features))

# Instantiating the model
model = Sequential()

# Activation
activ = "softsign"

# Input layer
model.add(LSTM(30, activation=activ, return_sequences=True, input_shape=(n_per_in, n_features)))

# Hidden layers
layer_maker(n_layers=6, n_nodes=12, activation=activ)

# Final Hidden layer
model.add(LSTM(10, activation=activ))

# Output layer
model.add(Dense(n_per_out))

# Model summary
model.summary()

plt.figure(figsize=(12,5))


#Im compiling the code here before training
model.compile()
#visualize_training_results(res)
res = model.fit(X, y, epochs=800, batch_size=32, validation_split=0.1)
#visualize_training_results(res)

我需要在这里做什么才能开始训练我的模型？

我也试过这段代码编译：

model.compile(
    optimizer='adam',
    loss='mean_squared_error',
    metrics=[
        metrics.MeanSquaredError(),
        metrics.AUC(),
    ]
)

但是当我运行这个时，我得到未定义的指标错误。 即使我在编译函数中定义了它？

我也这样做了

model.compile(loss='categorical_crossentropy', optimizer='adam')

开始训练，但我得到了这个输出

Epoch 6/800
10/10 [==============================] - 1s 53ms/step - loss: nan - val_loss: nan
Epoch 7/800
10/10 [==============================] - 0s 43ms/step - loss: nan - val_loss: nan

当我这样做时：

model.compile(loss='categorical_crossentropy',
              optimizer='sgd',
              metrics=['accuracy'])

我得到这个输出

0/10 [==============================] - 0s 49ms/step - loss: nan - accuracy: 0.1128 - val_loss: nan - val_accuracy: 0.0606
Epoch 121/800
10/10 [==============================] - 0s 48ms/step - loss: nan - accuracy: 0.1086 - val_loss: nan - val_accuracy: 0.0606

它训练并且准确度有效，但我仍然得到 nan 的 loss 和 val_loss

不管我用什么编译函数参数，这个东西都不计算损失

现在我试过了

model.compile(
    optimizer='adam',
    loss='mean_squared_error',
    metrics=["accuracy"],
)

这就是我得到的

10/10 [==============================] - 1s 74ms/step - loss: nan - accuracy: 0.0953 - val_loss: nan - val_accuracy: 0.0606
Epoch 5/800
10/10 [==============================] - 1s 74ms/step - loss: nan - accuracy: 0.1097 - val_loss: nan - val_accuracy: 0.0606
Epoch 6/800
10/10 [==============================] - 1s 72ms/step - loss: nan - accuracy: 0.0929 - val_loss: nan - val_accuracy: 0.0606

Answer 1

在训练模型时可以获得 NaN 的原因有很多。

• 数据
• 损失函数返回大于/小于 +infinity/-infinity

尝试获取 20 行，确保您自己用非常笨拙的眼睛查看了每一行，然后运行代码以查看您是否仍然获得 NAN。如果你不这样做，那么可能有 1-2 个损坏的条目。否则，看看split_sequence 仅使用这 20 列时返回的内容。

然后是无穷大问题，这导致了 NAN。尝试使用bounded 的损失函数。

Here is a list of Keras loss functions

虽然均方根偏差返回一个“与输入数据相比以自然方式表示误差”的值i.e. RMSE 30'000 <=> error in estimating house price by 30'000$，但它仍然计算平方，这肯定会很快上升到无穷大，所以它是如果这是这里的问题，那不是解决方案。您可以使用 MAE 将其保持在较低水平。

更新：

它对我有用。我使用了以下内容，以及美国另一个加密男孩硬币的类似数据集：

# The good old SGD and MAE. 
model.compile(optimizer="sgd", loss="mae") # optional: metrics=['accuracy']

在使用 Adams/Huber 或其他优化器/损失函数时要小心。您应该了解它们的运作方式，至少了解我们使用它们的方式和原因。现在，使用 SGD 并不是最好的，但从那里你有一个健康的基础。

祝你工作顺利！不要过拟合，用小数据集制作小型模型，以便在运行 800 个 epoch 之前快速训练和原型化，结果发现一个参数不可行！ 原型！未来你会感谢你的！ :)