tf.keras給出nan丟失和非驗證錯誤

Question

我正在嘗試使用tf.keras API編程深度神經網絡，我覺得我的模型是正確的，我刪除了所有的nan值，但是在訓練過程中仍然得到NAN值。 我一直在使用的數據集是由UCI WISCON癌症數據集在這里

我的代碼：

from tensorflow import keras
import pandas as pd
import tensorflow as tf

df = pd.read_csv('breastc.csv.csv')
df.dropna()
id_ = df['ID'].tolist()
del df['ID']
labels = df['Class'].tolist()
import numpy as np
del df['Class']
column_list='Compactness'
df[column_list] = df[column_list].apply(pd.to_numeric, errors='coerce')

model = keras.Sequential()
model.add(keras.layers.Dense(64,activation='relu',input_shape = (9,)))
model.add(keras.layers.Dense(64,activation='relu'))
model.add(keras.layers.Dense(1,activation='softmax'))

model.summary()

X=df.iloc[:].values

model.compile(optimizer=tf.train.AdamOptimizer(0.01),
              loss='mse',       # mean squared error
              metrics=['mae'])
model.fit(X,labels,batch_size=32,epochs=10,validation_split=0.2)

在fit語句之后，我得到以下結果

 Train on 559 samples, validate on 140 samples Epoch 1/10 559/559 [==============================] - 0s 599us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 2/10 559/559 [==============================] - 0s 82us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 3/10 559/559 [==============================] - 0s 86us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 4/10 559/559 [==============================] - 0s 84us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 5/10 559/559 [==============================] - 0s 87us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 6/10 559/559 [==============================] - 0s 83us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 7/10 559/559 [==============================] - 0s 80us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 8/10 559/559 [==============================] - 0s 77us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 9/10 559/559 [==============================] - 0s 73us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan Epoch 10/10 559/559 [==============================] - 0s 62us/step - loss: nan - mean_absolute_error: nan - val_loss: nan - val_mean_absolute_error: nan <tensorflow.python.keras._impl.keras.callbacks.History at 0x15c94a80cc0> 

如我們所見，沒有培訓在進行。 請指導我。

您誠摯的，Vidit Shah

Answer 1

當您遇到2個或更多類的分類問題並且只想選擇其中一個時，通常應將最后一層的輸出神經元數與類數相等，並將其激活函數設置為softmax（因此，將在所有類別上作為輸出分布）。 一旦獲得輸出分布，就可以將接收概率最高的輸出向量（即類）中的索引視為預測。

您應該更改的另一件事是損失函數，當您使用softmax作為輸出激活函數時，您需要使用交叉熵損失來測量兩個分布之間的距離（在這種情況下，網絡的輸出和黃金分布-全為0，對應正確類別的索引處為1）。 如果您將黃金標簽表示為正確類別的索引（在您的情況下為0或1），則可以通過將sparse_corssentropy設置為損失函數來直接使用這些索引（它將自動將索引轉換為一熱向量）。

要包裝所有內容，可以按以下方式轉換代碼：

from tensorflow import keras
import pandas as pd
import tensorflow as tf

df = pd.read_csv('breastc.csv.csv')
df.dropna()
id_ = df['ID'].tolist()
del df['ID']
labels = df['Class'].tolist()
import numpy as np
del df['Class']
column_list='Compactness'
df[column_list] = df[column_list].apply(pd.to_numeric, errors='coerce')

model = keras.Sequential()
model.add(keras.layers.Dense(64,activation='relu',input_shape = (9,)))
model.add(keras.layers.Dense(64,activation='relu'))
model.add(keras.layers.Dense(2,activation='softmax'))

model.summary()

X=df.iloc[:].values

model.compile(optimizer=tf.train.AdamOptimizer(0.01),
              loss='categorical_crossentropy',
              metrics=['accuracy'])
model.fit(X,labels,batch_size=32,epochs=10,validation_split=0.2)