Question

我正在尝试在Keras训练LSTM，但似乎不合时宜。

adam = Adam(learning_rate=0.1, beta_1=0.9, beta_2=0.999, amsgrad=False)
model = Sequential()
model.add(LSTM(124, input_shape=(trainX.shape[1], trainX.shape[2])))
model.add(Dense(62))
model.compile(loss='mse', optimizer=adam)
history = model.fit(
              trainX, trainY, epochs=100, batch_size=431, validation_data=(testX, testY), verbose=2, shuffle=False)

我的数据代表了一系列数据泄露事件，其中前0-7列是公司类型，而0-51列是美国各州的热点。我的目标是根据此数据集预测下一次违规。

      date     size    0    1    2    3    4    5    6    7    0    1  ...   40   41   42   43   44   45   46   47   48   49   50   51
0        1    32000  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  1.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
1        2     3500  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
2        3  1400000  0.0  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
3        4   120000  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
4        5    59000  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

每个时期花费的时间为0s，测试数据的损失基本上保持不变。这是我模型的最后10个时期。

Epoch 90/100
 - 0s - loss: 130445623221660.6562 - val_loss: 5098486354074.1934
Epoch 91/100
 - 0s - loss: 130445622565865.9375 - val_loss: 5098486346506.8730
Epoch 92/100
 - 0s - loss: 130445622191126.0625 - val_loss: 5098485465787.7988
Epoch 93/100
 - 0s - loss: 130445621816386.2344 - val_loss: 5098484149524.7568
Epoch 94/100
 - 0s - loss: 130445621347961.4219 - val_loss: 5098484141613.4668
Epoch 95/100
 - 0s - loss: 130445620785851.6406 - val_loss: 5098483698846.1465
Epoch 96/100
 - 0s - loss: 130445620317426.8281 - val_loss: 5098483255046.9189
Epoch 97/100
 - 0s - loss: 130445619942686.9844 - val_loss: 5098482810903.7217
Epoch 98/100
 - 0s - loss: 130445619380577.2188 - val_loss: 5098481494640.6797
Epoch 99/100
 - 0s - loss: 130445619005837.3750 - val_loss: 5098481050841.4521
Epoch 100/100
 - 0s - loss: 130445618724782.5000 - val_loss: 5098480170466.3477

此图显示了训练和测试过程中的损失。 标准化后 数据：

      date          size    0    1    2    3    4    5    6    7    0    1  ...   40   41   42   43   44   45   46   47   48   49   50   51
0        1  1.066633e-05  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  1.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
1        2  1.166333e-06  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
2        3  4.666663e-04  0.0  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
3        4  3.999967e-05  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
4        5  1.966633e-05  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  ...  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

输出：

Epoch 90/100
 - 0s - loss: 1425.8638 - val_loss: 8939.7717
Epoch 91/100
 - 0s - loss: 1397.5078 - val_loss: 8847.1606
Epoch 92/100
 - 0s - loss: 1403.1658 - val_loss: 8756.0660
Epoch 93/100
 - 0s - loss: 1453.6541 - val_loss: 9025.9873
Epoch 94/100
 - 0s - loss: 1435.4252 - val_loss: 8928.1981
Epoch 95/100
 - 0s - loss: 1414.3660 - val_loss: 8830.8435
Epoch 96/100
 - 0s - loss: 1391.6843 - val_loss: 8734.7518
Epoch 97/100
 - 0s - loss: 1365.5961 - val_loss: 8655.9235
Epoch 98/100
 - 0s - loss: 1431.5669 - val_loss: 8896.3086
Epoch 99/100
 - 0s - loss: 1409.6963 - val_loss: 9031.4488
Epoch 100/100
 - 0s - loss: 1466.6886 - val_loss: 9088.9677

损失：

如何在Keras训练欠拟合模型？

0 个答案: