Keras EarlyStopping:使用哪种min_delta和耐心？

Question

我是深度学习和Keras的新手,我尝试对我的模型训练过程进行的一项改进是利用Keras的keras.callbacks.EarlyStopping回调函数.

基于训练我的模型的输出,使用以下参数似乎是合理的EarlyStopping吗？

EarlyStopping(monitor='val_loss', min_delta=0.0001, patience=5, verbose=0, mode='auto')

另外,为什么它似乎比它应该等待连续5个历元的时间更早停止,其中差异val_loss小于min_delta0.0001？

训练LSTM模型时输出(不含EarlyStop)

运行所有100个时代

Epoch 1/100
10200/10200 [==============================] - 133s 12ms/step - loss: 1.1236 - val_loss: 0.6431
Epoch 2/100
10200/10200 [==============================] - 141s 13ms/step - loss: 0.2783 - val_loss: 0.0301
Epoch 3/100
10200/10200 [==============================] - 143s 13ms/step - loss: 0.1131 - val_loss: 0.1716
Epoch 4/100
10200/10200 [==============================] - 145s 13ms/step - loss: 0.0586 - val_loss: 0.3671
Epoch 5/100
10200/10200 [==============================] - 146s 13ms/step - loss: 0.0785 - val_loss: 0.0038
Epoch 6/100
10200/10200 [==============================] - 146s 13ms/step - loss: 0.0549 - val_loss: 0.0041
Epoch 7/100
10200/10200 [==============================] - 147s 13ms/step - loss: 4.7482e-04 - val_loss: 8.9437e-05
Epoch 8/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.5181e-05 - val_loss: 4.7367e-06
Epoch 9/100
10200/10200 [==============================] - 149s 14ms/step - loss: 9.1632e-07 - val_loss: 3.6576e-07
Epoch 10/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.4117e-07 - val_loss: 1.6058e-07
Epoch 11/100
10200/10200 [==============================] - 152s 14ms/step - loss: 1.2024e-07 - val_loss: 1.2804e-07
Epoch 12/100
10200/10200 [==============================] - 150s 14ms/step - loss: 0.0151 - val_loss: 0.4181
Epoch 13/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0701 - val_loss: 0.0057
Epoch 14/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0332 - val_loss: 5.0014e-04
Epoch 15/100
10200/10200 [==============================] - 147s 14ms/step - loss: 0.0367 - val_loss: 0.0020
Epoch 16/100
10200/10200 [==============================] - 151s 14ms/step - loss: 0.0040 - val_loss: 0.0739
Epoch 17/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0282 - val_loss: 6.4996e-05
Epoch 18/100
10200/10200 [==============================] - 147s 13ms/step - loss: 0.0346 - val_loss: 1.6545e-04
Epoch 19/100
10200/10200 [==============================] - 147s 14ms/step - loss: 4.6678e-05 - val_loss: 6.8101e-06
Epoch 20/100
10200/10200 [==============================] - 148s 14ms/step - loss: 1.7270e-06 - val_loss: 6.7108e-07
Epoch 21/100
10200/10200 [==============================] - 147s 14ms/step - loss: 2.4334e-07 - val_loss: 1.5736e-07
Epoch 22/100
10200/10200 [==============================] - 147s 14ms/step - loss: 0.0416 - val_loss: 0.0547
Epoch 23/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0413 - val_loss: 0.0145
Epoch 24/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0045 - val_loss: 1.1096e-04
Epoch 25/100
10200/10200 [==============================] - 149s 14ms/step - loss: 0.0218 - val_loss: 0.0083
Epoch 26/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0029 - val_loss: 5.0954e-05
Epoch 27/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0316 - val_loss: 0.0035
Epoch 28/100
10200/10200 [==============================] - 148s 14ms/step - loss: 0.0032 - val_loss: 0.2343
Epoch 29/100
10200/10200 [==============================] - 149s 14ms/step - loss: 0.0299 - val_loss: 0.0021
Epoch 30/100
10200/10200 [==============================] - 150s 14ms/step - loss: 0.0171 - val_loss: 9.3622e-04
Epoch 31/100
10200/10200 [==============================] - 149s 14ms/step - loss: 0.0167 - val_loss: 0.0023
Epoch 32/100
10200/10200 [==============================] - 148s 14ms/step - loss: 7.3654e-04 - val_loss: 4.1998e-05
Epoch 33/100
10200/10200 [==============================] - 149s 14ms/step - loss: 7.3300e-06 - val_loss: 1.9043e-06
Epoch 34/100
10200/10200 [==============================] - 148s 14ms/step - loss: 6.6648e-07 - val_loss: 2.3814e-07
Epoch 35/100
10200/10200 [==============================] - 147s 14ms/step - loss: 1.5611e-07 - val_loss: 1.3155e-07
Epoch 36/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.2159e-07 - val_loss: 1.2398e-07
Epoch 37/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1940e-07 - val_loss: 1.1977e-07
Epoch 38/100
10200/10200 [==============================] - 150s 14ms/step - loss: 1.1939e-07 - val_loss: 1.1935e-07
Epoch 39/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1935e-07
Epoch 40/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1935e-07
Epoch 41/100
10200/10200 [==============================] - 150s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 42/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 43/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 44/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 45/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 46/100
10200/10200 [==============================] - 151s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 47/100
10200/10200 [==============================] - 151s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07
Epoch 48/100
10200/10200 [==============================] - 151s 14ms/step - loss: 1.1921e-07 - val_loss: 1.1921e-07

使用EarlyStop输出

11个节后停止(太早了？)

10200/10200 [==============================] - 134s 12ms/step - loss: 1.2733 - val_loss: 0.9022
Epoch 2/100
10200/10200 [==============================] - 144s 13ms/step - loss: 0.5429 - val_loss: 0.4093
Epoch 3/100
10200/10200 [==============================] - 144s 13ms/step - loss: 0.1644 - val_loss: 0.0552
Epoch 4/100
10200/10200 [==============================] - 144s 13ms/step - loss: 0.0263 - val_loss: 0.9872
Epoch 5/100
10200/10200 [==============================] - 145s 13ms/step - loss: 0.1297 - val_loss: 0.1175
Epoch 6/100
10200/10200 [==============================] - 146s 13ms/step - loss: 0.0287 - val_loss: 0.0136
Epoch 7/100
10200/10200 [==============================] - 145s 13ms/step - loss: 0.0718 - val_loss: 0.0270
Epoch 8/100
10200/10200 [==============================] - 145s 13ms/step - loss: 0.0272 - val_loss: 0.0530
Epoch 9/100
10200/10200 [==============================] - 150s 14ms/step - loss: 3.3879e-04 - val_loss: 0.0575
Epoch 10/100
10200/10200 [==============================] - 146s 13ms/step - loss: 1.6789e-05 - val_loss: 0.0766
Epoch 11/100
10200/10200 [==============================] - 149s 14ms/step - loss: 1.4124e-06 - val_loss: 0.0981

Training stops early here.

---

 EarlyStopping(monitor='val_loss', min_delta=0, patience=5, verbose=0, mode='min')

尝试设置min_delta为0.为什么即使val_loss从0.0011增加到0.1045也会停止？

10200/10200 [==============================] - 140s 13ms/step - loss: 1.1938 - val_loss: 0.5941
Epoch 2/100
10200/10200 [==============================] - 150s 14ms/step - loss: 0.3307 - val_loss: 0.0989
Epoch 3/100
10200/10200 [==============================] - 151s 14ms/step - loss: 0.0946 - val_loss: 0.0213
Epoch 4/100
10200/10200 [==============================] - 149s 14ms/step - loss: 0.0521 - val_loss: 0.0011
Epoch 5/100
10200/10200 [==============================] - 150s 14ms/step - loss: 0.0793 - val_loss: 0.0313
Epoch 6/100
10200/10200 [==============================] - 154s 14ms/step - loss: 0.0367 - val_loss: 0.0369
Epoch 7/100
10200/10200 [==============================] - 154s 14ms/step - loss: 0.0323 - val_loss: 0.0014
Epoch 8/100
10200/10200 [==============================] - 153s 14ms/step - loss: 0.0408 - val_loss: 0.0011
Epoch 9/100
10200/10200 [==============================] - 154s 14ms/step - loss: 0.0379 - val_loss: 0.1045

Training stops early here.

Answer 1

从keras 文档中可以清楚地看到两个参数的作用.

min_delta:监控数量的最小变化,即有资格作为改进,即绝对变化小于min_delta,将被视为无改进.

耐心:没有改善的时期数,之后将停止训练.

实际上这些参数没有标准值.您需要分析培训过程的参与者(数据集,环境,模型类型)以确定其值.

(1).忍耐

• 数据集 - 如果数据集对于不同的类别没有那么好的变化(例如 - 年龄组25-30和30-35的人脸).损失的变化将是缓慢的并且也是随机的.- 在这种情况下,拥有更高的价值是件好事patience.反之亦然,这是一个良好而清晰的数据集.
• 模型类型 - 在训练GAN模型时,精度变化会很低(最大情况),并且时期运行会消耗大量的GPU.在这种情况下,最好checkpoint files在具有较低值的特定数量的历元之后保存patience.然后根据需要使用检查点进一步改进.对其他模型类型进行类似的分析.
• 运行时环境 - 在CPU上进行培训时,时间运行会非常耗时.所以,我们更喜欢较小的价值patience.并且可以尝试使用GPU更大的价值.

(2).min_delta

• 要决定min_delta,请运行几个时期并查看错误和验证准确性的变化.根据变化率,应该定义.0在许多情况下,默认值非常有效.

Answer 2

你的参数是有效的第一选择。

然而，正如 Akash 所指出的，这取决于数据集以及您如何拆分数据，例如您的交叉验证方案。您可能希望首先观察模型验证错误的行为，然后相应地选择这些参数。

关于min_delta：我发现 0 或像你这样选择<< 1 很多时候效果很好。再次，首先看看你的错误变化有多疯狂。

关于耐心：如果您将其设置为 n，那么您将在最佳模型之后获得 n 个时期的模型。常见的选择介于 0 和 10 之间，但同样，这将取决于您的数据集，尤其是数据集中的可变性。

最后， EarlyStopping 在您提供的示例中表现正常。在 epoch 4 中找到了最终触发提前停止的最优值：val_loss: 0.0011。之后，训练发现另外 5 个验证损失，它们都高于或等于该最优值，并最终在 5 个 epoch 后终止。