使用 keras 训练多类 nn 时损失不能再往下可能是什么原因答案

【问题标题】：What may be the reason for loss can not go further down when training multi-class nn using keras使用 keras 训练多类 nn 时损失不能再往下可能是什么原因
【发布时间】：2017-11-08 13:20:02
【问题描述】：

我正在使用 keras 训练一个多类神经网络（后端是 tensorflow）。我将在最终位置给出我的设置和一些代码。

描述是：当我进行 10 个文件夹交叉验证时，训练损失和验证损失在前 10-15 个时期下降，但在 15 个时期后不能进一步下降并保持大约（损失：1.0606 - acc：0.6301 - val_loss：1.1577 - val_acc: 0.5774)。

我已尝试对我的设置进行多项更改。例如，添加隐藏层，添加normalization.BatchNormalization()，将优化器从adam 更改为sgd 或rmsprop，将损失函数从categorical_crossentropy 更改为其他。但没有效果。

我想讨论一下这种事情的可能原因是什么。如果这里有摘要文件或演示文稿，我将非常高兴。

我的数据有 10000 行。并且特征有507个属性为0/1。标签是多类的，类数为 7。类之间的平衡几乎可以，因为我从更大的数据集中选择了 10000 个数据。

我的模型如下：

model = Sequential()
model.add(Dense(500, activation='relu', input_dim=self.feature_dim, 
kernel_regularizer=regularizers.l2(0.01)))
model.add(Dense(100, activation='relu'))
model.add(Dense(self.label_dim, activation='softmax'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

部分日志如下：

Running Fold 1/10
Train on 9534 samples, validate on 1060 samples
Epoch 1/100
1000/9534 [==>...........................] - ETA: 7s - loss: 6.9644 - acc: 0.1150
2000/9534 [=====>........................] - ETA: 3s - loss: 6.8357 - acc: 0.1715
3000/9534 [========>.....................] - ETA: 2s - loss: 6.7147 - acc: 0.2243
4000/9534 [===========>..................] - ETA: 1s - loss: 6.5922 - acc: 0.2683
5000/9534 [==============>...............] - ETA: 1s - loss: 6.4779 - acc: 0.2908
6000/9534 [=================>............] - ETA: 0s - loss: 6.3618 - acc: 0.3097
7000/9534 [=====================>........] - ETA: 0s - loss: 6.2513 - acc: 0.3244
8000/9534 [========================>.....] - ETA: 0s - loss: 6.1465 - acc: 0.3340
9000/9534 [===========================>..] - ETA: 0s - loss: 6.0439 - acc: 0.3411
9534/9534 [==============================] - 1s - loss: 5.9900 - acc: 0.3442 - val_loss: 4.8716 - val_acc: 0.4377
Epoch 2/100
1000/9534 [==>...........................] - ETA: 0s - loss: 4.8370 - acc: 0.4340
2000/9534 [=====>........................] - ETA: 0s - loss: 4.7593 - acc: 0.4415
3000/9534 [========>.....................] - ETA: 0s - loss: 4.6923 - acc: 0.4423
4000/9534 [===========>..................] - ETA: 0s - loss: 4.6176 - acc: 0.4557
5000/9534 [==============>...............] - ETA: 0s - loss: 4.5517 - acc: 0.4642
6000/9534 [=================>............] - ETA: 0s - loss: 4.4809 - acc: 0.4703
7000/9534 [=====================>........] - ETA: 0s - loss: 4.4036 - acc: 0.4804
8000/9534 [========================>.....] - ETA: 0s - loss: 4.3364 - acc: 0.4821
9000/9534 [===========================>..] - ETA: 0s - loss: 4.2652 - acc: 0.4901
9534/9534 [==============================] - 1s - loss: 4.2316 - acc: 0.4928 - val_loss: 3.5151 - val_acc: 0.5179
Epoch 3/100
1000/9534 [==>...........................] - ETA: 1s - loss: 3.4892 - acc: 0.5370
2000/9534 [=====>........................] - ETA: 1s - loss: 3.4573 - acc: 0.5395
3000/9534 [========>.....................] - ETA: 0s - loss: 3.4006 - acc: 0.5450
4000/9534 [===========>..................] - ETA: 0s - loss: 3.3430 - acc: 0.5435
5000/9534 [==============>...............] - ETA: 0s - loss: 3.2929 - acc: 0.5448
6000/9534 [=================>............] - ETA: 0s - loss: 3.2414 - acc: 0.5448
7000/9534 [=====================>........] - ETA: 0s - loss: 3.1959 - acc: 0.5446
8000/9534 [========================>.....] - ETA: 0s - loss: 3.1489 - acc: 0.5485
9000/9534 [===========================>..] - ETA: 0s - loss: 3.1021 - acc: 0.5501
9534/9534 [==============================] - 1s - loss: 3.0832 - acc: 0.5481 - val_loss: 2.6184 - val_acc: 0.5349
Epoch 4/100
1000/9534 [==>...........................] - ETA: 1s - loss: 2.5950 - acc: 0.5640
2000/9534 [=====>........................] - ETA: 1s - loss: 2.5570 - acc: 0.5705
3000/9534 [========>.....................] - ETA: 0s - loss: 2.5197 - acc: 0.5743
4000/9534 [===========>..................] - ETA: 0s - loss: 2.4929 - acc: 0.5650
5000/9534 [==============>...............] - ETA: 0s - loss: 2.4703 - acc: 0.5646
6000/9534 [=================>............] - ETA: 0s - loss: 2.4388 - acc: 0.5648
7000/9534 [=====================>........] - ETA: 0s - loss: 2.4054 - acc: 0.5680
8000/9534 [========================>.....] - ETA: 0s - loss: 2.3798 - acc: 0.5649
9000/9534 [===========================>..] - ETA: 0s - loss: 2.3522 - acc: 0.5662
9534/9534 [==============================] - 1s - loss: 2.3342 - acc: 0.5685 - val_loss: 2.0442 - val_acc: 0.5491
Epoch 5/100
1000/9534 [==>...........................] - ETA: 0s - loss: 2.0090 - acc: 0.5830
2000/9534 [=====>........................] - ETA: 0s - loss: 1.9990 - acc: 0.5865
3000/9534 [========>.....................] - ETA: 0s - loss: 1.9812 - acc: 0.5833
4000/9534 [===========>..................] - ETA: 0s - loss: 1.9558 - acc: 0.5835
5000/9534 [==============>...............] - ETA: 0s - loss: 1.9377 - acc: 0.5832
6000/9534 [=================>............] - ETA: 0s - loss: 1.9173 - acc: 0.5832
7000/9534 [=====================>........] - ETA: 0s - loss: 1.8968 - acc: 0.5850
8000/9534 [========================>.....] - ETA: 0s - loss: 1.8759 - acc: 0.5851
9000/9534 [===========================>..] - ETA: 0s - loss: 1.8582 - acc: 0.5846
9534/9534 [==============================] - 1s - loss: 1.8501 - acc: 0.5834 - val_loss: 1.6868 - val_acc: 0.5500
Epoch 6/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.6716 - acc: 0.5790
2000/9534 [=====>........................] - ETA: 0s - loss: 1.6387 - acc: 0.5910
3000/9534 [========>.....................] - ETA: 0s - loss: 1.6163 - acc: 0.5910
4000/9534 [===========>..................] - ETA: 0s - loss: 1.6130 - acc: 0.5882
5000/9534 [==============>...............] - ETA: 0s - loss: 1.5982 - acc: 0.5890
6000/9534 [=================>............] - ETA: 0s - loss: 1.5861 - acc: 0.5892
7000/9534 [=====================>........] - ETA: 0s - loss: 1.5724 - acc: 0.5914
8000/9534 [========================>.....] - ETA: 0s - loss: 1.5578 - acc: 0.5922
9000/9534 [===========================>..] - ETA: 0s - loss: 1.5492 - acc: 0.5904
9534/9534 [==============================] - 0s - loss: 1.5468 - acc: 0.5893 - val_loss: 1.4677 - val_acc: 0.5585
Epoch 7/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.4380 - acc: 0.5790
2000/9534 [=====>........................] - ETA: 0s - loss: 1.4332 - acc: 0.5900
3000/9534 [========>.....................] - ETA: 0s - loss: 1.4208 - acc: 0.5957
4000/9534 [===========>..................] - ETA: 0s - loss: 1.4073 - acc: 0.5985
5000/9534 [==============>...............] - ETA: 0s - loss: 1.4027 - acc: 0.5960
6000/9534 [=================>............] - ETA: 0s - loss: 1.3922 - acc: 0.5950
7000/9534 [=====================>........] - ETA: 0s - loss: 1.3842 - acc: 0.5951
8000/9534 [========================>.....] - ETA: 0s - loss: 1.3729 - acc: 0.5988
9000/9534 [===========================>..] - ETA: 0s - loss: 1.3611 - acc: 0.6012
9534/9534 [==============================] - 1s - loss: 1.3588 - acc: 0.6015 - val_loss: 1.3387 - val_acc: 0.5717
Epoch 8/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.3429 - acc: 0.5750
2000/9534 [=====>........................] - ETA: 0s - loss: 1.3071 - acc: 0.5980
3000/9534 [========>.....................] - ETA: 0s - loss: 1.2915 - acc: 0.6007
4000/9534 [===========>..................] - ETA: 0s - loss: 1.2834 - acc: 0.5977
5000/9534 [==============>...............] - ETA: 0s - loss: 1.2791 - acc: 0.6008
6000/9534 [=================>............] - ETA: 0s - loss: 1.2636 - acc: 0.6043
7000/9534 [=====================>........] - ETA: 0s - loss: 1.2521 - acc: 0.6049
8000/9534 [========================>.....] - ETA: 0s - loss: 1.2495 - acc: 0.6041
9000/9534 [===========================>..] - ETA: 0s - loss: 1.2506 - acc: 0.6031
9534/9534 [==============================] - 1s - loss: 1.2491 - acc: 0.6022 - val_loss: 1.2617 - val_acc: 0.5698
Epoch 9/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.1627 - acc: 0.6240
2000/9534 [=====>........................] - ETA: 0s - loss: 1.1709 - acc: 0.6235
3000/9534 [========>.....................] - ETA: 0s - loss: 1.2001 - acc: 0.6127
4000/9534 [===========>..................] - ETA: 0s - loss: 1.2000 - acc: 0.6098
5000/9534 [==============>...............] - ETA: 0s - loss: 1.2002 - acc: 0.6096
6000/9534 [=================>............] - ETA: 0s - loss: 1.1969 - acc: 0.6085
7000/9534 [=====================>........] - ETA: 0s - loss: 1.1894 - acc: 0.6117
9534/9534 [==============================] - 1s - loss: 1.1793 - acc: 0.6094 - val_loss: 1.2151 - val_acc: 0.5679
Epoch 10/100
1000/9534 [==>...........................] - ETA: 1s - loss: 1.1436 - acc: 0.6190
2000/9534 [=====>........................] - ETA: 0s - loss: 1.1369 - acc: 0.6260
3000/9534 [========>.....................] - ETA: 0s - loss: 1.1366 - acc: 0.6207
4000/9534 [===========>..................] - ETA: 0s - loss: 1.1293 - acc: 0.6210
5000/9534 [==============>...............] - ETA: 0s - loss: 1.1276 - acc: 0.6232
6000/9534 [=================>............] - ETA: 0s - loss: 1.1289 - acc: 0.6217
7000/9534 [=====================>........] - ETA: 0s - loss: 1.1321 - acc: 0.6180
8000/9534 [========================>.....] - ETA: 0s - loss: 1.1352 - acc: 0.6150
9000/9534 [===========================>..] - ETA: 0s - loss: 1.1341 - acc: 0.6141
9534/9534 [==============================] - 0s - loss: 1.1349 - acc: 0.6129 - val_loss: 1.1946 - val_acc: 0.5632
Epoch 11/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.1684 - acc: 0.5930
2000/9534 [=====>........................] - ETA: 0s - loss: 1.1338 - acc: 0.6075
3000/9534 [========>.....................] - ETA: 0s - loss: 1.1177 - acc: 0.6140
4000/9534 [===========>..................] - ETA: 0s - loss: 1.1293 - acc: 0.6075
5000/9534 [==============>...............] - ETA: 0s - loss: 1.1235 - acc: 0.6154
6000/9534 [=================>............] - ETA: 0s - loss: 1.1188 - acc: 0.6173
7000/9534 [=====================>........] - ETA: 0s - loss: 1.1147 - acc: 0.6179
8000/9534 [========================>.....] - ETA: 0s - loss: 1.1068 - acc: 0.6196
9000/9534 [===========================>..] - ETA: 0s - loss: 1.1090 - acc: 0.6190
9534/9534 [==============================] - 0s - loss: 1.1092 - acc: 0.6177 - val_loss: 1.1788 - val_acc: 0.5689
Epoch 12/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0702 - acc: 0.6280
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0742 - acc: 0.6280
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0821 - acc: 0.6237
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0868 - acc: 0.6233
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0807 - acc: 0.6258
6000/9534 [=================>............] - ETA: 0s - loss: 1.0884 - acc: 0.6208
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0905 - acc: 0.6187
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0895 - acc: 0.6205
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0899 - acc: 0.6200
9534/9534 [==============================] - 1s - loss: 1.0900 - acc: 0.6205 - val_loss: 1.1598 - val_acc: 0.5830
Epoch 13/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0730 - acc: 0.6340
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0649 - acc: 0.6445
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0600 - acc: 0.6430
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0718 - acc: 0.6350
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0821 - acc: 0.6280
6000/9534 [=================>............] - ETA: 0s - loss: 1.0779 - acc: 0.6295
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0713 - acc: 0.6316
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0737 - acc: 0.6289
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0767 - acc: 0.6261
9534/9534 [==============================] - 1s - loss: 1.0752 - acc: 0.6259 - val_loss: 1.1589 - val_acc: 0.5642
Epoch 14/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0148 - acc: 0.6520
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0395 - acc: 0.6430
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0503 - acc: 0.6377
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0521 - acc: 0.6382
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0529 - acc: 0.6388
6000/9534 [=================>............] - ETA: 0s - loss: 1.0519 - acc: 0.6392
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0561 - acc: 0.6359
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0547 - acc: 0.6332
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0591 - acc: 0.6313
9534/9534 [==============================] - 0s - loss: 1.0606 - acc: 0.6301 - val_loss: 1.1577 - val_acc: 0.5774
Epoch 15/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0513 - acc: 0.6410
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0635 - acc: 0.6245
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0500 - acc: 0.6280
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0530 - acc: 0.6257
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0585 - acc: 0.6232
6000/9534 [=================>............] - ETA: 0s - loss: 1.0562 - acc: 0.6233
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0507 - acc: 0.6267
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0540 - acc: 0.6267
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0513 - acc: 0.6286
9534/9534 [==============================] - 0s - loss: 1.0492 - acc: 0.6290 - val_loss: 1.1608 - val_acc: 0.5802
Epoch 16/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0553 - acc: 0.6300
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0582 - acc: 0.6305
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0341 - acc: 0.6407
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0312 - acc: 0.6398
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0454 - acc: 0.6324
6000/9534 [=================>............] - ETA: 0s - loss: 1.0438 - acc: 0.6332
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0445 - acc: 0.6323
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0426 - acc: 0.6331
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0439 - acc: 0.6323
9534/9534 [==============================] - 0s - loss: 1.0427 - acc: 0.6323 - val_loss: 1.1544 - val_acc: 0.5764
Epoch 17/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0633 - acc: 0.6190
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0407 - acc: 0.6300
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0417 - acc: 0.6343
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0322 - acc: 0.6402
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0283 - acc: 0.6426
6000/9534 [=================>............] - ETA: 0s - loss: 1.0355 - acc: 0.6400
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0361 - acc: 0.6413
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0336 - acc: 0.6392
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0309 - acc: 0.6394
9534/9534 [==============================] - 0s - loss: 1.0342 - acc: 0.6382 - val_loss: 1.1575 - val_acc: 0.5755
Epoch 18/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0289 - acc: 0.6510
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0233 - acc: 0.6505
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0176 - acc: 0.6507
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0194 - acc: 0.6500
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0242 - acc: 0.6442
6000/9534 [=================>............] - ETA: 0s - loss: 1.0239 - acc: 0.6423
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0249 - acc: 0.6413
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0264 - acc: 0.6404
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0277 - acc: 0.6406
9534/9534 [==============================] - 0s - loss: 1.0299 - acc: 0.6389 - val_loss: 1.1597 - val_acc: 0.5708
Epoch 19/100
1000/9534 [==>...........................] - ETA: 0s - loss: 1.0271 - acc: 0.6420
2000/9534 [=====>........................] - ETA: 0s - loss: 1.0114 - acc: 0.6445
3000/9534 [========>.....................] - ETA: 0s - loss: 1.0046 - acc: 0.6510
4000/9534 [===========>..................] - ETA: 0s - loss: 1.0137 - acc: 0.6453
5000/9534 [==============>...............] - ETA: 0s - loss: 1.0074 - acc: 0.6492
6000/9534 [=================>............] - ETA: 0s - loss: 1.0112 - acc: 0.6490
7000/9534 [=====================>........] - ETA: 0s - loss: 1.0072 - acc: 0.6504
8000/9534 [========================>.....] - ETA: 0s - loss: 1.0093 - acc: 0.6496
9000/9534 [===========================>..] - ETA: 0s - loss: 1.0137 - acc: 0.6452
9534/9534 [==============================] - 0s - loss: 1.0159 - acc: 0.6451 - val_loss: 1.1603 - val_acc: 0.5651

【问题讨论】：

您能发布一个数据集的样本吗？或图形表示。
感谢您的回复，米切托努。在机器学习领域，我总是对如何找到训练数据集的上限性能感到困惑。顺便问一下，我怎样才能将我的数据传递给您？
对于初学者来说，单行是什么样的？是 507 个二进制元素吗？没有直接的方法可以找到给定数据集的可能性能上限。如果数据集都是二进制的，那么在预处理方面就没什么可做的了。您是否在将数据输入网络之前对其进行洗牌？如果是这样，怎么做？您是否可能混淆了标签？
一行看起来像该特征被编码为长度为 507 的 (value=0 或 value=1) 数组。标签被编码为 1000000,0100000,0100000...
我在进行交叉验证时通过 sklearn 对数据进行了混洗。代码为 KFold(n_splits=cross_validation, shuffle=True)。什么是混淆标签？我认为标签类 A 和类 B 有时与其他类相似。混合标签意味着我将classA和classB标记为一个新的大类并将标签计数从7转换为6？

标签： machine-learning neural-network keras gradient-descent convergence

【解决方案1】：

交叉验证代码如下：

skf = KFold(n_splits=cross_validation, shuffle=True)
for train_index, test_index in skf.split(X):
    X_train, X_test = X[train_index], X[test_index]
    y_train, y_test = Y[train_index], Y[test_index]
    model = None
    model = self.__create_model()
    model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=(X_test, y_test))

X 和 Y 是形状为 (10000, 507) 和 (10000, 7) 的两个矩阵

【讨论】：

【解决方案2】：

对于任何模型，您可以达到多少准确度显然很大程度上取决于您的数据集和标签的准确度。

使用经过训练的模型进行预测并生成混淆矩阵。看看假阳性和假阴性预测的具体例子。这些预测实际上是错误的还是模型预测的比标签更准确？我在项目中多次遇到这种情况。

我建议您首先尝试训练您的模型，直到它过拟合。据我所知，您的模型仍在学习或处于过度拟合的边缘。添加更多的时期，直到验证损失和/或准确性再次变得更糟。然后，对于后续尝试，如果需要，请应用正则化。从 0.1 或 .2 到 0.5 最大的 dropout 开始。

设置 Tensorboard，以便您可以跟踪准确度和验证准确度之间的差异。

有这么多分类变量：您确定避免了虚拟变量陷阱吗？您需要确保虚拟变量的数量比您的类别数量少一：http://www.algosome.com/articles/dummy-variable-trap-regression.html

【讨论】：

【解决方案3】：

非常感谢 petezurich 的回复。我完全同意你的看法。你真好。

我正在尝试我的另一个多标签分类项目。

数据是 503 个二进制特征和 64 个二进制标签。我在输出层使用 sigmoid，在损失函数上使用 binary_crossentrophy，在第一个隐藏层使用正则化 l2。隐藏层结构为 500*100。所以整个网络就像503*500*100*64。我在这里发布我的数据： https://ufile.io/f4tvf

当我在 10 个文件夹交叉验证中分别进行 10000 个 epoch 时，我得到了性能分数：

coverage error: 10.485887, 
ranking average precision: 0.766574, 
ranking loss: 0.045134.

如果我为每个标签设置 0.5 个阈值，我会得到基于标签的指标

zero one loss: 0.848790, 
hamming loss: 0.037109, 
macro precision: 0.426696, 
micro precision: 0.672705, 
macro recall: 0.371033, 
micro recall: 0.636571, 
macro f1: 0.383845, 
micro f1: 0.654140.

对结果有好处吗？有人有兴趣训练它吗？

【讨论】：