后端模型训练结果的比较

Question

我在两个cnn网络上训练了6000张图片，它们似乎给了我类似的结果。该模型进行二值分类训练，两个模型之间的差异在稠密层中，模型A具有下面的稠密层配置

fc1 (Dense)                  (None, 512)               12845568  
_________________________________________________________________
fc2 (Dense)                  (None, 256)               131328    
_________________________________________________________________
output (Dense)               (None, 2)                 514 

和结果

Train on 4800 samples, validate on 1200 samples
Epoch 1/30
4800/4800 [===] - 98s - loss: 0.7923 - acc: 0.6865 - val_loss: 0.4599 - val_acc: 0.7858

Epoch 2/30
4800/4800 [===] - 80s - loss: 0.4263 - acc: 0.7996 - val_loss: 0.5913 - val_acc: 0.6350

Epoch 3/30
4800/4800 [===] - 80s - loss: 0.3912 - acc: 0.8133 - val_loss: 0.3199 - val_acc: 0.8625

Epoch 4/30
4800/4800 [===] - 80s - loss: 0.3562 - acc: 0.8402 - val_loss: 0.3086 - val_acc: 0.8708

Epoch 5/30
4800/4800 [===] - 80s - loss: 0.3251 - acc: 0.8558 - val_loss: 0.2784 - val_acc: 0.8817

Epoch 6/30
4800/4800 [===] - 80s - loss: 0.3150 - acc: 0.8631 - val_loss: 0.2792 - val_acc: 0.8817

Epoch 7/30
4800/4800 [===] - 80s - loss: 0.2997 - acc: 0.8692 - val_loss: 0.3615 - val_acc: 0.8342

Epoch 8/30
4800/4800 [===] - 80s - loss: 0.2990 - acc: 0.8662 - val_loss: 0.2630 - val_acc: 0.8908

Epoch 9/30
4800/4800 [===] - 80s - loss: 0.2594 - acc: 0.8867 - val_loss: 0.3102 - val_acc: 0.8700

Epoch 10/30
4800/4800 [===] - 80s - loss: 0.2846 - acc: 0.8785 - val_loss: 0.4234 - val_acc: 0.7842

Epoch 11/30
4800/4800 [===] - 80s - loss: 0.2510 - acc: 0.8969 - val_loss: 0.2952 - val_acc: 0.8742

Epoch 12/30
4800/4800 [===] - 80s - loss: 0.2288 - acc: 0.9090 - val_loss: 0.2680 - val_acc: 0.8858

Epoch 13/30
4800/4800 [===] - 80s - loss: 0.2277 - acc: 0.9044 - val_loss: 0.3745 - val_acc: 0.8600

Epoch 14/30
4800/4800 [===] - 80s - loss: 0.2659 - acc: 0.8873 - val_loss: 0.2438 - val_acc: 0.9025

Epoch 15/30
4800/4800 [===] - 80s - loss: 0.2101 - acc: 0.9133 - val_loss: 0.3176 - val_acc: 0.8667

Epoch 16/30
4800/4800 [===] - 80s - loss: 0.2094 - acc: 0.9146 - val_loss: 0.2763 - val_acc: 0.8875

Epoch 17/30
4800/4800 [===] - 80s - loss: 0.2058 - acc: 0.9125 - val_loss: 0.2677 - val_acc: 0.8925

Epoch 18/30
4800/4800 [===] - 80s - loss: 0.1839 - acc: 0.9296 - val_loss: 0.2449 - val_acc: 0.9117

Epoch 19/30
4800/4800 [===] - 80s - loss: 0.1918 - acc: 0.9221 - val_loss: 0.2471 - val_acc: 0.8992

Epoch 20/30
4800/4800 [===] - 80s - loss: 0.2014 - acc: 0.9225 - val_loss: 0.2709 - val_acc: 0.8808

Epoch 21/30
4800/4800 [===] - 80s - loss: 0.1540 - acc: 0.9425 - val_loss: 0.2541 - val_acc: 0.8933

Epoch 22/30
4800/4800 [===] - 80s - loss: 0.1803 - acc: 0.9294 - val_loss: 0.2289 - val_acc: 0.9058

Epoch 23/30
4800/4800 [===] - 80s - loss: 0.1548 - acc: 0.9425 - val_loss: 0.2417 - val_acc: 0.9175

Epoch 24/30
4800/4800 [===] - 80s - loss: 0.1754 - acc: 0.9294 - val_loss: 0.4914 - val_acc: 0.8183

Epoch 25/30
4800/4800 [===] - 80s - loss: 0.1449 - acc: 0.9419 - val_loss: 0.2281 - val_acc: 0.9125

Epoch 26/30
4800/4800 [===] - 80s - loss: 0.1529 - acc: 0.9385 - val_loss: 0.2328 - val_acc: 0.9217

Epoch 27/30
4800/4800 [===] - 80s - loss: 0.1237 - acc: 0.9533 - val_loss: 0.2646 - val_acc: 0.9167

Epoch 28/30
4800/4800 [===] - 80s - loss: 0.1236 - acc: 0.9531 - val_loss: 0.2485 - val_acc: 0.9100

Epoch 29/30
4800/4800 [===] - 80s - loss: 0.1301 - acc: 0.9500 - val_loss: 0.2726 - val_acc: 0.9042

Epoch 30/30
4800/4800 [===] - 80s - loss: 0.1335 - acc: 0.9500 - val_loss: 0.2803 - val_acc: 0.9183

Training time: 2440.315860271454
1200/1200 [===] - 27s    
[INFO] loss=0.2803, accuracy: 91.8333%

=================================================================

Model B has following final dense layer config

fc1 (Dense)                  (None, 1024)              25691136  
_________________________________________________________________
fc2 (Dense)                  (None, 512)               524800    
_________________________________________________________________
output (Dense)               (None, 2)                 1026      


Result
Train on 4800 samples, validate on 1200 samples
Epoch 1/30
4800/4800 [===] - 87s - loss: 0.4743 - acc: 0.7708 - val_loss: 0.4073 - val_acc: 0.8233

Epoch 2/30
4800/4800 [===] - 87s - loss: 0.3732 - acc: 0.8263 - val_loss: 0.3359 - val_acc: 0.8525

Epoch 3/30
4800/4800 [===] - 87s - loss: 0.3383 - acc: 0.8500 - val_loss: 0.3017 - val_acc: 0.8658

Epoch 4/30
4800/4800 [===] - 87s - loss: 0.3094 - acc: 0.8637 - val_loss: 0.3024 - val_acc: 0.8683

Epoch 5/30
4800/4800 [===] - 87s - loss: 0.3036 - acc: 0.8669 - val_loss: 0.3848 - val_acc: 0.8058

Epoch 6/30
4800/4800 [===] - 87s - loss: 0.2848 - acc: 0.8802 - val_loss: 0.2730 - val_acc: 0.8883

Epoch 7/30
4800/4800 [===] - 87s - loss: 0.2630 - acc: 0.8877 - val_loss: 0.3234 - val_acc: 0.8667

Epoch 8/30
4800/4800 [===] - 87s - loss: 0.2491 - acc: 0.8952 - val_loss: 0.2758 - val_acc: 0.8933

Epoch 9/30
4800/4800 [===] - 87s - loss: 0.2484 - acc: 0.8992 - val_loss: 0.3271 - val_acc: 0.8467

Epoch 10/30
4800/4800 [===] - 87s - loss: 0.2427 - acc: 0.8992 - val_loss: 0.2743 - val_acc: 0.8808

Epoch 11/30
4800/4800 [===] - 87s - loss: 0.2346 - acc: 0.9017 - val_loss: 0.2379 - val_acc: 0.9008

Epoch 12/30
4800/4800 [===] - 87s - loss: 0.2250 - acc: 0.9108 - val_loss: 0.2432 - val_acc: 0.9017

Epoch 13/30
4800/4800 [===] - 87s - loss: 0.1993 - acc: 0.9221 - val_loss: 0.2892 - val_acc: 0.8858

Epoch 14/30
4800/4800 [===] - 87s - loss: 0.2148 - acc: 0.9125 - val_loss: 0.3201 - val_acc: 0.8842

Epoch 15/30
4800/4800 [===] - 87s - loss: 0.1823 - acc: 0.9287 - val_loss: 0.5481 - val_acc: 0.8133

Epoch 16/30
4800/4800 [===] - 87s - loss: 0.1873 - acc: 0.9281 - val_loss: 0.2449 - val_acc: 0.9092

Epoch 17/30
4800/4800 [===] - 87s - loss: 0.1622 - acc: 0.9392 - val_loss: 0.2373 - val_acc: 0.9092

Epoch 18/30
4800/4800 [===] - 87s - loss: 0.1782 - acc: 0.9304 - val_loss: 0.2856 - val_acc: 0.8725

Epoch 19/30
4800/4800 [===] - 87s - loss: 0.1632 - acc: 0.9369 - val_loss: 0.2518 - val_acc: 0.9067

Epoch 20/30
4800/4800 [===] - 87s - loss: 0.1577 - acc: 0.9381 - val_loss: 0.2629 - val_acc: 0.9050

Epoch 21/30
4800/4800 [===] - 87s - loss: 0.1395 - acc: 0.9481 - val_loss: 0.2278 - val_acc: 0.9133

Epoch 22/30
4800/4800 [===] - 87s - loss: 0.1422 - acc: 0.9444 - val_loss: 0.2232 - val_acc: 0.9158

Epoch 23/30
4800/4800 [===] - 87s - loss: 0.1436 - acc: 0.9448 - val_loss: 0.2862 - val_acc: 0.9042

Epoch 24/30
4800/4800 [===] - 87s - loss: 0.1402 - acc: 0.9448 - val_loss: 0.3186 - val_acc: 0.8842

Epoch 25/30
4800/4800 [===] - 86s - loss: 0.1261 - acc: 0.9542 - val_loss: 0.2762 - val_acc: 0.9092

Epoch 26/30
4800/4800 [===] - 86s - loss: 0.1143 - acc: 0.9529 - val_loss: 0.2442 - val_acc: 0.9125

Epoch 27/30
4800/4800 [===] - 86s - loss: 0.1141 - acc: 0.9565 - val_loss: 0.3128 - val_acc: 0.8658

Epoch 28/30
4800/4800 [===] - 86s - loss: 0.1092 - acc: 0.9606 - val_loss: 0.2669 - val_acc: 0.9092

Epoch 29/30
4800/4800 [===] - 86s - loss: 0.0939 - acc: 0.9642 - val_loss: 0.2535 - val_acc: 0.8975

Epoch 30/30
4800/4800 [===] - 86s - loss: 0.1098 - acc: 0.9615 - val_loss: 0.2594 - val_acc: 0.9008

Training time: 2615.465226173401
1200/1200 [==============================] - 30s    
[INFO] loss=0.2594, accuracy: 90.0833%

这两种模型似乎都给了类似的results.Is一个很好的结果，或者有什么异常我无法检测到？或者这个模型是个好模型？附加信息批号128个，loss=categorical交叉熵，优化器-adadelta，任何改进建议也是见效的。

Answer 1

接下来我要考虑的步骤是：

• 你的训练损失是怎么回事。这可能是一个有用的指标。如果您看到您的培训损失达到100%，尝试添加更多的正则化，例如L2正则化。
• 您的卷积网络是否使用剩余层和批处理规范化？在大多数情况下，使用批处理规范的剩余网络似乎是最先进的。
• 测试每个卷积层中不同数目的滤波器。如果你有太多的过滤器，如果你的过滤器太少，你就不适合。这里面有个好地方，这很不一样。
• 您的批处理大小也会影响您的结果，使用它，我已经对小数据集和大型数据集进行了广泛的超参数搜索，并提出了非常不同的理想批处理大小。
• 您可以测试Adam优化器。然而，AdaDelta也是坚实的。
• 训练一些模型，随机初始化会产生稍微不同的最终结果。使用一组模型将做得更好。
• 随机初始化与哈维尔初始化程序可能会给你一个小凸起的准确性(有一个不同的conv层和fc层)。
• 在验证错误稳定后降低学习速度，这通常会在每次降低时提高性能。

总有很多事情你可以尝试。开始运行实验，用这些建议中的一些不同的价值来训练几个模型，看看你在哪里得到了改进。