search

lmartak / distill-nn-tree

Distillation of Neural Network Into a Soft Decision Tree
https://vgg.fiit.stuba.sk/people/martak/distill-nn-tree
MIT License
64 stars 29 forks source link

Hello, I used your model to train my own data set, one-dimensional signal data set, but when I trained, I found that the accuracy rate was very low, so I did not move 0.08. Why? #3

Closed 1448643857 closed 5 years ago

1448643857 commented 5 years ago

thank u very much

1448643857 commented 5 years ago

Epoch 1/500 24000/24000 [==============================] - 718s 30ms/step - loss: 14.0100 - acc: 0.0835 - val_loss: 13.9278 - val_acc: 0.0883 Epoch 2/500 24000/24000 [==============================] - 719s 30ms/step - loss: 13.9117 - acc: 0.1274 - val_loss: 13.9540 - val_acc: 0.0847 Epoch 3/500 24000/24000 [==============================] - 718s 30ms/step - loss: 13.7736 - acc: 0.2407 - val_loss: 14.1120 - val_acc: 0.0883 Epoch 4/500 24000/24000 [==============================] - 716s 30ms/step - loss: 13.4118 - acc: 0.3258 - val_loss: 14.5342 - val_acc: 0.0895 Epoch 5/500 24000/24000 [==============================] - 719s 30ms/step - loss: 12.7447 - acc: 0.3750 - val_loss: 14.8735 - val_acc: 0.0938 Epoch 6/500 24000/24000 [==============================] - 718s 30ms/step - loss: 11.8372 - acc: 0.3967 - val_loss: 15.1256 - val_acc: 0.0912 Epoch 7/500 24000/24000 [==============================] - 720s 30ms/step - loss: 10.8221 - acc: 0.4143 - val_loss: 15.2756 - val_acc: 0.0888 Epoch 8/500 24000/24000 [==============================] - 719s 30ms/step - loss: 9.8527 - acc: 0.4263 - val_loss: 15.3327 - val_acc: 0.0922 Epoch 9/500 24000/24000 [==============================] - 718s 30ms/step - loss: 9.0603 - acc: 0.4356 - val_loss: 15.3739 - val_acc: 0.0935 Epoch 10/500 24000/24000 [==============================] - 718s 30ms/step - loss: 8.4028 - acc: 0.4433 - val_loss: 15.3867 - val_acc: 0.0940 Epoch 11/500 24000/24000 [==============================] - 720s 30ms/step - loss: 7.9059 - acc: 0.4503 - val_loss: 15.4083 - val_acc: 0.0947 Epoch 12/500 24000/24000 [==============================] - 719s 30ms/step - loss: 7.4681 - acc: 0.4538 - val_loss: 15.4072 - val_acc: 0.0932 Epoch 13/500 24000/24000 [==============================] - 719s 30ms/step - loss: 7.1606 - acc: 0.4591 - val_loss: 15.4000 - val_acc: 0.0897 Epoch 14/500 24000/24000 [==============================] - 720s 30ms/step - loss: 6.9462 - acc: 0.4662 - val_loss: 15.3871 - val_acc: 0.0942 Epoch 15/500 24000/24000 [==============================] - 722s 30ms/step - loss: 6.7773 - acc: 0.4675 - val_loss: 15.3829 - val_acc: 0.0920 Epoch 16/500 24000/24000 [==============================] - 719s 30ms/step - loss: 6.6513 - acc: 0.4730 - val_loss: 15.3680 - val_acc: 0.0892 Epoch 17/500 24000/24000 [==============================] - 718s 30ms/step - loss: 6.5321 - acc: 0.4774 - val_loss: 15.3624 - val_acc: 0.0913 Epoch 18/500 24000/24000 [==============================] - 714s 30ms/step - loss: 6.4380 - acc: 0.4801 - val_loss: 15.3570 - val_acc: 0.0942

Like this

lmartak commented 5 years ago

As your training accuracy increases and training loss decreases, this looks like a problem of generalization related to the data you're using. As the problem is not with this code but with how its used, I'm closing the issue.