22
$\begingroup$

I setup a grid search for a bunch of params. I am trying to find the best parameters for a Keras neural net that does binary classification. The output is either a 1 or a 0. There are about 200 features. When I did a grid search, I got a bunch of models and their parameters. The best model had these parameters: 

Epochs           : 20
Batch Size       : 10
First Activation : sigmoid
Learning Rate    : 1
First Init       : uniform

and the results for that model were : 

        loss       acc  val_loss   val_acc
1   0.477424  0.768542  0.719960  0.722550
2   0.444588  0.788861  0.708650  0.732130
3   0.435809  0.794336  0.695768  0.732682
4   0.427056  0.798784  0.684516  0.721137
5   0.420828  0.803048  0.703748  0.720707
6   0.418129  0.806206  0.730803  0.723717
7   0.417522  0.805206  0.778434  0.721936
8   0.415197  0.807549  0.802040  0.733849
9   0.412922  0.808865  0.823036  0.731761
10  0.410463  0.810654  0.839087  0.730410
11  0.407369  0.813892  0.831844  0.725252
12  0.404436  0.815760  0.835217  0.723102
13  0.401728  0.816287  0.845178  0.722488
14  0.399623  0.816471  0.842231  0.717514
15  0.395746  0.819498  0.847118  0.719541
16  0.393361  0.820366  0.858291  0.714873
17  0.390947  0.822025  0.850880  0.723348
18  0.388478  0.823341  0.858591  0.721014
19  0.387062  0.822735  0.862971  0.721936
20  0.383744  0.825762  0.880477  0.721322

So I reran that model with more epochs (150 of them) and these are the results I got. I am not sure why this is happening, is this normal or what am I doing wrong?

Plot of the loss, acc, val_locc, and val_acc

         loss       acc  val_loss   val_acc
1    0.476387  0.769279  0.728492  0.722550
2    0.442604  0.789941  0.701136  0.730472
3    0.431936  0.796915  0.676995  0.723655
4    0.426349  0.800258  0.728562  0.721997
5    0.421143  0.803653  0.739789  0.716900
6    0.416389  0.807575  0.720850  0.711373
7    0.413163  0.809154  0.751340  0.718128
8    0.409013  0.811418  0.780856  0.723409
9    0.405871  0.813576  0.789046  0.719295
10   0.402579  0.815524  0.804526  0.720278
11   0.400152  0.816813  0.811905  0.719541
12   0.400304  0.817261  0.787449  0.713154
13   0.397917  0.817945  0.804222  0.721567
14   0.395266  0.819524  0.801722  0.723348
15   0.393957  0.820156  0.793889  0.719049
16   0.391780  0.821103  0.794179  0.721199
17   0.390206  0.822393  0.806803  0.722611
18   0.388075  0.823604  0.817850  0.723901
19   0.385985  0.824762  0.841883  0.722058
20   0.383762  0.826867  0.857071  0.720830
21   0.381493  0.827947  0.864432  0.718005
22   0.379520  0.829210  0.872835  0.720400
23   0.377488  0.830526  0.879962  0.721383
24   0.375619  0.830736  0.887850  0.723839
25   0.373684  0.832000  0.891267  0.724822
26   0.372023  0.832368  0.891562  0.724638
27   0.370155  0.833184  0.892528  0.724883
28   0.368511  0.834684  0.887061  0.724699
29   0.366522  0.835606  0.883541  0.724883
30   0.364500  0.836422  0.882823  0.724515
31   0.362612  0.836737  0.882611  0.722427
32   0.360742  0.837448  0.884282  0.720769
33   0.359093  0.838738  0.884339  0.719418
34   0.357436  0.839080  0.888006  0.716470
35   0.355723  0.840633  0.892658  0.713830
36   0.354305  0.840764  0.897303  0.710575
37   0.352758  0.841343  0.901147  0.709408
38   0.351414  0.842054  0.899546  0.707934
39   0.349619  0.843370  0.905133  0.704864
40   0.347993  0.844475  0.910400  0.701363
41   0.346402  0.845581  0.915086  0.699337
42   0.345014  0.845818  0.918697  0.697617
43   0.343708  0.846607  0.923413  0.695652
44   0.342335  0.847292  0.930816  0.693441
45   0.340745  0.848081  0.940737  0.689020
46   0.339623  0.848713  0.948633  0.685274
47   0.338846  0.849845  0.952492  0.683923
48   0.337724  0.850134  0.961147  0.683984
49   0.336247  0.850976  0.967792  0.683309
50   0.334444  0.851529  0.984107  0.680238
51   0.333086  0.852029  1.001179  0.678273
52   0.331756  0.853240  1.016130  0.674589
53   0.330738  0.854003  1.024875  0.673606
54   0.329548  0.854030  1.040597  0.670044
55   0.328813  0.855372  1.041871  0.668509
56   0.327120  0.855898  1.050617  0.668755
57   0.325962  0.855819  1.064525  0.666667
58   0.324602  0.856898  1.078078  0.662859
59   0.323560  0.857241  1.085016  0.661938
60   0.322243  0.858662  1.093114  0.661140
61   0.320680  0.858872  1.117269  0.656841
62   0.319267  0.860004  1.138825  0.654815
63   0.318132  0.860636  1.154959  0.653648
64   0.316956  0.861531  1.180216  0.649718
65   0.315543  0.862320  1.198216  0.648428
66   0.314405  0.862610  1.218663  0.647384
67   0.313501  0.863873  1.245123  0.644252
68   0.312513  0.864558  1.262998  0.643147
69   0.311567  0.865347  1.283213  0.641918
70   0.310069  0.866505  1.302089  0.640752
71   0.309087  0.866611  1.318972  0.641857
72   0.307767  0.867321  1.361531  0.638787
73   0.306750  0.866742  1.382162  0.638357
74   0.305760  0.867242  1.378694  0.641611
75   0.305289  0.867769  1.393187  0.642594
76   0.304089  0.868479  1.435852  0.635532
77   0.302472  0.869006  1.435019  0.639892
78   0.301118  0.869400  1.447060  0.639216
79   0.300629  0.870058  1.488730  0.634918
80   0.299364  0.870295  1.488376  0.636576
81   0.298380  0.870822  1.504260  0.634611
82   0.297253  0.871664  1.525655  0.634058
83   0.296760  0.871875  1.538717  0.632891
84   0.295502  0.872585  1.551178  0.633751
85   0.294569  0.872927  1.562323  0.633137
86   0.294780  0.872585  1.555390  0.629944
87   0.293796  0.872743  1.587800  0.627057
88   0.293029  0.873427  1.608010  0.627549
89   0.291822  0.874006  1.626047  0.627303
90   0.290643  0.874533  1.651658  0.626689
91   0.289920  0.875270  1.681202  0.623925
92   0.289661  0.875375  1.683188  0.626505
93   0.288103  0.876323  1.706517  0.625031
94   0.287917  0.876770  1.722031  0.624417
95   0.287020  0.877270  1.743283  0.624478
96   0.286750  0.877639  1.762506  0.624048
97   0.285712  0.877481  1.780433  0.622267
98   0.284635  0.878639  1.789917  0.622206
99   0.283627  0.879191  1.862468  0.616925
100  0.282214  0.879455  1.915643  0.612810
101  0.281749  0.879244  1.881444  0.615205
102  0.281710  0.879639  1.916390  0.614223
103  0.280293  0.880350  1.938470  0.612810
104  0.279233  0.881008  1.979127  0.609187
105  0.279204  0.880297  1.997384  0.606546
106  0.278264  0.881876  2.009851  0.607652
107  0.277511  0.882876  2.038530  0.606116
108  0.277521  0.881771  2.034664  0.604888
109  0.276264  0.882534  2.058179  0.604827
110  0.275230  0.883587  2.078912  0.604274
111  0.275147  0.883034  2.073272  0.603537
112  0.273717  0.883797  2.100150  0.600958
113  0.273372  0.883692  2.114416  0.601634
114  0.272626  0.883692  2.129778  0.601941
115  0.272001  0.883929  2.138462  0.601326
116  0.271344  0.884508  2.148771  0.602923
117  0.270134  0.884692  2.115114  0.604581
118  0.269494  0.885140  2.135719  0.603107
119  0.268803  0.885587  2.162380  0.601695
120  0.268593  0.886219  2.183793  0.599239
121  0.267141  0.886035  2.195810  0.600221
122  0.266565  0.886772  2.192426  0.600528
123  0.265715  0.886561  2.260088  0.596598
124  0.264788  0.887693  2.253029  0.597335
125  0.263643  0.887693  2.289285  0.597028
126  0.263612  0.887956  2.311600  0.596536
127  0.261996  0.888588  2.339754  0.595063
128  0.263069  0.887588  2.364881  0.594449
129  0.261684  0.889272  2.321568  0.596598
130  0.261304  0.889509  2.389324  0.591562
131  0.260336  0.889640  2.403542  0.593098
132  0.259131  0.890272  2.413964  0.592115
133  0.258756  0.890193  2.422454  0.591992
134  0.257794  0.891009  2.454598  0.591255
135  0.257187  0.891009  2.459366  0.590088
136  0.257249  0.891088  2.448625  0.591624
137  0.256344  0.891404  2.495104  0.589167
138  0.255590  0.891720  2.495032  0.589781
139  0.254596  0.892299  2.496050  0.589229
140  0.254308  0.892588  2.510471  0.589536
141  0.253694  0.892509  2.519580  0.589720
142  0.252973  0.893088  2.527464  0.590273
143  0.252714  0.893194  2.553902  0.589106
144  0.252190  0.893720  2.536494  0.590457
145  0.251870  0.893352  2.553102  0.588799
146  0.250437  0.893694  2.565141  0.589597
147  0.250066  0.894141  2.575599  0.588553
148  0.249596  0.894273  2.590722  0.588123
149  0.248569  0.894983  2.596031  0.588676
150  0.248096  0.895273  2.602810  0.588860
Improve this question
$\endgroup$
3
  • $\begingroup$ You case is strange because your validation loss never got smaller. Your learning rate is suspiciously high, typical learning rates are about 0.001. What range of learning rates did you use in the grid search? $\endgroup$
    – Hugh
    Commented Feb 6, 2017 at 22:28
  • 1
    $\begingroup$ I used [1.000, 0.100, 0.010, 0.001] $\endgroup$
    – user1367204
    Commented Feb 7, 2017 at 1:44
  • $\begingroup$ This can be a bit late, but are you sure that your data is what you think it is? Specifically it is very odd that your validation accuracy is stagnating, while the validation loss is increasing, because those two values should always move together, eg. the decrease in the loss value should be coupled with proportional increase in accuracy. You can see that in the case of training loss. As the training loss is decreasing so is the accuracy increasing. However this is not the case of the validation data you have. Therefore I would definitely looked into how you are getting validation loss and ac $\endgroup$
    – matt_m
    Commented May 19, 2018 at 18:07

2 Answers 2

Reset to default
30
$\begingroup$

(this may be a duplicate) It looks like your model is over fitting, that is just memorizing the training data. In general a model that over fits can be improved by adding more dropout, or training and validating on a larger data set. Explain more about the data/features and the model for further ideas.

Improve this answer
$\endgroup$
2
  • 8
    $\begingroup$ It's a very peculiar overfitting. val_loss never decreases as if there was no fitting before overfitting began. Do you know what could explain that? $\endgroup$
    – Hugh
    Commented Feb 6, 2017 at 22:27
  • $\begingroup$ this is the wrong answer. it shouldn't be overfitting right from the beginning $\endgroup$
    – Fine-Tuning
    Commented Mar 7 at 23:21
11
$\begingroup$

Perhaps your training dataset has different properties than your validation dataset. It's like training a network to distinguish between a chicken and an airplane, but then you show it an apple. The more you train it, the better it is at distinguishing chickens from airplanes, but also the worse it is when it is shown an apple. I'm having the same situation and am thinking of using a Generative Adversarial Network to identify if a validation data point is "alien" to the training dataset or not

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.