jacknudelman / facial-similarity- Goto Github PK

Jack Nudelman

https://drive.google.com/open?id=19WUt1OnSRTTXDCUQD72zvEHj4OWT_HEI


1a)
  Without Augmentation:
    - I chose to have 25 epochs. I realized that no matter how much the train
    loss decreased, the test loss maintained around .5 accuracy so it didn't
    matter how many epochs I did as long as I did more than about 10. This is
    because the network starts to memorize the training data after about 10
    epochs since it has already seen the same images multiple times.
    However, I do 25 epochs for the augmentation part so I did 25 here just to
    be consistent.

    final training accuracy = 1.0
    final testing accuracy = 0.528


  with augmentation:
  - I chose to pick 25 epochs here because I first did about 30 epochs and
  saw that the test accuracy would incrementally increase up until about 25
  epochs and then begin to slowly drop. I wanted to have the highest test
  accuracy possible so I cut it off after 25 epochs. It is important to note that
  the test accuracy fluctuated a lot because my learning rate was a bit too big
  so it would keep overfitting. However, 25 epochs seemed to consistently
  give me the best results so I stuck to it. 

    final training accuracy = 0.8131818181818182
    final testing accuracy = 0.548

    - comparing with no augmentation: We could see that the training accuracy is
    way higher without augmentation than with and that the loss function reaches
    a very low value way quicker without augmentation. This is because the
    augmentation doesn't allow for the net to memorize the data since it is
    essentially 'creating new data' by manipulating the images into images the net
    most likely hasn't seen before. On the other hand, you could see that,
    on average, the test accuracy does increase by a little with augmentation
    as opposed to no average increase without augmentation. This is because the
    net is getting better at classifying new data since it was trained with 'new'
    data. It is important to note that the increase however is tiny and this is
    because of the nature of BCELoss. This loss function just looks for
    specific features but doesn't look at the relationship between the images.



b)
  - I chose a margin of 2 because I tried with 1 and progressively increased up
  until 7 but two gave me the best results.

  Without Augmentation:
    - I chose to pick 20 epochs here because I first did about 30 epochs and
    saw that the test accuracy would incrementally increase up until about 20
    epochs and then begin to slowly drop. I wanted to have the highest test
    accuracy possible so I cut it off after 20 epochs. This still gave me a high
    training accuracy and the highest possible testing accuracy

    final training accuracy = 0.9121823
    final testing accuracy = 0.624

  With Augmentation















0
train accuracy on epoch  0  is  0.551818181818
current average testing accuracy is  0.5352
1
train accuracy on epoch  1  is  0.786818181818
current average testing accuracy is  0.542
2
train accuracy on epoch  2  is  0.896818181818
current average testing accuracy is  0.546066666667
3
train accuracy on epoch  3  is  0.945909090909
current average testing accuracy is  0.54445
4
train accuracy on epoch  4  is  0.969090909091
current average testing accuracy is  0.54516
5
train accuracy on epoch  5  is  0.978181818182
current average testing accuracy is  0.5445
6
train accuracy on epoch  6  is  0.983636363636
current average testing accuracy is  0.5446
7
train accuracy on epoch  7  is  0.993181818182
current average testing accuracy is  0.5443
8
train accuracy on epoch  8  is  0.995454545455
current average testing accuracy is  0.544066666667
9
train accuracy on epoch  9  is  0.995909090909
current average testing accuracy is  0.54322







augmenting
0
train accuracy on epoch  0  is  0.528
current average testing accuracy is  0.581
1
train accuracy on epoch  1  is  0.617
current average testing accuracy is  0.593
2
train accuracy on epoch  2  is  0.635
current average testing accuracy is  0.600666666667
3
train accuracy on epoch  3  is  0.67
current average testing accuracy is  0.6115
4
train accuracy on epoch  4  is  0.688
current average testing accuracy is  0.625
5
train accuracy on epoch  5  is  0.708
current average testing accuracy is  0.629285714286
6
train accuracy on epoch  6  is  0.721
current average testing accuracy is  0.632875
7
train accuracy on epoch  7  is  0.729
current average testing accuracy is  0.635666666667
8
train accuracy on epoch  8  is  0.763
current average testing accuracy is  0.6403
9
train accuracy on epoch  9  is  0.776
current average testing accuracy is  0.649083333333
10
train accuracy on epoch  10  is  0.774
current average testing accuracy is  0.651307692308
11
train accuracy on epoch  11  is  0.789
current average testing accuracy is  0.654071428571
12
train accuracy on epoch  12  is  0.796
current average testing accuracy is  0.657466666667
13
train accuracy on epoch  13  is  0.801
current average testing accuracy is  0.663882352941
14
train accuracy on epoch  14  is  0.814
current average testing accuracy is  0.666833333333
15
train accuracy on epoch  15  is  0.811
current average testing accuracy is  0.669578947368
16
train accuracy on epoch  16  is  0.815
current average testing accuracy is  0.6718
17
train accuracy on epoch  17  is  0.834
current average testing accuracy is  0.673666666667
18
train accuracy on epoch  18  is  0.809
current average testing accuracy is  0.678130434783
19
train accuracy on epoch  19  is  0.82
current average testing accuracy is  0.679541666667
20
train accuracy on epoch  20  is  0.83
current average testing accuracy is  0.6814
21
train accuracy on epoch  21  is  0.827
current average testing accuracy is  0.682769230769
22
train accuracy on epoch  22  is  0.843
current average testing accuracy is  0.68525
23
train accuracy on epoch  23  is  0.84
current average testing accuracy is  0.686137931034
24
train accuracy on epoch  24  is  0.831
current average testing accuracy is  0.6877
25
train accuracy on epoch  25  is  0.849
current average testing accuracy is  0.689419354839
26
train accuracy on epoch  26  is  0.858
current average testing accuracy is  0.69090625
27
train accuracy on epoch  27  is  0.868
current average testing accuracy is  0.693029411765
28
train accuracy on epoch  28  is  0.852
current average testing accuracy is  0.694285714286
29
train accuracy on epoch  29  is  0.847
current average testing accuracy is  0.695444444444

final average train accuracy is  0.7781
average test accuracy is  0.695444444444










PART B
0
train accuracy on epoch  0  is  0.503181818182
current average testing accuracy is  0.503
1
train accuracy on epoch  1  is  0.518181818182
current average testing accuracy is  0.5048
2
train accuracy on epoch  2  is  0.546363636364
current average testing accuracy is  0.507875
3
train accuracy on epoch  3  is  0.603636363636
current average testing accuracy is  0.512818181818
4
train accuracy on epoch  4  is  0.646818181818
current average testing accuracy is  0.5185
5
train accuracy on epoch  5  is  0.696818181818
current average testing accuracy is  0.521625
6
train accuracy on epoch  6  is  0.75
current average testing accuracy is  0.527
7
train accuracy on epoch  7  is  0.793636363636
current average testing accuracy is  0.531363636364
8
train accuracy on epoch  8  is  0.838636363636
current average testing accuracy is  0.53628
9
train accuracy on epoch  9  is  0.870454545455
current average testing accuracy is  0.541714285714
10
train accuracy on epoch  10  is  0.884090909091
current average testing accuracy is  0.546516129032
11
train accuracy on epoch  11  is  0.902272727273
current average testing accuracy is  0.55
12
train accuracy on epoch  12  is  0.917727272727
current average testing accuracy is  0.554166666667
13
train accuracy on epoch  13  is  0.918636363636
current average testing accuracy is  0.558743589744
14
train accuracy on epoch  14  is  0.919090909091
current average testing accuracy is  0.56230952381
15
train accuracy on epoch  15  is  0.925454545455
current average testing accuracy is  0.566866666667
16
train accuracy on epoch  16  is  0.917727272727
current average testing accuracy is  0.569468085106
17
train accuracy on epoch  17  is  0.914090909091
current average testing accuracy is  0.57322
18
train accuracy on epoch  18  is  0.918181818182
current average testing accuracy is  0.576396226415
19
train accuracy on epoch  19  is  0.913636363636
current average testing accuracy is  0.579142857143
20
train accuracy on epoch  20  is  0.908636363636
current average testing accuracy is  0.582169491525
21
train accuracy on epoch  21  is  0.904545454545
current average testing accuracy is  0.584532258065
22
train accuracy on epoch  22  is  0.893181818182
current average testing accuracy is  0.585671875
23
train accuracy on epoch  23  is  0.891363636364
current average testing accuracy is  0.587641791045
24
train accuracy on epoch  24  is  0.891818181818
current average testing accuracy is  0.589557142857
25
train accuracy on epoch  25  is  0.893636363636
current average testing accuracy is  0.591342465753
26
train accuracy on epoch  26  is  0.879090909091
current average testing accuracy is  0.593
27
train accuracy on epoch  27  is  0.886818181818
current average testing accuracy is  0.594064102564
28
train accuracy on epoch  28  is  0.881363636364
current average testing accuracy is  0.595765432099
29
train accuracy on epoch  29  is  0.890909090909
current average testing accuracy is  0.597226190476
total train correct =  54604
total train  =  66000
total test correct =  50167
total test  =  84000
average train accuracy is  0.827333333333
average test accuracy is  0.597226190476
3384
84
0



augmenting
0
train accuracy on epoch  0  is  0.502
current average testing accuracy is  0.505
1
train accuracy on epoch  1  is  0.509
current average testing accuracy is  0.504
2
train accuracy on epoch  2  is  0.518
current average testing accuracy is  0.504666666667
3
train accuracy on epoch  3  is  0.527
current average testing accuracy is  0.505
4
train accuracy on epoch  4  is  0.527
current average testing accuracy is  0.506
5
train accuracy on epoch  5  is  0.522
current average testing accuracy is  0.507428571429
6
train accuracy on epoch  6  is  0.539
current average testing accuracy is  0.509875
7
train accuracy on epoch  7  is  0.568
current average testing accuracy is  0.5142
8
train accuracy on epoch  8  is  0.58
current average testing accuracy is  0.516545454545
9
train accuracy on epoch  9  is  0.582
current average testing accuracy is  0.518416666667
10
train accuracy on epoch  10  is  0.59
current average testing accuracy is  0.521
11
train accuracy on epoch  11  is  0.604
current average testing accuracy is  0.522666666667
12
train accuracy on epoch  12  is  0.61
current average testing accuracy is  0.5235625
13
train accuracy on epoch  13  is  0.62
current average testing accuracy is  0.525352941176
14
train accuracy on epoch  14  is  0.634
current average testing accuracy is  0.528263157895
15
train accuracy on epoch  15  is  0.618
current average testing accuracy is  0.52995
16
train accuracy on epoch  16  is  0.632
current average testing accuracy is  0.53119047619
17
train accuracy on epoch  17  is  0.653
current average testing accuracy is  0.534260869565
18
train accuracy on epoch  18  is  0.623
current average testing accuracy is  0.535375
19
train accuracy on epoch  19  is  0.653
current average testing accuracy is  0.53668
20
train accuracy on epoch  20  is  0.67
current average testing accuracy is  0.538192307692
21
train accuracy on epoch  21  is  0.654
current average testing accuracy is  0.540714285714
22
train accuracy on epoch  22  is  0.63
current average testing accuracy is  0.541793103448
23
train accuracy on epoch  23  is  0.667
current average testing accuracy is  0.5433
24
train accuracy on epoch  24  is  0.685
current average testing accuracy is  0.54659375
25
train accuracy on epoch  25  is  0.663
current average testing accuracy is  0.54796969697
26
train accuracy on epoch  26  is  0.655
current average testing accuracy is  0.549382352941
27
train accuracy on epoch  27  is  0.634
current average testing accuracy is  0.550714285714
28
train accuracy on epoch  28  is  0.656
current average testing accuracy is  0.55327027027
29
train accuracy on epoch  29  is  0.657
current average testing accuracy is  0.554394736842
total train correct =  18182
total train  =  30000
total test correct =  21067
total test  =  38000
average train accuracy is  0.606066666667
average test accuracy is  0.554394736842
1538
38
0