We wanted to check our hypothesis that increasing the augmentation will always give us a better score. Started two models, one with 8 rotations instead of 4 and one turning on random crops as well. These both used a very large proportion of RAM because of the high augmentation factor.



In [1]:

    
cd ..









    



/afs/inf.ed.ac.uk/user/s08/s0805516/repos/neukrill-net-work



In [4]:

    
%run check_test_score.py -v run_settings/alexnet_based_16aug.json









    



Loading settings..
Loading model...
Loading data...
Applying normalisation: global
Finding batch size...
    chosen batch size 3089 for 16 batches
Compiling forward prop...
Making predictions...
    Batch 1 of 16
    Batch 2 of 16
    Batch 3 of 16
    Batch 4 of 16
    Batch 5 of 16
    Batch 6 of 16
    Batch 7 of 16
    Batch 8 of 16
    Batch 9 of 16
    Batch 10 of 16
    Batch 11 of 16
    Batch 12 of 16
    Batch 13 of 16
    Batch 14 of 16
    Batch 15 of 16
    Batch 16 of 16
Collapsing predictions...
Log loss: 0.751659367652



In [5]:

    
%run check_test_score.py -v run_settings/alexnet_based_40aug.json









    



Loading settings..
Loading model...
Loading data...
Applying normalisation: global
Finding batch size...
    chosen batch size 3089 for 40 batches
Compiling forward prop...
Making predictions...
    Batch 1 of 40
    Batch 2 of 40
    Batch 3 of 40
    Batch 4 of 40
    Batch 5 of 40
    Batch 6 of 40
    Batch 7 of 40
    Batch 8 of 40
    Batch 9 of 40
    Batch 10 of 40
    Batch 11 of 40
    Batch 12 of 40
    Batch 13 of 40
    Batch 14 of 40
    Batch 15 of 40
    Batch 16 of 40
    Batch 17 of 40
    Batch 18 of 40
    Batch 19 of 40
    Batch 20 of 40
    Batch 21 of 40
    Batch 22 of 40
    Batch 23 of 40
    Batch 24 of 40
    Batch 25 of 40
    Batch 26 of 40
    Batch 27 of 40
    Batch 28 of 40
    Batch 29 of 40
    Batch 30 of 40
    Batch 31 of 40
    Batch 32 of 40
    Batch 33 of 40
    Batch 34 of 40
    Batch 35 of 40
    Batch 36 of 40
    Batch 37 of 40
    Batch 38 of 40
    Batch 39 of 40
    Batch 40 of 40
Collapsing predictions...
Log loss: 0.739534953247