Model with 80% accuracy 17 classes

[thenomemac]

here is the confusion matrix pretty insightful:

[thenomemac]

another random draw from the same 17*20 images giving a slightly different view:

[thenomemac]

from both of the CM's the misses are not very suprising. In the case of nearly all misses they are both fish that "A" look similar. And "B" are often miss cross classified in the "training" data with substantial data leakage.

[thenomemac]

Okay, you should use data set "trainfilter" "testfilter" when modeling from here on out for best results. Note, I have to rename all of the images in "trainfilter" s3 folder since all the images have an extra number appended to the front of the file name that'll mess up our data pipeline.

[thenomemac]

For the purposes of documentation here's an 86% accurate InceptionV3 net. Wonder the correlation between this CM and resnet to understand ensemble potential.

[[ 0.45833333  0.          0.          0.          0.          0.44444444
   0.05555556  0.02777778  0.          0.          0.01388889  0.          0.
   0.          0.          0.          0.        ]
 [ 0.          0.90697674  0.          0.          0.          0.          0.
   0.          0.02325581  0.          0.          0.          0.          0.
   0.          0.06976744  0.        ]
 [ 0.          0.          0.86585366  0.          0.          0.          0.
   0.          0.          0.          0.          0.09756098  0.          0.
   0.          0.03658537  0.        ]
 [ 0.          0.          0.          0.6875      0.          0.          0.
   0.          0.275       0.          0.          0.          0.          0.
   0.0375      0.          0.        ]
 [ 0.          0.          0.          0.          0.88421053  0.          0.
   0.          0.          0.          0.          0.11578947  0.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.93103448
   0.          0.          0.          0.          0.          0.
   0.06896552  0.          0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.04819277
   0.95180723  0.          0.          0.          0.          0.          0.
   0.          0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   1.          0.          0.          0.          0.          0.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          1.          0.          0.          0.          0.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.
   0.03896104  0.          0.          0.96103896  0.          0.          0.
   0.          0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          0.          0.          1.          0.          0.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.05952381  0.          0.08333333  0.          0.
   0.          0.          0.02380952  0.          0.83333333  0.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          0.          0.          0.          0.          1.          0.
   0.          0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          0.          0.          0.          0.          0.
   0.73333333  0.          0.26666667  0.        ]
 [ 0.          0.43661972  0.          0.          0.          0.          0.
   0.          0.          0.          0.          0.          0.          0.
   0.56338028  0.          0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          0.          0.          0.          0.02941176  0.
   0.07352941  0.          0.89705882  0.        ]
 [ 0.          0.          0.          0.          0.          0.          0.
   0.          0.          0.          0.          0.          0.          0.
   0.          0.          1.        ]]

[matthew-sochor-zz]

black crappie/white crappie confusions is most likely due to mislabeling, I had a very hard time going through the white crappies but tried my damndest to keep out black crappies. This is pretty awesome

[thenomemac]

Long story short with as bad as our data is this is way better than I could ever do competing against the machine.

[thenomemac]

Think our "TOP-2" accuracy would be very high

[matthew-sochor-zz]

also the black bullhead/channel catfish confusion is totally understandable. Both cat fish. Both look the damn same to me

[matthew-sochor-zz]

Closing this as we have essentially settled on the 17 fish model.

In the future I'd love to see a performance analysis between resnet and inception. If they are the same acc and inception is faster we need to dig this back up!