Scripts/notebooks for The Nature Conservancy's fish classification competition

• Current best score before blending: 0.886 log loss on LB from ensemble_model.ipynb using vgg16bn architecture.
• Current best score after blending: 0.822 log loss on LB from blending ensemble_model.ipynb results with conv_model.ipynb results

Fits a group of CNNs of (optionally) varying architectures on the image data; predicts on augmented test data

Computes and saves the convolutional features from a pretrained VGG model w/ batch normalization; uses these features to train a new model on the image data

Uses pretrained PyTorch models with both train-time and test-time augmentation. Have only tested resnet architecture so far, but currently produces third best single model (after "ensemble_model" and "conv_model").

Trains a model to distinguish images that contain fish from those that don't

Similar to ensemble model, but uses the "fish detector" model to segment out the "NoF" class first

Uses annotations to crop training images down to a bounding box surrounding the fish

Trains model on cropped data to predict coordinates of bounding box in test images

Similar to end_to_end but uses the "bb regressor" model to try to crop test images first

Mutli-input CNN using keras functional API to incorporate bounding box coordinates as a feature in the model; predicts both coordinates and class probabibilities

Uses conv features of image data as input to SVM model

Same as above, but using gradient boosting instead of SVM

Uses a sliding window to feed subsets of image into CNN for classification

Transformations that could be useful for preprocessing / image segmentation.

Combines submission files into simple ensemble

• Unable to get any improvement over vanilla CNNs using any of the bounding box, cropping, sliding window, or pre-filtering strategies above. In particular, the bb_end_to_end model scored significantly worse -- my theories include:
  • not enough accuracy from bb regressor to make the stategy work
  • size/aspect ratio mismatches introduced at some point in the pipeline
  • user error in the pipeline somewhere (ie filenames not aligned with predictions, etc)
• Relabeling the dataset also didn't lead to improved performance using these methods
• Unable to get comparable performance from Resnet and Inception models. Might just be an issue of my implementation

(credit to Jeremy Howard, Pai Peng, Naive Shuai, Craig Glastonbury, and others for portions of this code)