Code used for competing in National Data Science Bowl (http://www.kaggle.com/c/datasciencebowl). The final solution used Convolutional Neural Networks.
Dependices are found in requirements.txt. I created a even folded where these libraries are stored.
create Data_converted/train/ by running python gen_train.py
create Data_converted/test/ by running python gen_test.py
Within Fish Bowl.ipynb run steps 1 & 2 to create the training and testing set needed as inputs to the CNNs.
We can perform unsupervised pre-training on the network by running the exact network used for training BUT we use regression and the labels are the same as the input features. The pre-training is ran with the full test + train set and the weights of the network are saved into a pickled object. These weights are then used to initialize the true training network.
python2.7 run_autoencoder.py fit
THIS MADE NO IMPROVEMENT SO WILL NOT CONTINUE TO DO IT
To train the best single mode, run:
python2.7 run_analysis.py fit
This will create a pickled object net-specialists.pickle which contains the neccessary weights to create predictions.
To generate predictions which are averaged across multiple transformations of the input, run:
python2.7 run_predict.py predict
This will create multiple a csv file with predictions for each test set observation
To generate predictions for a single model run step 3 within Fish Bowl.ipynb
To generate predictions for a multiple models averaged together run step 4 within Fish Bowl.ipynb
In the end I trained 8 different models, to train and predict all of these at once run ./run_all.sh and ./predict_all.sh
Throughout the competition I had repeated issues with data augmentation, I was only able to achieve good results with the rotation of [0,90,180,270].
The background in the images is white (255) and opencv/scikit-image assume by default that it is black (0). Therefore, we can invert the images with
im = np.invert(im) when loading in the images.