NOAA

Learning to use Theano by playing with the NOAA whale recognition data (link).

Prerequisites

Ensure all dependencies are met. You will need at least the equivalent of:

• Anaconda (python 2.7)
• OpenCV

Download this repository and mv config-example.py config.py. Edit config.py to match your local directory structure (to replicate results, leave the SEED variable unchanged).

Download the data files from Kaggle and unzip them into /your/base/directory/noaa/data. Make sure to copy w_7489.jpg into /your/base/directory/noaa/data/imgs/.

OpenCV

cd /your/base/directory/noaa/ and run:

$ python -c 'from noaa.cascade.generate_opencv_training_data import main; main()'
$ cd data
# in the following command, change -numPos to match the number of samples you generated
$ opencv_createsamples -info head_examples.info -num 2000 -w 48 -h 48 -vec heads.vec
# in the following command, change -numPos and -numNeg to match the number of samples you generated
$ opencv_traincascade -data heads -vec heads.vec -bg head_backgrounds.info \
    -numPos 1700 -numNeg 8000 -numStages 10 -w 48 -h 48 -featureType HAAR -mode ALL \
    -precalcValBufSize 2048 -precalcIdxBufSize 2048

Note that -numPos in opencv_traincascade should be smaller than -num from opencv_createsamples. According to this thread, the relationship is num >= (numPos + (numStages-1) * (1 - minHitRate) * numPos) + S, where S is the count of examples identified as background images during training. Another comment on the thread suggests using numPos = .85 * num.

Additional steps

...

NNet

The current nnet does not use pre-processing other than shrinking the images. To shrink the images, first edit proc-imgs.sh to match your directory structure. Then cd /your/base/directory/noaa/ and run source noaa/proc-imgs.sh.

Next, load the processed data into memmap arrays. cd /your/base/directory/noaa/ and run:

$ python -c 'from noaa.nnet.preproc import build_memmap_arrays; build_memmap_arrays()'

To train the nnet, first cd /your/base/directory/noaa/, open the python interpreter, and run the following commands:

>>> from noaa.nnet.deep_convolutional_neural_network import fit_lenet
>>> ln = fit_lenet()

Notes

Status: unstable, under active development. If using numpy 1.10.0, you may need to edit Theano's source code per this commit.