Download the training data, and training patch list. Upacking files like this structure:
+data
+training
+HG
+LG
+training_list
-trainval-balanced.txt
-trainval.txt
For each row in training list, it gives the sampleID, index and label for the indexed pixel.
#ID x y z label
HG/0005 70 95 128 0
HG/0004 77 117 137 0
...
You can construct training batches (data pathch and labels) according to this list within your own data_loader.
Use create_h5.py to generate hdf5 file. Run "python create_h5.py -h" for usage.
python ./create_h5.py --data_dir=/path/to/data --output_path=/path/to/h5_file
How to use:
import h5py
f = h5py.File('training.h5','r')
img_patch = f['HG/0001'][:, x-16:x+16+1, y-16:y+16+1, z] #sample a 33x33 patch