Inference

Having a trained model, we can now process the new data. The main feature here is that the processing is carried out by sequential sampling of the image patches as we cannot read the whole image at once. The pathches overlap each other to avoid the boundary effects as possible.

from keras.models import load_model
from aeronet.dataset import Predictor

# configuration
INPUT_BC = '/path/to/test/element/'
channels = ['RED', 'GRN', 'BLU']
labels = ['100']

# Load the model. Keras is for example, you can use any
model = load_model('path/to/model/file.h5', compile=False)

# Make a prediction function that processes a BandSample
def processing_fn(sample):
    # Extracting the data from BandSample
    x = sample.numpy().astype(np.float32)

    # Transform the data to fit the model
    x = x.transpose(1,2,0)
    x = np.expand_dims(x, 0)

    # prediction
    y = model.predict(x)

    # Thresholding the output to get a mask
    if threshold is not None:
        y = (y > 0.5).astype(np.uint8)
    return y.squeeze(0).transpose(2,0,1)

# Wrap the function into Predictor
# `bound` means the width of samples overlap
predictor = Predictor(channels,
                  labels,
                  processing_fn=processing_fn,
                  sample_size=(2048,2048),
                  bound=512
                  ))

# Open the imagery and process it
bc = BandCollection(parse_direcotry(INPUT_BC, channels))
bc.process(bc, '/path/to/output/')

# Make polygons
vector_data = polygonize(mask2[0], properties={'class': '100'}