def main():
"""Generate image tiles corresponding to the predictions."""
with open(CACHE_PATH + 'raster_data_paths.pickle', 'r') as infile:
raster_data_paths = pickle.load(infile)
with open(CACHE_PATH + METADATA_PATH, 'r') as infile:
training_info = pickle.load(infile)
with open(CACHE_PATH + 'model_metadata.pickle', 'r') as infile:
model_info = pickle.load(infile)
model = load_model(model_info['neural_net_type'], model_info['tile_size'],
len(model_info['bands']))
bands = training_info['bands']
for path in raster_data_paths:
labels, images = load_all_training_tiles(path, bands)
path_parts = path.split('/')
filename = path_parts[len(path_parts) - 1]
wp_preds = list_findings_3(labels, images, model)
start_xs = [label[1] for label in labels]
start_ys = [label[2] for label in labels]
x_values = numpy.unique(start_xs).size
y_values = numpy.unique(start_ys).size
img = numpy.reshape(wp_preds,
(x_values, y_values)) # wrong orientation!
im_uint8 = Image.fromarray(img * 255).convert('RGB')
png_file = os.path.splitext(filename)[0] + "_tf_pred.png"
rotate_and_flip(im_uint8).save(os.path.join(tf_preds_dir, png_file))
raster_dataset, bands_data = read_naip(path, bands) # expensive!
raster_prj = raster_dataset.GetProjection()
epsg = osr.SpatialReference(wkt=raster_prj).GetAttrValue(
'AUTHORITY', 1)
drv = gdal.GetDriverByName("GTiff")
geotiff_file = os.path.splitext(filename)[0] + "_tf_pred.tif"
ds = drv.Create(os.path.join(tf_preds_dir, geotiff_file), x_values,
y_values, 1, gdal.GDT_Float32)
geo_transform = raster_dataset.GetGeoTransform()
ds.SetGeoTransform((geo_transform[0] + NAIP_PIXEL_BUFFER, 64.0, 0.0,
geo_transform[3] - NAIP_PIXEL_BUFFER, 0.0, -64.0))
srs = osr.SpatialReference()
srs.ImportFromEPSG(int(epsg))
ds.SetProjection(srs.ExportToWkt())
im_flt32 = Image.fromarray(img)
ds.GetRasterBand(1).WriteArray(numpy.asarray(
rotate_and_flip(im_flt32)))
ds.FlushCache()
del ds
def main():
"""Post test results to an S3 bucket."""
with open(CACHE_PATH + 'raster_data_paths.pickle', 'r') as infile:
raster_data_paths = pickle.load(infile)
with open(CACHE_PATH + METADATA_PATH, 'r') as infile:
training_info = pickle.load(infile)
with open(CACHE_PATH + MODEL_METADATA_PATH, 'r') as infile:
model_info = pickle.load(infile)
model = load_model(model_info['neural_net_type'], model_info['tile_size'],
len(model_info['bands']))
post_findings_to_s3(raster_data_paths, model, training_info, model_info['bands'], False)
def main():
"""Do each state one by one."""
# randomize_naips = False
naip_year = 2013
naip_states = {'de': ['http://download.geofabrik.de/north-america/us/delaware-latest.osm.pbf'],
'me': ['http://download.geofabrik.de/north-america/us/maine-latest.osm.pbf']}
number_of_naips = 25
extract_type = 'highway'
bands = [1, 1, 1, 1]
tile_size = 64
pixels_to_fatten_roads = 3
tile_overlap = 1
neural_net = 'one_layer_relu_conv'
number_of_epochs = 25
for state, filenames in naip_states:
naiper = NAIPDownloader(number_of_naips, state, naip_year)
raster_data_paths = naiper.download_naips()
cache_paths(raster_data_paths)
create_tiled_training_data(raster_data_paths, extract_type, bands, tile_size,
pixels_to_fatten_roads, filenames,
tile_overlap, state)
with open(CACHE_PATH + METADATA_PATH, 'r') as infile:
training_info = pickle.load(infile)
test_images, model = train_on_cached_data(raster_data_paths, neural_net,
training_info['bands'], training_info['tile_size'],
number_of_epochs)
with open(CACHE_PATH + MODEL_METADATA_PATH, 'r') as infile:
model_info = pickle.load(infile)
model = load_model(model_info['neural_net_type'], model_info['tile_size'],
len(model_info['bands']))
post_findings_to_s3(raster_data_paths, model, training_info)