def test_train_test_split(self):
ts = dtr.TrainingSelector(img_filepaths=self.img_filepaths)
X = ts.descr_feature_matrix
self.assertEqual(len(X), len(self.img_filepaths))
# test `method` argument
split_df = ts.train_test_split()
self.assertIn('train', split_df)
self.assertIn('img_cluster', split_df)
split_df = ts.train_test_split(method='cluster-I')
self.assertIn('train', split_df)
self.assertNotIn('img_cluster', split_df)
# test `return_evr` argument (expected variance ratio)
split_df, evr = ts.train_test_split(return_evr=True)
self.assertIsInstance(split_df, pd.DataFrame)
self.assertIsInstance(evr, float)
def test_init(self):
# if providing `img_filepaths`, `img_dir` and `img_filename_pattern`
# are ignored
ts = dtr.TrainingSelector(img_filepaths=self.img_filepaths)
self.assertEqual(
ts.img_filepaths,
dtr.TrainingSelector(img_filepaths=self.img_filepaths,
img_dir='foo').img_filepaths)
self.assertEqual(
ts.img_filepaths,
dtr.TrainingSelector(img_filepaths=self.img_filepaths,
img_filename_pattern='foo').img_filepaths)
self.assertEqual(
ts.img_filepaths,
dtr.TrainingSelector(img_filepaths=self.img_filepaths,
img_dir='foo',
img_filename_pattern='foo').img_filepaths)
# if not providing `img_filepaths`, inexistent `img_dir` or non-tif
# `img_filename_pattern` will result in an empty `img_filepaths`
# attribute
self.assertEqual(
len(dtr.TrainingSelector(img_dir='foo').img_filepaths), 0)
self.assertEqual(
len(
dtr.TrainingSelector(
img_dir=self.img_dir,
img_filename_pattern='foo').img_filepaths), 0)
# otherwise, there should be at least one img
self.assertGreater(
len(dtr.TrainingSelector(img_dir=self.img_dir).img_filepaths), 0)
# even when providing an integer in the `gabor_num_orientations`
# argument, the respective attribute will be a tuple after `__init__`
# is executed
self.assertIsInstance(
dtr.TrainingSelector(
self.img_filepaths,
gabor_num_orientations=8).gabor_num_orientations, tuple)
def train_test_split(ctx, img_filepaths, img_dir, img_filename_pattern,
gabor_frequencies, gabor_num_orientations,
response_bins_per_axis, num_color_bins, method,
num_components, num_img_clusters, train_prop,
output_filepath):
logger = ctx.obj['LOGGER']
ts = dtr.TrainingSelector(img_filepaths=img_filepaths,
img_dir=img_dir,
img_filename_pattern=img_filename_pattern,
gabor_frequencies=gabor_frequencies,
gabor_num_orientations=gabor_num_orientations,
response_bins_per_axis=response_bins_per_axis,
num_color_bins=num_color_bins)
num_imgs = len(ts.img_filepaths)
if img_dir is not None:
logger.info("Loaded %d images from %s", num_imgs, img_dir)
else:
logger.info("Loaded %d images", num_imgs)
tts_kws = {}
if method is not None:
tts_kws['method'] = method
if num_components is not None:
tts_kws['num_components'] = num_components
if num_img_clusters is not None:
tts_kws['num_img_clusters'] = num_img_clusters
if train_prop is not None:
tts_kws['train_prop'] = train_prop
df, evr = ts.train_test_split(return_evr=True, **tts_kws)
logger.info("Variance ratio explained by PCA: %f", evr)
if output_filepath is None:
output_filepath = 'split.csv'
df.to_csv(output_filepath)
logger.info("Dumped train/test split to %s", output_filepath)