def testEndToEndSuccess(self):
logging.info("Using testdata in %s", self.get_temp_dir())
avg_model_dir = self._export_global_average_model()
image_dir = self._write_cmy_dataset()
saved_model_dir = os.path.join(self.get_temp_dir(), "final_saved_model")
saved_model_expected_file = os.path.join(saved_model_dir, "saved_model.pb")
tflite_output_file = os.path.join(self.get_temp_dir(), "final_model.tflite")
labels_output_file = os.path.join(self.get_temp_dir(), "labels.txt")
# Make sure we don't test for pre-existing files.
self.assertFalse(os.path.isfile(saved_model_expected_file))
self.assertFalse(os.path.isfile(tflite_output_file))
self.assertFalse(os.path.isfile(labels_output_file))
with flagsaver.flagsaver(
image_dir=image_dir, tfhub_module=avg_model_dir,
# This dataset is expected to be fit perfectly.
assert_accuracy_at_least=0.9,
saved_model_dir=saved_model_dir,
tflite_output_file=tflite_output_file,
labels_output_file=labels_output_file,
**self.DEFAULT_FLAGS):
make_image_classifier.main([])
# Test that the SavedModel was written.
self.assertTrue(os.path.isfile(saved_model_expected_file))
# Test that the TFLite model works.
labels = self._load_labels(labels_output_file)
lite_model = self._load_lite_model(tflite_output_file)
for class_name, rgb in self.CMY_NAMES_AND_RGB_VALUES:
input_batch = (_fill_image(rgb, self.IMAGE_SIZE)[None, ...]
/ np.array(255., dtype=np.float32))
output_batch = lite_model(input_batch)
prediction = labels[np.argmax(output_batch[0])]
self.assertEqual(class_name, prediction)
def testEndToEndAccuracyFailure(self):
logging.info("Using testdata in %s", self.get_temp_dir())
avg_model_dir = self._export_global_average_model()
image_dir = self._write_random_dataset()
with flagsaver.flagsaver(
image_dir=image_dir, tfhub_module=avg_model_dir,
# This is expeced to fail for this random dataset.
assert_accuracy_at_least=0.8, **self.DEFAULT_FLAGS):
with self.assertRaisesRegex(AssertionError, "ACCURACY FAILED"):
make_image_classifier.main([])
def testEndToEndAccuracyFailure(self, use_tf_data_input):
if use_tf_data_input and (LooseVersion(tf.__version__) <
LooseVersion("2.5.0")):
return
logging.info("Using testdata in %s", self.get_temp_dir())
avg_model_dir = self._export_global_average_model()
image_dir = self._write_random_dataset()
with flagsaver.flagsaver(
image_dir=image_dir,
tfhub_module=avg_model_dir,
# This is expected to fail for this random dataset.
assert_accuracy_at_least=0.9,
use_tf_data_input=use_tf_data_input,
**self.DEFAULT_FLAGS):
with self.assertRaisesRegex(AssertionError, "ACCURACY FAILED"):
make_image_classifier.main([])
def testEndToEndSuccess(self):
logging.info("Using testdata in %s", self.get_temp_dir())
avg_model_dir = self._export_global_average_model()
image_dir = self._write_cmy_dataset()
saved_model_dir = os.path.join(self.get_temp_dir(), "final_saved_model")
self.assertFalse( # Make sure we don't test for pre-existing files.
os.path.isfile(os.path.join(saved_model_dir, "saved_model.pb")))
with flagsaver.flagsaver(
image_dir=image_dir, tfhub_module=avg_model_dir,
# This dataset is expected to be fit perfectly.
assert_accuracy_at_least=0.9,
saved_model_dir=saved_model_dir, **self.DEFAULT_FLAGS):
make_image_classifier.main([])
# Test for main output artifact.
self.assertTrue(
os.path.isfile(os.path.join(saved_model_dir, "saved_model.pb")))