2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Save predictions directly before observations are grouped

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-09-14 13:43:28 +02:00
parent 6a990b204a
commit 8230488031
1 changed files with 11 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/twomartens/masterthesis/ssd.py
View File

@ -322,17 +322,23 @@ def _predict_loop(generator: Generator,
saved_images_prediction = True
if use_bayesian:
decoded_predictions = callables.decode_func_dropout(predictions)
callables.save_func(decoded_predictions, original_labels, filenames,
batch_nr=batch_counter,
suffix="_prediction",
save_labels=False)
observations = callables.get_observations_func(decoded_predictions)
callables.save_func(observations, original_labels, filenames,
batch_nr=batch_counter,
suffix="_prediction")
suffix="_observation",
save_labels=False)
for entropy_threshold in entropy_thresholds:
if use_bayesian:
decoded_predictions = callables.apply_entropy_threshold_func(observations,
entropy_threshold=entropy_threshold)
callables.save_func(decoded_predictions, original_labels, filenames,
batch_nr=batch_counter, entropy_threshold=entropy_threshold,
suffix="_entropy")
suffix="_entropy",
save_labels=False)
decoded_predictions = callables.apply_top_k_func(decoded_predictions)
else:
decoded_predictions = callables.decode_func(predictions, entropy_threshold=entropy_threshold)
@ -483,6 +489,7 @@ def _save_predictions(transformed_predictions: Union[np.ndarray, Sequence[np.nda
original_labels: np.ndarray, filenames: Sequence[str],
output_file: str, label_output_file: str,
batch_nr: int, nr_digits: int, suffix: str,
save_labels: Optional[bool] = True,
entropy_threshold: Optional[float] = 0) -> None:
counter_str = str(batch_nr).zfill(nr_digits)
@ -492,7 +499,8 @@ def _save_predictions(transformed_predictions: Union[np.ndarray, Sequence[np.nda
with open(filename, "wb") as file, open(label_filename, "wb") as label_file:
pickle.dump(transformed_predictions, file)
pickle.dump({"labels": original_labels, "filenames": filenames}, label_file)
if save_labels:
pickle.dump({"labels": original_labels, "filenames": filenames}, label_file)
def _predict_save_images(inputs: np.ndarray, predictions: np.ndarray,