search

Borda / keras-yolo3

A Keras implementation of YOLOv3 (Tensorflow backend) a successor of qqwweee/keras-yolo3
MIT License
31 stars 10 forks source link

inconsistency of exported trained models #5

Closed Borda closed 5 years ago

Borda commented 5 years ago

While training model completely random weights from the beginning and exporting just weights https://github.com/Borda/keras-yolo3/blob/b46e258ae35cd0021c85fe9392fb7c7ecbe6dedb/scripts/train.py#L112 and complete model https://github.com/Borda/keras-yolo3/blob/b46e258ae35cd0021c85fe9392fb7c7ecbe6dedb/scripts/train.py#L121 the model does not predict the same detections.

For both predistions, I have used the same predict.py script. Overall the model from weights after a short training on VOC dataset is able to predict some people on sample volleyball video compare to fully exported model predicts nothing...

Borda commented 5 years ago

using simple save and load_model got similar error to https://github.com/qqwweee/keras-yolo3/issues/48

Borda commented 5 years ago

dropping this exporting since it would require some deeper conversion... probably reuse model creation from convert_weights.py but the still the YOLO class us used co there is no significant gain from going deeper into this feature... see c3339b55d2654faff30f2b802b3dde3d9f41690a