Search code examples
pythonmachine-learningtf.kerasyolov4

How to Continue Training from the last acquired best weights?

Keras implementation of YOLOv4

Is it possible in this Keras implementation of YOLOv4 to somehow continue training from the last saved best weights? Something like the following:

model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath=checkpoint_filepath,
    monitor='val_binary_accuracy',
    mode='max',
    save_best_only=True)

model.load_weights(checkpoint_filepath)
Solution

  • According to these lines the repository automatically handles the weights on your path; So to load a pre-trained weights (either .h5 checkpoint or .weights to do transfer learning, and follow training notebooks for the rest;

    model = Yolov4(weight_path='mytraining.weights', 
               class_name_path=class_name_path)

    Update: (from the comments of OP)

    Pre-trained weights can be loaded for transfer learning by passing the path to ".h5" checkpoint file to weight_path argumentwith the following amendment in models.py: replace line 75:

    if load_pretrained and self.weight_path and self.weight_path.endswith('.weights'):

    with:

    if load_pretrained and self.weight_path and (self.weight_path.endswith('.weights') or self.weight_path.endswith('.h5')):

    This issue is addressed in this PR.