Initial weight for retraining?

[Yilin Yang]

What is the initial weight for retraining? Is that the weight in amp file?

[Muammar El Khatib]

Depending on what you are referring to:

1) If you are using the retries keyword argument, then randomize() is called with weights=True and scalings=True which will make your optimization to restart in another point of the loss function space.

2) If you would like to retrain, then there is this information on https://amp.readthedocs.io/en/latest/useamp.html?highlight=amp#re-training. There you open an .amp calculator file and use those weights as your starting point in the loss function to minimize it further.

Does that make sense?

[Yilin Yang]

you mean if I load the existing amp file to retrain a set of images, the starting point will be the weight recorded in this amp file?

[Muammar El Khatib]

Yeap. Look at this line: https://bitbucket.org/andrewpeterson/amp/src/3c78aff4591ba125e8da22e81919a783a6694068/amp/model/neuralnetwork.py#lines-205 When you instantiate the NeuralNetwork class from .load(), the weights are present and they are not going to be initialized.

[Yilin Yang]

Thanks, another question is if the fingerprints of images will be recalculated during retraining?

[Muammar El Khatib]

No, if you keep your .ampdb directories in the same location where you performed a previous training. In that case, Amp will try to open neighbor lists, fingerprints, their derivatives from those .ampdb databases instead of computing them again. Just be careful to not mix those databases!.

[Yilin Yang]

I see, thanks very much.

[andrew_peterson]

Not a bug...