Is there anthing equivalent to load and save model in ocaml torch about ocaml-torch HOT 4 CLOSED

I'm not sure to understand what you mean by "saving and loading a model". Is it about serializing the weights or the full model architecture. Both are somehow supported, to save load tensors you can find a bunch of examples in the codebase, e.g. here for tensor serialization. Models typically store their weights in a Var_store.t and these get serialized through the same api, e.g. here. Finally see this tutorial for the full model via the jit layer.

For converting list/array to tensors, there are a bunch function doing so, maybe have a look at the tests in this file to get a feeling of what is available.

from ocaml-torch.

This what exactly what I meant. But for the saving and loading a model. Your example shows you loading a python torch trained model. My project has to be done solely in ocaml. Is this library capable of saving a model and loading a model using only ocaml?

I guess what I am looking for It pretty well demonstrated by this tensorflow (I know we are using torch;) )

https://keras.io/guides/serialization_and_saving/

from ocaml-torch.

You can save all the tensor values saved in a Var_store.t and load them, all this via the Serialized api (see the yolo example from my previous message for loading the weights, writing them is similar). Note that there is no way to serialize the model architecture so the yolo code generates the model then load the weights - but if your code is fully in ocaml this should not be an issue.

from ocaml-torch.

Thank you for your very detailed responses. Using your responses I was able to do everything in your library that I needed for my project. Hopefully these discussions will be able to help others.

from ocaml-torch.