Load the pre-trained neural network model and save the trained model.

[beomyeol]

dolphin DNN has not supported the feature of loading and saving the neural network model.
We discussed this feature in #77. I'd like to continue our discussion such as how to implement this and how to load and save the model.

[dongjoon-hyun]

I think we can implement saving the model by using 'output' service.
For loading, I have no idea at all so far.

[beomyeol]

As we discussed in #77, we can implement this by saving parameter of each layer and loading the saved parameter using parameter initializer.

We can save the parameters in Numpy compatible plain text file format like we are using this format as input matrix format. The layer parameter consists of the weight matrix and the bias vector. These can be put into a single matrix. However, in some case, bias vector may not be used. So, I think we can introduce another delimiter to distinguish weight matrix and bias vector explicitly. (The delimiter must be different to one that is used as the delimiter of the matrix.)

[dongjoon-hyun]

Do you think each layer save and load its own file, e.g. layer1.txt? If we can handle separate files for each layers, why don't we split them into two subfiles? e.g. layer1.weight.txt, layer1.bias.txt? It's just question, @beomyeol .

[beomyeol]

@dongjoon-hyun Thank you for your idea. It will be better if we can handle multiple files. With this, we can rule out the possibility of the problem when the delimiter for matrices is same to the one for parameters. :)