MLP Changes

[tedmoore]

In MLP.py, this could be a dict lookup (which is faster and easier on the eyes) than a bunch of elifs:

MMMAudio/mmm_audio/MLP.py

Lines 71 to 83 in 227da52

	for i, vals in enumerate(layers):
	print(f"Layer {i}: {vals}")
	val, activation = vals
	if activation is not None:
	if activation == 'relu':
	activation = nn.ReLU()
	elif activation == 'sigmoid':
	activation = nn.Sigmoid()
	elif activation == 'tanh':
	activation = nn.Tanh()
	else:
	raise ValueError("Activation function not recognized.")
	layers[i] = [val, activation]

[spluta]

fine by me.

[tedmoore]

Also, the train_nn function creates a new MLP each time it is called. Instead of this, what would be more flexible is to pass an MLP to the function with some data and have it update the weights of the passed model. It might also make sense to make train_nn a method of the MLP class.

[spluta]

I could totally be wrong, but I think you are overthinking this. It makes a new MLP, trains it, then saves the training file. Then you load the new training into the MLP. The temporary MLP just gets garbage collected.

[tedmoore]

The way it works now is perfect for the specific use-case where you have a small amount of data. It will definitely converge to something and you basically want to overfit every time. When using it to control parameters of a synth like this, it's super straight forward. But it's not flexible enough to accomodate other use cases. If this were in the examples folder it'd be ok, but here it is presented as the MLP interface for MMMAudio. Some users are (hopefully) going to have other uses in mind!

[spluta]

i see what you are saying. maybe the fn train_nn just needs a different name?

[tedmoore]

ideas:

make_and_train_nn
new_training
train_new