Support for recursive parameter assignment on module about skorch HOT 8 OPEN

class Seq2Seq:
    def __init__(self, encoder, decoder, **kwargs):
        self.encoder = encoder(skorch.utils.params_for('encoder', kwargs))
        self.decoder = decoder(skorch.utils.params_for('decoder', kwargs))

ef = NeuralNet(
        module=Seq2Seq(encoder=AttentionEncoderRNN, decoder=DecoderRNN),
        module__encoder__num_hidden=23,
    )

from operator import methodcaller

def set_params_in_module(module, **kwargs):
    for k, v in kwargs.items():
        set_param_in_module(module, k, v)

def set_param_in_module(module, param, value):
    name, key = param.rsplit('__', 1)
    name = name.replace('__', '.')
    for n, p in module.named_parameters():
        if n.startswith(name):
            methodcaller(f'{key}_', value)(p)

class Encoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.embeddings = nn.Embedding(10, 10)

class MyModule(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(10, 1, 1)
        self.encoder = Encoder()

module = MyModule()
set_params_in_module(module, conv__weight__requires_grad=False)
set_params_in_module(module, conv__weight__copy=torch.ones((1, 10, 1, 1)))
set_params_in_module(module, conv__weight__add=torch.ones((1, 10, 1, 1)))
set_params_in_module(module, encoder__requires_grad=False)
set_params_in_module(module, 
                     conv__requires_grad=False,
                     encoder__embeddings__weight__requires_grad=False)