[Question] Setting up a classifier with params

[jpbarddal]

Hi,

I am interested in deep-river. Is the project still being incremented?
Also, I am attempting to pass a module with specific params. A possible MWE would be as follows, however, it seems the Classifier class requires as input a non-instantiated nn.Module.
Am I missing something?

MLP code:

class MLP(nn.Module):

    def __init__(self,
                 input_dim: int = 512,
                 hidden_dim: int = 256,
                 output_dim: int = 2):
        super(MLP, self).__init__()
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.output_dim = output_dim

        self.linear = nn.Linear(self.input_dim, self.hidden_dim)
        self.activation = nn.ReLU()
        self.linear_2 = nn.Linear(self.hidden_dim, self.output_dim)

    def forward(self, x):
        out = self.linear(x)
        out = self.activation(out)
        out = self.linear_2(out)
        return out

Pipeline construction:

    model = MLP(input_dim=dataset.n_features - 1,
                hidden_dim=(dataset.n_features - 1) // 2,
                output_dim=2)
    model_pipeline = compose.Pipeline(
        preprocessing.StandardScaler(),
        classification.Classifier(module=model,
                                  loss_fn='binary_cross_entropy',
                                  optimizer_fn='adam',
                                  )
    )

[kulbachcedric]

Hi @jpbarddal,
this project is still being incremented!
You are right the classifier requires a non-instantiated nn.Module.
The reason for that is that in data streams the number of classes to predict or the number of features are might not be known at the beginning of the data stream. For that reason we initialize the model within the first event of the data stream.

However, to pass specific params to the Module, you can still do that by adding it to the Classifiers arguments:

    class MLP(nn.Module):

    def __init__(self,
                 n_features: int = 512,
                 hidden_dim: int = 256,
                 output_dim: int = 2):
        super(MLP, self).__init__()
        self.input_dim = n_features
        self.hidden_dim = hidden_dim
        self.output_dim = output_dim

        self.linear = nn.Linear(self.input_dim, self.hidden_dim)
        self.activation = nn.ReLU()
        self.linear_2 = nn.Linear(self.hidden_dim, self.output_dim)

    def forward(self, x):
        out = self.linear(x)
        out = self.activation(out)
        out = self.linear_2(out)
        return out


    model_pipeline = compose.Pipeline(
        preprocessing.StandardScaler(),
        classification.Classifier(module=MLP,
                                  loss_fn='binary_cross_entropy',
                                  optimizer_fn='adam',
                                  hidden_dim=(dataset.n_features - 1) // 2,
                                  output_dim=2
                                  )
    )

Note, that I removed the input_dim argument and added the n_features field as this variable is set with the first event of the data stream.

Kind regards
Cedric