[<Library component: Model|Core|etc...>] Adding Multivariate Model.

[skmanzg]

What happened + What you expected to happen

I am trying to change(?) LSTM model into multivariate version based on multivariate correlation + autocorrelation in my local enviroment.

so, I tried to change like this:

1. use BaseMultivariate instead of recurrent thing
   class LSTM(BaseMultivariate):

2. set class attributes

    # Class attributes
    SAMPLING_TYPE = "multivariate"
    EXOGENOUS_FUTR = True
    EXOGENOUS_HIST = True
    EXOGENOUS_STAT = True

3. add 'n_series' in init

    def __init__(
        self,
        h: int,
        input_size,
        n_series,
        inference_input_size: int = -1,
        valid_batch_size: Optional[int] = None,
...
...
)


super(LSTM, self).__init__(
            h=h,
            input_size=input_size,
            n_series=n_series,
            inference_input_size=inference_input_size,
            valid_batch_size=valid_batch_size,
...
..


)

        self.n_series = n_series

4. add multivariate feature i intended as a new class.

        self.cross_correlation = CrossCorrelationLayer(
            n_series, 
            self.encoder_hidden_size
        )
        self.temporal_attention = TemporalAttention(
            self.encoder_hidden_size
        )


class CrossCorrelationLayer(nn.Module):
    def __init__(self, n_series, hidden_size):
        super().__init__()
        self.correlation_weights = nn.Parameter(
            torch.randn(n_series, n_series)
        )
        self.hidden_transform = nn.Linear(hidden_size * 2, hidden_size)
    
    def forward(self, x):
        # x shape: [batch, seq_len, n_series]
        corr_weighted = torch.matmul(x, self.correlation_weights)
        combined = torch.cat([x, corr_weighted], dim=-1)



class TemporalAttention(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.attention = nn.MultiheadAttention(
            hidden_size, 
            num_heads=8
        )
    
    def forward(self, x):
        # x shape: [batch, seq_len, hidden_size]
        attn_output, _ = self.attention(x, x, x)
        return attn_output
    

5. change input_encoder

        # LSTM input size (1 for target variable y)
        input_encoder = n_series + self.hist_exog_size + self.stat_exog_size

6. fix the forward function accordingly:

def forward(self, windows_batch):

        # Parse windows_batch
        encoder_input = windows_batch["insample_y"]  # [B, seq_len, n_series]
        futr_exog = windows_batch["futr_exog"]
        hist_exog = windows_batch["hist_exog"]
        stat_exog = windows_batch["stat_exog"]
        print('encoder_input:', encoder_input.shape)  # for debug



..

..

        # Cross-correlation 
        corr_features = self.cross_correlation(encoder_input)
        
        # LSTM encoding
        hidden_state, _ = self.hist_encoder(corr_features)
        
        # Apply temporal attention
        hidden_state = self.temporal_attention(hidden_state)
        

        print('hidden_state:',hidden_state.shape)
        # Context adapter
        context = self.context_adapter(hidden_state)
        context = context.reshape(batch_size, seq_len, self.h, self.context_size)
        print('context:',context.shape)
        # Residual connection with futr_exog
        if self.futr_exog_size > 0:
            context = torch.cat((context, futr_exog), dim=-1)

        # Final forecast
        output = self.mlp_decoder(context)
        output = self.loss.domain_map(output)
        
        print(output.shape)

        return output

7. I set the hyperparameters for five features time series.

model = LSTM(h=288, 
            input_size=1440,
            n_series=5,
            loss=MAE(),
            scaler_type='minmax',
            encoder_n_layers=5,
            encoder_hidden_size=128,
            context_size=10,
            decoder_hidden_size=128,
            decoder_layers=5,
            max_steps=500,
                 )

nf = NeuralForecast(models=[model], freq='10min') 

and the, when I try to do 'nf.fit(df=Y_train_df), the error occurred.

RuntimeError: input.size(-1) must be equal to input_size. Expected 5, got 2

This is mainly because it accepts 2 features instead of 5 features when I print encoder_input
encoder_input: encoder_input:encoder_input:torch.Size([32, 1440, 2])

I do not know why the size became 2 when I set n_series = 5. This is really frustrating tbh.

In addition,

            inference_input_size=inference_input_size,
            valid_batch_size=valid_batch_size,

These also cause problem with TypeError: Trainer.__init__() got an unexpected keyword argument. so I had to remove those two when I run the code.

Those errors seems very common when I try to do it with different model, ModernTCN, not to mention the guide document is very limited.

Versions / Dependencies

python 3.10.14
reinstalled neuralforecast today.

Reproduction script

This was already described above

Issue Severity

High: It blocks me from completing my task.