TimeXer 분석

Official implementation for "TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables" (NeurIPS 2024)

Timexer는 시계열 작업을 효율적으로 처리하기 위한 경량 라이브러리입니다.
본 글에서는 Timexer의 코드 구조와 동작 원리를 중심으로 리뷰합니다.
데이터는 코드에서 제공한 데이터(NP.csv, PJM.csv, BE.csv,FR.csv, DE.csv)로 진행했습니다.

Model code

class Model(nn.Module):

    def __init__(self, configs):
        super(Model, self).__init__()
        self.task_name = configs.task_name
        self.features = configs.features
        self.seq_len = configs.seq_len
        self.pred_len = configs.pred_len
        self.use_norm = configs.use_norm
        self.patch_len = configs.patch_len
        self.patch_num = int(configs.seq_len // configs.patch_len)
        self.n_vars = 1 if configs.features == 'MS' else configs.enc_in
        # Embedding
        self.en_embedding = EnEmbedding(self.n_vars, configs.d_model, self.patch_len, configs.dropout)

        self.ex_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
                                                   configs.dropout)

        # Encoder-only architecture
        self.encoder = Encoder(
            [
                EncoderLayer(
                    AttentionLayer(
                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
                                      output_attention=False),
                        configs.d_model, configs.n_heads),
                    AttentionLayer(
                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
                                      output_attention=False),
                        configs.d_model, configs.n_heads),
                    configs.d_model,
                    configs.d_ff,
                    dropout=configs.dropout,
                    activation=configs.activation,
                )
                for l in range(configs.e_layers)
            ],
            norm_layer=torch.nn.LayerNorm(configs.d_model)
        )
        self.head_nf = configs.d_model * (self.patch_num + 1)
        self.head = FlattenHead(configs.enc_in, self.head_nf, configs.pred_len,
                                head_dropout=configs.dropout)

    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
        if self.use_norm:
            # Normalization from Non-stationary Transformer
            means = x_enc.mean(1, keepdim=True).detach()
            x_enc = x_enc - means
            stdev = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5)
            x_enc /= stdev

        _, _, N = x_enc.shape

        en_embed, n_vars = self.en_embedding(x_enc[:, :, -1].unsqueeze(-1).permute(0, 2, 1))
        ex_embed = self.ex_embedding(x_enc[:, :, :-1], x_mark_enc)

        enc_out = self.encoder(en_embed, ex_embed)
        enc_out = torch.reshape(
            enc_out, (-1, n_vars, enc_out.shape[-2], enc_out.shape[-1]))
        # z: [bs x nvars x d_model x patch_num]
        enc_out = enc_out.permute(0, 1, 3, 2)

        dec_out = self.head(enc_out)  # z: [bs x nvars x target_window]
        dec_out = dec_out.permute(0, 2, 1)

        if self.use_norm:
            # De-Normalization from Non-stationary Transformer
            dec_out = dec_out * (stdev[:, 0, -1:].unsqueeze(1).repeat(1, self.pred_len, 1))
            dec_out = dec_out + (means[:, 0, -1:].unsqueeze(1).repeat(1, self.pred_len, 1))

        return dec_out