v1.0

model.properties

+# 模型编码
+modelCode=612
+# 模型名称
+modelName=itransformer_pytorch
+# 模型描述
+modelDescription=iTransformer能高效利用长程时序特征，能同时预测多个指标。
+# 应用场景
+appScenario=推理,训练,金融,运维,电商,制造,能源,医疗
+# 框架类型
+frameType=pytorch

model/Flashformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Decoder, DecoderLayer, Encoder, EncoderLayer, ConvLayer
+from layers.SelfAttention_Family import FlashAttention, AttentionLayer, FullAttention
+from layers.Embed import DataEmbedding
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        # Embedding
+        self.enc_embedding = DataEmbedding(configs.enc_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        # Encoder
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FlashAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                       output_attention=configs.output_attention), 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)
+        )
+        # Decoder
+        self.dec_embedding = DataEmbedding(configs.dec_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        self.decoder = Decoder(
+            [
+                DecoderLayer(
+                    AttentionLayer(
+                        FullAttention(True, 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.d_layers)
+            ],
+            norm_layer=torch.nn.LayerNorm(configs.d_model),
+            projection=nn.Linear(configs.d_model, configs.c_out, bias=True)
+        )
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.dec_embedding(x_dec, x_mark_dec)
+        dec_out = self.decoder(dec_out, enc_out, x_mask=None, cross_mask=None)
+        return dec_out
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/Flowformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Decoder, DecoderLayer, Encoder, EncoderLayer, ConvLayer
+from layers.SelfAttention_Family import FullAttention, AttentionLayer, FlowAttention
+from layers.Embed import DataEmbedding
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+
+        if configs.channel_independence:
+            self.enc_in = 1
+            self.dec_in = 1
+            self.c_out = 1
+        else:
+            self.enc_in = configs.enc_in
+            self.dec_in = configs.dec_in
+            self.c_out = configs.c_out
+
+        # Embedding
+        self.enc_embedding = DataEmbedding(self.enc_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        # Encoder
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FlowAttention(attention_dropout=configs.dropout), 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)
+        )
+        # Decoder
+        self.dec_embedding = DataEmbedding(self.dec_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        self.decoder = Decoder(
+            [
+                DecoderLayer(
+                    AttentionLayer(
+                        FullAttention(True, 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.d_layers)
+            ],
+            norm_layer=torch.nn.LayerNorm(configs.d_model),
+            projection=nn.Linear(configs.d_model, configs.c_out, bias=True)
+        )
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.dec_embedding(x_dec, x_mark_dec)
+        dec_out = self.decoder(dec_out, enc_out, x_mask=None, cross_mask=None)
+        return dec_out
+
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/Informer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Decoder, DecoderLayer, Encoder, EncoderLayer, ConvLayer
+from layers.SelfAttention_Family import ProbAttention, AttentionLayer
+from layers.Embed import DataEmbedding
+
+
+class Model(nn.Module):
+    """
+    Informer with Propspare attention in O(LlogL) complexity
+    Paper link: https://ojs.aaai.org/index.php/AAAI/article/view/17325/17132
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+
+        self.pred_len = configs.pred_len
+        self.label_len = configs.label_len
+
+        if configs.channel_independence:
+            self.enc_in = 1
+            self.dec_in = 1
+            self.c_out = 1
+        else:
+            self.enc_in = configs.enc_in
+            self.dec_in = configs.dec_in
+            self.c_out = configs.c_out
+
+        # Embedding
+        self.enc_embedding = DataEmbedding(self.enc_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        self.dec_embedding = DataEmbedding(self.dec_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+
+        # Encoder
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        ProbAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=configs.output_attention),
+                        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)
+            ],
+            [
+                ConvLayer(
+                    configs.d_model
+                ) for l in range(configs.e_layers - 1)
+            ] if configs.distil else None,
+            norm_layer=torch.nn.LayerNorm(configs.d_model)
+        )
+        # Decoder
+        self.decoder = Decoder(
+            [
+                DecoderLayer(
+                    AttentionLayer(
+                        ProbAttention(True, configs.factor, attention_dropout=configs.dropout, output_attention=False),
+                        configs.d_model, configs.n_heads),
+                    AttentionLayer(
+                        ProbAttention(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.d_layers)
+            ],
+            norm_layer=torch.nn.LayerNorm(configs.d_model),
+            projection=nn.Linear(configs.d_model, configs.c_out, bias=True)
+        )
+
+
+    def long_forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        dec_out = self.dec_embedding(x_dec, x_mark_dec)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.decoder(dec_out, enc_out, x_mask=None, cross_mask=None)
+
+        return dec_out  # [B, L, D]
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.long_forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/Reformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import ReformerLayer
+from layers.Embed import DataEmbedding
+
+
+class Model(nn.Module):
+    """
+    Reformer with O(LlogL) complexity
+    Paper link: https://openreview.net/forum?id=rkgNKkHtvB
+    """
+
+    def __init__(self, configs, bucket_size=4, n_hashes=4):
+        """
+        bucket_size: int, 
+        n_hashes: int, 
+        """
+        super(Model, self).__init__()
+        self.pred_len = configs.pred_len
+        self.seq_len = configs.seq_len
+
+        if configs.channel_independence:
+            self.enc_in = 1
+            self.dec_in = 1
+            self.c_out = 1
+        else:
+            self.enc_in = configs.enc_in
+            self.dec_in = configs.dec_in
+            self.c_out = configs.c_out
+
+        self.enc_embedding = DataEmbedding(self.enc_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        # Encoder
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    ReformerLayer(None, configs.d_model, configs.n_heads,
+                                  bucket_size=bucket_size, n_hashes=n_hashes),
+                    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.projection = nn.Linear(
+            configs.d_model, configs.c_out, bias=True)
+
+    def long_forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # add placeholder
+        x_enc = torch.cat([x_enc, x_dec[:, -self.pred_len:, :]], dim=1)
+        if x_mark_enc is not None:
+            x_mark_enc = torch.cat(
+                [x_mark_enc, x_mark_dec[:, -self.pred_len:, :]], dim=1)
+
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)  # [B,T,C]
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+        dec_out = self.projection(enc_out)
+
+        return dec_out  # [B, L, D]
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.long_forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/Transformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Decoder, DecoderLayer, Encoder, EncoderLayer, ConvLayer
+from layers.SelfAttention_Family import FullAttention, AttentionLayer
+from layers.Embed import DataEmbedding
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+
+        if configs.channel_independence:
+            self.enc_in = 1
+            self.dec_in = 1
+            self.c_out = 1
+        else:
+            self.enc_in = configs.enc_in
+            self.dec_in = configs.dec_in
+            self.c_out = configs.c_out
+
+        # Embedding
+        self.enc_embedding = DataEmbedding(self.enc_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        # Encoder
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=configs.output_attention), 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)
+        )
+        # Decoder
+        self.dec_embedding = DataEmbedding(self.dec_in, configs.d_model, configs.embed, configs.freq,
+                                           configs.dropout)
+        self.decoder = Decoder(
+            [
+                DecoderLayer(
+                    AttentionLayer(
+                        FullAttention(True, 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.d_layers)
+            ],
+            norm_layer=torch.nn.LayerNorm(configs.d_model),
+            projection=nn.Linear(configs.d_model, configs.c_out, bias=True)
+        )
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.dec_embedding(x_dec, x_mark_dec)
+        dec_out = self.decoder(dec_out, enc_out, x_mask=None, cross_mask=None)
+        return dec_out
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/iFlashformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import FlashAttention, AttentionLayer
+from layers.Embed import DataEmbedding_inverted
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.seq_len = configs.seq_len
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        # Embedding
+        self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
+                                                    configs.dropout)
+        # Encoder-only architecture
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FlashAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                       output_attention=configs.output_attention), 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.projector = nn.Linear(configs.d_model, configs.pred_len, bias=True)
+
+
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # 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
+
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N]
+        # De-Normalization from Non-stationary Transformer
+        dec_out = dec_out * (stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        dec_out = dec_out + (means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        return dec_out
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/iFlowformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import FlowAttention, AttentionLayer
+from layers.Embed import DataEmbedding_inverted
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.seq_len = configs.seq_len
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        # Embedding
+        self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
+                                                    configs.dropout)
+        # Encoder-only architecture
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FlowAttention(attention_dropout=configs.dropout), 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.projector = nn.Linear(configs.d_model, configs.pred_len, bias=True)
+
+
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # 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
+
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N]
+        # De-Normalization from Non-stationary Transformer
+        dec_out = dec_out * (stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        dec_out = dec_out + (means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        return dec_out
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/iInformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import ProbAttention, AttentionLayer
+from layers.Embed import DataEmbedding_inverted
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.seq_len = configs.seq_len
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        # Embedding
+        self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
+                                                    configs.dropout)
+        # Encoder-only architecture
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        ProbAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=configs.output_attention), 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.projector = nn.Linear(configs.d_model, configs.pred_len, bias=True)
+
+
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # 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
+
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N]
+        # De-Normalization from Non-stationary Transformer
+        dec_out = dec_out * (stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        dec_out = dec_out + (means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        return dec_out
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]

model/iReformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import ReformerLayer
+from layers.Embed import DataEmbedding_inverted
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Vanilla Transformer
+    with O(L^2) complexity
+    Paper link: https://proceedings.neurips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.seq_len = configs.seq_len
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        # Embedding
+        self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
+                                                    configs.dropout)
+        # Encoder-only architecture
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    ReformerLayer(None, configs.d_model, configs.n_heads,
+                                  bucket_size=4, n_hashes=4),
+                    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.projector = nn.Linear(configs.d_model, configs.pred_len, bias=True)
+
+
+    def forecast(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
+        # 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
+
+        # Embedding
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N]
+        # De-Normalization from Non-stationary Transformer
+        dec_out = dec_out * (stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        dec_out = dec_out + (means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+        return dec_out
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]
\ No newline at end of file

model/iTransformer.py

+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from layers.Transformer_EncDec import Encoder, EncoderLayer
+from layers.SelfAttention_Family import FullAttention, AttentionLayer
+from layers.Embed import DataEmbedding_inverted
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Paper link: https://arxiv.org/abs/2310.06625
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.seq_len = configs.seq_len
+        self.pred_len = configs.pred_len
+        self.output_attention = configs.output_attention
+        self.use_norm = configs.use_norm
+        # Embedding
+        self.enc_embedding = DataEmbedding_inverted(configs.seq_len, configs.d_model, configs.embed, configs.freq,
+                                                    configs.dropout)
+        self.class_strategy = configs.class_strategy
+        # Encoder-only architecture
+        self.encoder = Encoder(
+            [
+                EncoderLayer(
+                    AttentionLayer(
+                        FullAttention(False, configs.factor, attention_dropout=configs.dropout,
+                                      output_attention=configs.output_attention), 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.projector = nn.Linear(configs.d_model, configs.pred_len, bias=True)
+
+    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 # B L N
+        # B: batch_size;    E: d_model; 
+        # L: seq_len;       S: pred_len;
+        # N: number of variate (tokens), can also includes covariates
+
+        # Embedding
+        # B L N -> B N E                (B L N -> B L E in the vanilla Transformer)
+        enc_out = self.enc_embedding(x_enc, x_mark_enc) # covariates (e.g timestamp) can be also embedded as tokens
+        
+        # B N E -> B N E                (B L E -> B L E in the vanilla Transformer)
+        # the dimensions of embedded time series has been inverted, and then processed by native attn, layernorm and ffn modules
+        enc_out, attns = self.encoder(enc_out, attn_mask=None)
+
+        # B N E -> B N S -> B S N 
+        dec_out = self.projector(enc_out).permute(0, 2, 1)[:, :, :N] # filter the covariates
+
+        if self.use_norm:
+            # De-Normalization from Non-stationary Transformer
+            dec_out = dec_out * (stdev[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+            dec_out = dec_out + (means[:, 0, :].unsqueeze(1).repeat(1, self.pred_len, 1))
+
+        return dec_out
+
+
+    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec, mask=None):
+        dec_out = self.forecast(x_enc, x_mark_enc, x_dec, x_mark_dec)
+        return dec_out[:, -self.pred_len:, :]  # [B, L, D]
\ No newline at end of file

requirements.txt

+pandas==1.5.3
+scikit-learn==1.2.2
+numpy==1.23.5
+matplotlib==3.7.0
+# torch==2.0.0
+reformer-pytorch==1.4.4

run.py

+import argparse
+import torch
+from experiments.exp_long_term_forecasting import Exp_Long_Term_Forecast
+from experiments.exp_long_term_forecasting_partial import Exp_Long_Term_Forecast_Partial
+import random
+import numpy as np
+
+if __name__ == '__main__':
+    fix_seed = 2023
+    random.seed(fix_seed)
+    torch.manual_seed(fix_seed)
+    np.random.seed(fix_seed)
+
+    parser = argparse.ArgumentParser(description='iTransformer')
+
+    # basic config
+    parser.add_argument('--is_training', type=int, required=True, default=1, help='status')
+    parser.add_argument('--model_id', type=str, required=True, default='test', help='model id')
+    parser.add_argument('--model', type=str, required=True, default='iTransformer',
+                        help='model name, options: [iTransformer, iInformer, iReformer, iFlowformer, iFlashformer]')
+
+    # data loader
+    parser.add_argument('--data', type=str, required=True, default='custom', help='dataset type')
+    parser.add_argument('--root_path', type=str, default='./data/electricity/', help='root path of the data file')
+    parser.add_argument('--data_path', type=str, default='electricity.csv', help='data csv file')
+    parser.add_argument('--features', type=str, default='M',
+                        help='forecasting task, options:[M, S, MS]; M:multivariate predict multivariate, S:univariate predict univariate, MS:multivariate predict univariate')
+    parser.add_argument('--target', type=str, default='OT', help='target feature in S or MS task')
+    parser.add_argument('--freq', type=str, default='h',
+                        help='freq for time features encoding, options:[s:secondly, t:minutely, h:hourly, d:daily, b:business days, w:weekly, m:monthly], you can also use more detailed freq like 15min or 3h')
+    parser.add_argument('--checkpoints', type=str, default='./checkpoints/', help='location of model checkpoints')
+
+    # forecasting task
+    parser.add_argument('--seq_len', type=int, default=96, help='input sequence length')
+    parser.add_argument('--label_len', type=int, default=48, help='start token length') # no longer needed in inverted Transformers
+    parser.add_argument('--pred_len', type=int, default=96, help='prediction sequence length')
+
+    # model define
+    parser.add_argument('--enc_in', type=int, default=7, help='encoder input size')
+    parser.add_argument('--dec_in', type=int, default=7, help='decoder input size')
+    parser.add_argument('--c_out', type=int, default=7, help='output size') # applicable on arbitrary number of variates in inverted Transformers
+    parser.add_argument('--d_model', type=int, default=512, help='dimension of model')
+    parser.add_argument('--n_heads', type=int, default=8, help='num of heads')
+    parser.add_argument('--e_layers', type=int, default=2, help='num of encoder layers')
+    parser.add_argument('--d_layers', type=int, default=1, help='num of decoder layers')
+    parser.add_argument('--d_ff', type=int, default=2048, help='dimension of fcn')
+    parser.add_argument('--moving_avg', type=int, default=25, help='window size of moving average')
+    parser.add_argument('--factor', type=int, default=1, help='attn factor')
+    parser.add_argument('--distil', action='store_false',
+                        help='whether to use distilling in encoder, using this argument means not using distilling',
+                        default=True)
+    parser.add_argument('--dropout', type=float, default=0.1, help='dropout')
+    parser.add_argument('--embed', type=str, default='timeF',
+                        help='time features encoding, options:[timeF, fixed, learned]')
+    parser.add_argument('--activation', type=str, default='gelu', help='activation')
+    parser.add_argument('--output_attention', action='store_true', help='whether to output attention in ecoder')
+    parser.add_argument('--do_predict', action='store_true', help='whether to predict unseen future data')
+
+    # optimization
+    parser.add_argument('--num_workers', type=int, default=10, help='data loader num workers')
+    parser.add_argument('--itr', type=int, default=1, help='experiments times')
+    parser.add_argument('--train_epochs', type=int, default=10, help='train epochs')
+    parser.add_argument('--batch_size', type=int, default=32, help='batch size of train input data')
+    parser.add_argument('--patience', type=int, default=3, help='early stopping patience')
+    parser.add_argument('--learning_rate', type=float, default=0.0001, help='optimizer learning rate')
+    parser.add_argument('--des', type=str, default='test', help='exp description')
+    parser.add_argument('--loss', type=str, default='MSE', help='loss function')
+    parser.add_argument('--lradj', type=str, default='type1', help='adjust learning rate')
+    parser.add_argument('--use_amp', action='store_true', help='use automatic mixed precision training', default=False)
+
+    # GPU
+    parser.add_argument('--use_gpu', type=bool, default=True, help='use gpu')
+    parser.add_argument('--gpu', type=int, default=0, help='gpu')
+    parser.add_argument('--use_multi_gpu', action='store_true', help='use multiple gpus', default=False)
+    parser.add_argument('--devices', type=str, default='0,1,2,3', help='device ids of multile gpus')
+
+    # iTransformer
+    parser.add_argument('--exp_name', type=str, required=False, default='MTSF',
+                        help='experiemnt name, options:[MTSF, partial_train]')
+    parser.add_argument('--channel_independence', type=bool, default=False, help='whether to use channel_independence mechanism')
+    parser.add_argument('--inverse', action='store_true', help='inverse output data', default=False)
+    parser.add_argument('--class_strategy', type=str, default='projection', help='projection/average/cls_token')
+    parser.add_argument('--target_root_path', type=str, default='./data/electricity/', help='root path of the data file')
+    parser.add_argument('--target_data_path', type=str, default='electricity.csv', help='data file')
+    parser.add_argument('--efficient_training', type=bool, default=False, help='whether to use efficient_training (exp_name should be partial train)') # See Figure 8 of our paper for the detail
+    parser.add_argument('--use_norm', type=int, default=True, help='use norm and denorm')
+    parser.add_argument('--partial_start_index', type=int, default=0, help='the start index of variates for partial training, '
+                                                                           'you can select [partial_start_index, min(enc_in + partial_start_index, N)]')
+
+    args = parser.parse_args()
+    args.use_gpu = True if torch.cuda.is_available() and args.use_gpu else False
+
+    if args.use_gpu and args.use_multi_gpu:
+        args.devices = args.devices.replace(' ', '')
+        device_ids = args.devices.split(',')
+        args.device_ids = [int(id_) for id_ in device_ids]
+        args.gpu = args.device_ids[0]
+
+    print('Args in experiment:')
+    print(args)
+
+    if args.exp_name == 'partial_train': # See Figure 8 of our paper, for the detail
+        Exp = Exp_Long_Term_Forecast_Partial
+    else: # MTSF: multivariate time series forecasting
+        Exp = Exp_Long_Term_Forecast
+
+
+    if args.is_training:
+        for ii in range(args.itr):
+            # setting record of experiments
+            setting = '{}_{}_{}_{}_ft{}_sl{}_ll{}_pl{}_dm{}_nh{}_el{}_dl{}_df{}_fc{}_eb{}_dt{}_{}_{}'.format(
+                args.model_id,
+                args.model,
+                args.data,
+                args.features,
+                args.seq_len,
+                args.label_len,
+                args.pred_len,
+                args.d_model,
+                args.n_heads,
+                args.e_layers,
+                args.d_layers,
+                args.d_ff,
+                args.factor,
+                args.embed,
+                args.distil,
+                args.des,
+                args.class_strategy, ii)
+
+            exp = Exp(args)  # set experiments
+            print('>>>>>>>start training : {}>>>>>>>>>>>>>>>>>>>>>>>>>>'.format(setting))
+            exp.train(setting)
+
+            print('>>>>>>>testing : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.format(setting))
+            exp.test(setting)
+
+            if args.do_predict:
+                print('>>>>>>>predicting : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.format(setting))
+                exp.predict(setting, True)
+
+            torch.cuda.empty_cache()
+    else:
+        ii = 0
+        setting = '{}_{}_{}_{}_ft{}_sl{}_ll{}_pl{}_dm{}_nh{}_el{}_dl{}_df{}_fc{}_eb{}_dt{}_{}_{}'.format(
+            args.model_id,
+            args.model,
+            args.data,
+            args.features,
+            args.seq_len,
+            args.label_len,
+            args.pred_len,
+            args.d_model,
+            args.n_heads,
+            args.e_layers,
+            args.d_layers,
+            args.d_ff,
+            args.factor,
+            args.embed,
+            args.distil,
+            args.des,
+            args.class_strategy, ii)
+
+        exp = Exp(args)  # set experiments
+        print('>>>>>>>testing : {}<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<'.format(setting))
+        exp.test(setting, test=1)
+        # exp.predict(setting, load=True)
+        torch.cuda.empty_cache()

scripts/boost_performance/ECL/iFlowformer.sh

+export CUDA_VISIBLE_DEVICES=0
+
+model_name=iFlowformer
+# model_name=Flowformer
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_96 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_192 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_336 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_720 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1

scripts/boost_performance/ECL/iInformer.sh

+export CUDA_VISIBLE_DEVICES=0
+
+model_name=iInformer
+# model_name=Informer
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_96 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_192 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_336 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_720 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1

scripts/boost_performance/ECL/iReformer.sh

+export CUDA_VISIBLE_DEVICES=0
+
+model_name=iReformer
+# model_name=Reformer
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_96 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_192 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_336 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_720 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1

scripts/boost_performance/ECL/iTransformer.sh

+export CUDA_VISIBLE_DEVICES=0
+
+model_name=iTransformer
+# model_name=Transformer
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_96 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_192 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_336 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_720 \
+  --model $model_name \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --enc_in 321 \
+  --dec_in 321 \
+  --c_out 321 \
+  --des 'Exp' \
+  --itr 1

scripts/boost_performance/README.md

+# Inverted Transformers Work Better for Time Series Forecasting
+
+This folder contains the comparison of the vanilla Transformer-based forecasters and the inverted versions. If you are new to this repo, we recommend you to have a look at this [README](../multivariate_forecasting/README.md) first.
+
+## Scripts
+
+In each folder named after the dataset, we compare the performance of iTransformers and the vanilla Transformers.
+
+```
+# iTransformer on the Traffic Dataset with gradually enlarged lookback windows.
+
+bash ./scripts/boost_performance/Traffic/iTransformer.sh
+```
+
+You can change the ```model_name``` in the script to select one Transformer variant and its inverted version.
+
+## Results
+We compare the performance of Transformer and iTransformer on all six datasets, indicating that the attention and feed-forward network on the
+inverted dimensions greatly empower Transformers in multivariate time series forecasting.
+
+<p align="center">
+<img src="../../figures/boosting_trm.png" alt="" align=center />
+</p>
+
+We apply the proposed inverted framework to Transformer and its variants. It demonstrates that our iTransformers framework can consistently promote these Transformer variants,
+and take advantage of the booming efficient attention mechanisms.
+<p align="center">
+<img src="../../figures/boosting.png" alt="" align=center />
+</p>