add Reformer

exp/exp_main.py

@@ -1,6 +1,6 @@
 from data_provider.data_factory import data_provider
 from exp.exp_basic import Exp_Basic
-from models import Informer, Autoformer, Transformer
+from models import Informer, Autoformer, Transformer, Reformer
 from utils.tools import EarlyStopping, adjust_learning_rate, visual
 from utils.metrics import metric
 
@@ -28,6 +28,7 @@ def _build_model(self):
             'Autoformer': Autoformer,
             'Transformer': Transformer,
             'Informer': Informer,
+            'Reformer': Reformer,
         }
         model = model_dict[self.args.model].Model(self.args).float()
 

layers/SelfAttention_Family.py

@@ -8,6 +8,7 @@
 import math
 from math import sqrt
 from utils.masking import TriangularCausalMask, ProbMask
+from reformer_pytorch import LSHSelfAttention
 import os
 
 
@@ -164,3 +165,33 @@ def forward(self, queries, keys, values, attn_mask):
         out = out.view(B, L, -1)
 
         return self.out_projection(out), attn
+
+
+class ReformerLayer(nn.Module):
+    def __init__(self, attention, d_model, n_heads, d_keys=None,
+                 d_values=None, causal=False, bucket_size=4, n_hashes=4):
+        super().__init__()
+        self.bucket_size = bucket_size
+        self.attn = LSHSelfAttention(
+            dim=d_model,
+            heads=n_heads,
+            bucket_size=bucket_size,
+            n_hashes=n_hashes,
+            causal=causal
+        )
+
+    def fit_length(self, queries):
+        # inside reformer: assert N % (bucket_size * 2) == 0
+        B, N, C = queries.shape
+        if N % (self.bucket_size * 2) == 0:
+            return queries
+        else:
+            # fill the time series
+            fill_len = (self.bucket_size * 2) - (N % (self.bucket_size * 2))
+            return torch.cat([queries, torch.zeros([B, fill_len, C]).to(queries.device)], dim=1)
+
+    def forward(self, queries, keys, values, attn_mask):
+        # in Reformer: defalut queries=keys
+        B, N, C = queries.shape
+        queries = self.attn(self.fit_length(queries))[:, :N, :]
+        return queries, None

models/Reformer.py

@@ -0,0 +1,57 @@
+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 ReformerLayer
+from layers.Embed import DataEmbedding
+import numpy as np
+
+
+class Model(nn.Module):
+    """
+    Reformer with O(LlogL) complexity
+    - It is notable that Reformer is not proposed for time series forecasting, in that it cannot accomplish the cross attention.
+    - Here is only one adaption in BERT-style, other possible implementations can also be acceptable
+    - The hyper-parameters, such as bucket_size and n_hashes, need to be further tuned.
+    The official repo of Reformer (https://github.com/lucidrains/reformer-pytorch) can be very helpful, if you have any question.
+    """
+
+    def __init__(self, configs):
+        super(Model, self).__init__()
+        self.pred_len = configs.pred_len
+        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(
+                    ReformerLayer(None, configs.d_model, configs.n_heads, bucket_size=configs.bucket_size,
+                                  n_hashes=configs.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 forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec,
+                enc_self_mask=None, dec_self_mask=None, dec_enc_mask=None):
+        # add placeholder
+        x_enc = torch.cat([x_enc, x_dec[:, -self.pred_len:, :]], dim=1)
+        x_mark_enc = torch.cat([x_mark_enc, x_mark_dec[:, -self.pred_len:, :]], dim=1)
+        # Reformer: encoder only
+        enc_out = self.enc_embedding(x_enc, x_mark_enc)
+        enc_out, attns = self.encoder(enc_out, attn_mask=enc_self_mask)
+        enc_out = self.projection(enc_out)
+
+        if self.output_attention:
+            return enc_out[:, -self.pred_len:, :], attns
+        else:
+            return enc_out[:, -self.pred_len:, :]  # [B, L, D]

requirements.txt

@@ -0,0 +1,6 @@
+pandas
+sklearn
+torchvision
+numpy
+matplotlib
+reformer_pytorch

run.py

@@ -35,6 +35,8 @@
 parser.add_argument('--pred_len', type=int, default=96, help='prediction sequence length')
 
 # model define
+parser.add_argument('--bucket_size', type=int, default=4, help='for Reformer')
+parser.add_argument('--n_hashes', type=int, default=4, help='for Reformer')
 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')

scripts/ECL_script/Reformer.sh

@@ -0,0 +1,81 @@
+export CUDA_VISIBLE_DEVICES=2
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/electricity/ \
+  --data_path electricity.csv \
+  --model_id ECL_96_96 \
+  --model Reformer \
+  --data custom \
+  --features S \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 1 \
+  --dec_in 1 \
+  --c_out 1 \
+  --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 Reformer \
+  --data custom \
+  --features S \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 1 \
+  --dec_in 1 \
+  --c_out 1 \
+  --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 Reformer \
+  --data custom \
+  --features S \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 1 \
+  --dec_in 1 \
+  --c_out 1 \
+  --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 Reformer \
+  --data custom \
+  --features S \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 1 \
+  --dec_in 1 \
+  --c_out 1 \
+  --des 'Exp' \
+  --itr 1

scripts/ETT_script/Reformer.sh

@@ -0,0 +1,78 @@
+export CUDA_VISIBLE_DEVICES=0
+
+# multivariate setting of vanilla transformer
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/ETT-small/ \
+  --data_path ETTm1.csv \
+  --model_id ETTm1_96_96 \
+  --model Reformer \
+  --data ETTm1 \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --enc_in 7 \
+  --dec_in 7 \
+  --c_out 7 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/ETT-small/ \
+  --data_path ETTm1.csv \
+  --model_id ETTm1_96_192 \
+  --model Reformer \
+  --data ETTm1 \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --enc_in 7 \
+  --dec_in 7 \
+  --c_out 7 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/ETT-small/ \
+  --data_path ETTm1.csv \
+  --model_id ETTm1_96_336 \
+  --model Reformer \
+  --data ETTm1 \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --enc_in 7 \
+  --dec_in 7 \
+  --c_out 7 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/ETT-small/ \
+  --data_path ETTm1.csv \
+  --model_id ETTm1_96_720 \
+  --model Reformer \
+  --data ETTm1 \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --enc_in 7 \
+  --dec_in 7 \
+  --c_out 7 \
+  --des 'Exp' \
+  --itr 1

scripts/Exchange_script/Reformer.sh

@@ -0,0 +1,83 @@
+export CUDA_VISIBLE_DEVICES=4
+
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/exchange_rate/ \
+  --data_path exchange_rate.csv \
+  --model_id Exchange_96_96 \
+  --model Reformer \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 96 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 8 \
+  --dec_in 8 \
+  --c_out 8 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/exchange_rate/ \
+  --data_path exchange_rate.csv \
+  --model_id Exchange_96_192 \
+  --model Reformer \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 192 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 8 \
+  --dec_in 8 \
+  --c_out 8 \
+  --des 'Exp' \
+  --itr 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/exchange_rate/ \
+  --data_path exchange_rate.csv \
+  --model_id Exchange_96_336 \
+  --model Reformer \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 336 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 8 \
+  --dec_in 8 \
+  --c_out 8 \
+  --des 'Exp' \
+  --itr 1 \
+  --train_epochs 1
+
+python -u run.py \
+  --is_training 1 \
+  --root_path ./dataset/exchange_rate/ \
+  --data_path exchange_rate.csv \
+  --model_id Exchange_96_720 \
+  --model Reformer \
+  --data custom \
+  --features M \
+  --seq_len 96 \
+  --label_len 48 \
+  --pred_len 720 \
+  --e_layers 2 \
+  --d_layers 1 \
+  --factor 3 \
+  --enc_in 8 \
+  --dec_in 8 \
+  --c_out 8 \
+  --des 'Exp' \
+  --itr 1