File size: 13,374 Bytes
e1ccef5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
import math
import torch
import torch.nn as nn
from data_provider.data_factory import data_provider



class FourierFilter(nn.Module):
    """
    Fourier Filter: to time-variant and time-invariant term
    """
    def __init__(self, mask_spectrum):
        super(FourierFilter, self).__init__()
        self.mask_spectrum = mask_spectrum
        
    def forward(self, x):
        xf = torch.fft.rfft(x, dim=1)
        mask = torch.ones_like(xf)
        mask[:, self.mask_spectrum, :] = 0
        x_var = torch.fft.irfft(xf*mask, dim=1)
        x_inv = x - x_var
        
        return x_var, x_inv
    

class MLP(nn.Module):
    '''
    Multilayer perceptron to encode/decode high dimension representation of sequential data
    '''
    def __init__(self, 
                 f_in, 
                 f_out, 
                 hidden_dim=128, 
                 hidden_layers=2, 
                 dropout=0.05,
                 activation='tanh'): 
        super(MLP, self).__init__()
        self.f_in = f_in
        self.f_out = f_out
        self.hidden_dim = hidden_dim
        self.hidden_layers = hidden_layers
        self.dropout = dropout
        if activation == 'relu':
            self.activation = nn.ReLU()
        elif activation == 'tanh':
            self.activation = nn.Tanh()
        else:
            raise NotImplementedError
        
        layers = [nn.Linear(self.f_in, self.hidden_dim), 
                  self.activation, nn.Dropout(self.dropout)]
        for i in range(self.hidden_layers-2):
            layers += [nn.Linear(self.hidden_dim, self.hidden_dim),
                       self.activation, nn.Dropout(dropout)]
        
        layers += [nn.Linear(hidden_dim, f_out)]
        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        # x:     B x S x f_in
        # y:     B x S x f_out
        y = self.layers(x)
        return y
    

class KPLayer(nn.Module):
    """
    A demonstration of finding one step transition of linear system by DMD iteratively
    """
    def __init__(self): 
        super(KPLayer, self).__init__()
        
        self.K = None # B E E

    def one_step_forward(self, z, return_rec=False, return_K=False):
        B, input_len, E = z.shape
        assert input_len > 1, 'snapshots number should be larger than 1'
        x, y = z[:, :-1], z[:, 1:]

        # solve linear system
        self.K = torch.linalg.lstsq(x, y).solution # B E E
        if torch.isnan(self.K).any():
            print('Encounter K with nan, replace K by identity matrix')
            self.K = torch.eye(self.K.shape[1]).to(self.K.device).unsqueeze(0).repeat(B, 1, 1)

        z_pred = torch.bmm(z[:, -1:], self.K)
        if return_rec:
            z_rec = torch.cat((z[:, :1], torch.bmm(x, self.K)), dim=1)
            return z_rec, z_pred

        return z_pred
    
    def forward(self, z, pred_len=1):
        assert pred_len >= 1, 'prediction length should not be less than 1'
        z_rec, z_pred= self.one_step_forward(z, return_rec=True)
        z_preds = [z_pred]
        for i in range(1, pred_len):
            z_pred = torch.bmm(z_pred, self.K)
            z_preds.append(z_pred)
        z_preds = torch.cat(z_preds, dim=1)
        return z_rec, z_preds


class KPLayerApprox(nn.Module):
    """
    Find koopman transition of linear system by DMD with multistep K approximation
    """
    def __init__(self): 
        super(KPLayerApprox, self).__init__()
        
        self.K = None # B E E
        self.K_step = None # B E E

    def forward(self, z, pred_len=1):
        # z:       B L E, koopman invariance space representation
        # z_rec:   B L E, reconstructed representation
        # z_pred:  B S E, forecasting representation
        B, input_len, E = z.shape
        assert input_len > 1, 'snapshots number should be larger than 1'
        x, y = z[:, :-1], z[:, 1:]

        # solve linear system
        self.K = torch.linalg.lstsq(x, y).solution # B E E

        if torch.isnan(self.K).any():
            print('Encounter K with nan, replace K by identity matrix')
            self.K = torch.eye(self.K.shape[1]).to(self.K.device).unsqueeze(0).repeat(B, 1, 1)

        z_rec = torch.cat((z[:, :1], torch.bmm(x, self.K)), dim=1) # B L E
        
        if pred_len <= input_len:
            self.K_step = torch.linalg.matrix_power(self.K, pred_len)
            if torch.isnan(self.K_step).any():
                print('Encounter multistep K with nan, replace it by identity matrix')
                self.K_step = torch.eye(self.K_step.shape[1]).to(self.K_step.device).unsqueeze(0).repeat(B, 1, 1)
            z_pred = torch.bmm(z[:, -pred_len:, :], self.K_step)
        else:
            self.K_step = torch.linalg.matrix_power(self.K, input_len)
            if torch.isnan(self.K_step).any():
                print('Encounter multistep K with nan, replace it by identity matrix')
                self.K_step = torch.eye(self.K_step.shape[1]).to(self.K_step.device).unsqueeze(0).repeat(B, 1, 1)
            temp_z_pred, all_pred = z, []
            for _ in range(math.ceil(pred_len / input_len)):
                temp_z_pred = torch.bmm(temp_z_pred, self.K_step)
                all_pred.append(temp_z_pred)
            z_pred = torch.cat(all_pred, dim=1)[:, :pred_len, :]

        return z_rec, z_pred
    

class TimeVarKP(nn.Module):
    """
    Koopman Predictor with DMD (analysitical solution of Koopman operator)
    Utilize local variations within individual sliding window to predict the future of time-variant term
    """
    def __init__(self,
                 enc_in=8,
                 input_len=96,
                 pred_len=96,
                 seg_len=24,
                 dynamic_dim=128,
                 encoder=None,
                 decoder=None,
                 multistep=False,
                ):
        super(TimeVarKP, self).__init__()
        self.input_len = input_len
        self.pred_len = pred_len
        self.enc_in = enc_in
        self.seg_len = seg_len
        self.dynamic_dim = dynamic_dim
        self.multistep = multistep
        self.encoder, self.decoder = encoder, decoder            
        self.freq = math.ceil(self.input_len / self.seg_len)  # segment number of input
        self.step = math.ceil(self.pred_len / self.seg_len)   # segment number of output
        self.padding_len = self.seg_len * self.freq - self.input_len
        # Approximate mulitstep K by KPLayerApprox when pred_len is large
        self.dynamics = KPLayerApprox() if self.multistep else KPLayer() 

    def forward(self, x):
        # x: B L C
        B, L, C = x.shape

        res = torch.cat((x[:, L-self.padding_len:, :], x) ,dim=1)

        res = res.chunk(self.freq, dim=1)     # F x B P C, P means seg_len
        res = torch.stack(res, dim=1).reshape(B, self.freq, -1)   # B F PC

        res = self.encoder(res) # B F H
        x_rec, x_pred = self.dynamics(res, self.step) # B F H, B S H

        x_rec = self.decoder(x_rec) # B F PC
        x_rec = x_rec.reshape(B, self.freq, self.seg_len, self.enc_in)
        x_rec = x_rec.reshape(B, -1, self.enc_in)[:, :self.input_len, :]  # B L C
        
        x_pred = self.decoder(x_pred)     # B S PC
        x_pred = x_pred.reshape(B, self.step, self.seg_len, self.enc_in)
        x_pred = x_pred.reshape(B, -1, self.enc_in)[:, :self.pred_len, :] # B S C

        return x_rec, x_pred


class TimeInvKP(nn.Module):
    """
    Koopman Predictor with learnable Koopman operator
    Utilize lookback and forecast window snapshots to predict the future of time-invariant term
    """
    def __init__(self,
                 input_len=96,
                 pred_len=96,
                 dynamic_dim=128,
                 encoder=None,
                 decoder=None):
        super(TimeInvKP, self).__init__()
        self.dynamic_dim = dynamic_dim
        self.input_len = input_len
        self.pred_len = pred_len
        self.encoder = encoder
        self.decoder = decoder

        K_init = torch.randn(self.dynamic_dim, self.dynamic_dim)
        U, _, V = torch.svd(K_init) # stable initialization
        self.K = nn.Linear(self.dynamic_dim, self.dynamic_dim, bias=False)
        self.K.weight.data = torch.mm(U, V.t())
    
    def forward(self, x):
        # x: B L C
        res = x.transpose(1, 2) # B C L
        res = self.encoder(res) # B C H
        res = self.K(res) # B C H
        res = self.decoder(res) # B C S
        res = res.transpose(1, 2) # B S C

        return res


class Model(nn.Module):
    '''
    Paper link: https://arxiv.org/pdf/2305.18803.pdf
    '''
    def __init__(self, configs, dynamic_dim=128, hidden_dim=64, hidden_layers=2, num_blocks=3, multistep=False):
        """
        mask_spectrum: list, shared frequency spectrums
        seg_len: int, segment length of time series
        dynamic_dim: int, latent dimension of koopman embedding
        hidden_dim: int, hidden dimension of en/decoder
        hidden_layers: int, number of hidden layers of en/decoder
        num_blocks: int, number of Koopa blocks
        multistep: bool, whether to use approximation for multistep K
        alpha: float, spectrum filter ratio
        """
        super(Model, self).__init__()
        self.task_name = configs.task_name
        self.enc_in = configs.enc_in
        self.input_len = configs.seq_len
        self.pred_len = configs.pred_len

        self.seg_len = self.pred_len
        self.num_blocks = num_blocks
        self.dynamic_dim = dynamic_dim
        self.hidden_dim = hidden_dim
        self.hidden_layers = hidden_layers
        self.multistep = multistep
        self.alpha = 0.2
        self.mask_spectrum = self._get_mask_spectrum(configs)

        self.disentanglement = FourierFilter(self.mask_spectrum)

        # shared encoder/decoder to make koopman embedding consistent
        self.time_inv_encoder = MLP(f_in=self.input_len, f_out=self.dynamic_dim, activation='relu',
                    hidden_dim=self.hidden_dim, hidden_layers=self.hidden_layers)
        self.time_inv_decoder = MLP(f_in=self.dynamic_dim, f_out=self.pred_len, activation='relu',
                           hidden_dim=self.hidden_dim, hidden_layers=self.hidden_layers)
        self.time_inv_kps = self.time_var_kps = nn.ModuleList([
                                TimeInvKP(input_len=self.input_len,
                                    pred_len=self.pred_len, 
                                    dynamic_dim=self.dynamic_dim,
                                    encoder=self.time_inv_encoder, 
                                    decoder=self.time_inv_decoder)
                                for _ in range(self.num_blocks)])

        # shared encoder/decoder to make koopman embedding consistent
        self.time_var_encoder = MLP(f_in=self.seg_len*self.enc_in, f_out=self.dynamic_dim, activation='tanh',
                           hidden_dim=self.hidden_dim, hidden_layers=self.hidden_layers)
        self.time_var_decoder = MLP(f_in=self.dynamic_dim, f_out=self.seg_len*self.enc_in, activation='tanh',
                           hidden_dim=self.hidden_dim, hidden_layers=self.hidden_layers)
        self.time_var_kps = nn.ModuleList([
                    TimeVarKP(enc_in=configs.enc_in,
                        input_len=self.input_len,
                        pred_len=self.pred_len,
                        seg_len=self.seg_len,
                        dynamic_dim=self.dynamic_dim,
                        encoder=self.time_var_encoder,
                        decoder=self.time_var_decoder,
                        multistep=self.multistep)
                    for _ in range(self.num_blocks)])

    def _get_mask_spectrum(self, configs):
        """
        get shared frequency spectrums
        """
        train_data, train_loader = data_provider(configs, 'train')
        amps = 0.0
        for data in train_loader:
            lookback_window = data[0]
            amps += abs(torch.fft.rfft(lookback_window, dim=1)).mean(dim=0).mean(dim=1)
        mask_spectrum = amps.topk(int(amps.shape[0]*self.alpha)).indices
        return mask_spectrum # as the spectrums of time-invariant component
    
    def forecast(self, x_enc):
        # Series Stationarization adopted from NSformer
        mean_enc = x_enc.mean(1, keepdim=True).detach() # B x 1 x E
        x_enc = x_enc - mean_enc
        std_enc = torch.sqrt(torch.var(x_enc, dim=1, keepdim=True, unbiased=False) + 1e-5).detach()
        x_enc = x_enc / std_enc

        # Koopman Forecasting
        residual, forecast = x_enc, None
        for i in range(self.num_blocks):
            time_var_input, time_inv_input = self.disentanglement(residual)
            time_inv_output = self.time_inv_kps[i](time_inv_input)
            time_var_backcast, time_var_output = self.time_var_kps[i](time_var_input)
            residual = residual - time_var_backcast
            if forecast is None:
                forecast = (time_inv_output + time_var_output)
            else:
                forecast += (time_inv_output + time_var_output)

        # Series Stationarization adopted from NSformer
        res = forecast * std_enc + mean_enc

        return res        
    
    def forward(self, x_enc, x_mark_enc, x_dec, x_mark_dec):
        if self.task_name == 'long_term_forecast':
            dec_out = self.forecast(x_enc)
            return dec_out[:, -self.pred_len:, :] # [B, L, D]