File size: 15,464 Bytes
d03866e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from jaxtyping import Float, Int
from einops import rearrange


class RMSNorm(nn.Module):
    """Root Mean Square Normalization layer."""

    def __init__(self, size: int, dim: int = -1, eps: float = 1e-5) -> None:
        super().__init__()
        self.scale = nn.Parameter(torch.ones(size))
        self.eps = eps
        self.dim = dim

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        norm_x = x.to(torch.float32).pow(2).mean(dim=self.dim, keepdim=True)
        x_normed = x * torch.rsqrt(norm_x + self.eps)
        return (self.scale * x_normed).type_as(x)


class RotaryEmbedding(nn.Module):
    """Rotary Positional Embedding for injecting positional information."""

    def __init__(self, dim):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, seq_len):
        t = torch.arange(seq_len, device=self.inv_freq.device).type_as(self.inv_freq)
        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        return freqs  # Shape: (seq_len, dim // 2)


class BinaryAttentionBias(nn.Module):
    """Binary Variate Attention for time series data."""

    def __init__(self,
                 num_heads: Int):
        super().__init__()
        self.num_heads = num_heads
        self.emd = nn.Embedding(2, num_heads)

    def forward(self,
                query_id: Int[torch.Tensor, "batch_size q_len"],
                kv_id: Int[torch.Tensor, "batch_size kv_len"],
                ) -> Float[torch.Tensor, "batch_size num_heads q_len kv_len"]:
        ind = torch.eq(query_id.unsqueeze(-1), kv_id.unsqueeze(-2))
        ind = ind.unsqueeze(1)  # (batch_size, 1, q_len, kv_len)
        weight = rearrange(self.emd.weight, "two num_heads -> two num_heads 1 1")  # (2, num_heads, 1, 1)
        bias = ~ind * weight[:1] + ind * weight[1:]  # (batch_size, num_heads, q_len, kv_len)
        return bias


class MultiheadAttentionWithRoPE(nn.Module):
    """Multi-head Attention with Rotary Positional Encoding (RoPE), non-causal by default."""
    "========== NOtice that this applies BinaryAttentionBias ==========="

    def __init__(self, embed_dim, num_heads, num_features):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.head_dim = embed_dim // num_heads
        self.num_features = num_features
        assert self.head_dim * num_heads == embed_dim, "embed_dim must be divisible by num_heads"

        # Linear projections for Q, K, V, and output
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=False)
        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=False)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=False)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=False)

        # Binary attention bias for time series
        if num_features > 1:
            self.binary_attention_bias = BinaryAttentionBias(num_heads)

    def apply_rope(self, x, freqs):
        """Apply Rotary Positional Encoding to the input tensor."""
        B, seq_len, embed_dim = x.shape
        assert embed_dim == self.embed_dim, "Embedding dimension mismatch"
        assert freqs.shape == (seq_len, embed_dim // 2), "freqs shape mismatch"

        # Reshape for rotation: split embed_dim into pairs
        x_ = x.view(B, seq_len, embed_dim // 2, 2)
        cos = freqs.cos().unsqueeze(0)  # (1, seq_len, embed_dim // 2, 1)
        sin = freqs.sin().unsqueeze(0)  # (1, seq_len, embed_dim // 2, 1)

        # Apply rotation to each pair
        x_rot = torch.stack(
            [
                x_[..., 0] * cos - x_[..., 1] * sin,
                x_[..., 0] * sin + x_[..., 1] * cos,
            ],
            dim=-1
        )
        return x_rot.view(B, seq_len, embed_dim)

    def forward(self, query, key, value, freqs, query_id=None, kv_id=None, attn_mask=None):
        """
        Forward pass for multi-head attention with RoPE.

        Args:
            query (Tensor): Shape (B, T, C)
            key (Tensor): Shape (B, T, C)
            value (Tensor): Shape (B, T, C)
            freqs (Tensor): RoPE frequencies, shape (T, embed_dim // 2)
            query_id (Tensor, optional): Shape (B, q_len), feature IDs for query
            kv_id (Tensor, optional): Shape (B, kv_len), feature IDs for key/value
            attn_mask (Tensor, optional): Shape (B, T), True for valid positions, False for padding.

        Returns:
            Tensor: Attention output, shape (B, T, C)
        """
        B, T, C = query.shape
        assert key.shape == (B, T, C) and value.shape == (B, T, C), "query, key, value shapes must match"

        # Project inputs to Q, K, V
        Q = self.q_proj(query)
        K = self.k_proj(key)
        V = self.v_proj(value)

        # Apply RoPE to Q and K
        Q_rot = self.apply_rope(Q, freqs)
        K_rot = self.apply_rope(K, freqs)

        # Reshape for multi-head attention
        Q_rot = Q_rot.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)  # (B, nh, T, hs)
        K_rot = K_rot.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)  # (B, nh, T, hs)
        V = V.view(B, T, self.num_heads, self.head_dim).transpose(1, 2)  # (B, nh, T, hs)

        # Prepare attention mask for padding
        if attn_mask is not None:
            attn_mask = attn_mask.unsqueeze(1).unsqueeze(2)  # (B, 1, 1, T)
        else:
            attn_mask = None

        if query_id is not None and kv_id is not None:
            # Add binary attention bias
            attn_bias = self.binary_attention_bias(query_id, kv_id)  # (B, num_heads, q_len, kv_len)
            scores = torch.matmul(Q_rot, K_rot.transpose(-2, -1)) / math.sqrt(
                self.head_dim)  # (B, num_heads, q_len, kv_len)
            scores += attn_bias
            if attn_mask is not None:
                scores = scores.masked_fill(~attn_mask, float('-inf'))
            attn_weights = F.softmax(scores, dim=-1)  # (B, num_heads, q_len, kv_len)
            y = torch.matmul(attn_weights, V)  # (B, num_heads, q_len, hs)

        else:
            # Compute scaled dot-product attention (non-causal) without binary bias
            # for param in self.binary_attention_bias.parameters():
            #     param.requires_grad = False
            y = F.scaled_dot_product_attention(
                Q_rot, K_rot, V,
                attn_mask=attn_mask,
                is_causal=False  # Non-causal attention for encoder
            )  # (B, nh, T, hs)

        # Reshape and project output
        y = y.transpose(1, 2).contiguous().view(B, T, C)
        y = self.out_proj(y)
        return y


class LlamaMLP(nn.Module):
    def __init__(self, d_model, dim_feedforward=2048):
        super().__init__()
        self.hidden_size = d_model
        self.intermediate_size = dim_feedforward
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=True)
        self.act_fn = F.gelu

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


class TransformerEncoderLayerWithRoPE(nn.Module):
    """Transformer Encoder Layer with RoPE and RMSNorm."""

    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation="relu", num_features=1):
        super().__init__()
        self.self_attn = MultiheadAttentionWithRoPE(d_model, nhead, num_features)
        self.dropout = nn.Dropout(dropout)
        self.input_norm = RMSNorm(d_model)
        self.output_norm = RMSNorm(d_model)
        self.mlp = LlamaMLP(d_model, dim_feedforward)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.activation = F.relu if activation == "relu" else F.gelu

    def forward(self, src, freqs, src_id=None, attn_mask=None):
        residual = src
        src = self.input_norm(src)
        src = self.self_attn(src, src, src, freqs, src_id, src_id, attn_mask=attn_mask)
        src = src + residual
        residual = src
        src = self.output_norm(src)
        src = self.mlp(src)
        src = residual + self.dropout2(src)
        return src


class CustomTransformerEncoder(nn.Module):
    """Stack of Transformer Encoder Layers."""

    def __init__(self, d_model, nhead, dim_feedforward, dropout, activation, num_layers, num_features):
        super().__init__()
        self.layers = nn.ModuleList([
            TransformerEncoderLayerWithRoPE(
                d_model=d_model,
                nhead=nhead,
                dim_feedforward=dim_feedforward,
                dropout=dropout,
                activation=activation,
                num_features=num_features
            ) for _ in range(num_layers)
        ])

    def forward(self, src, freqs, src_id=None, attn_mask=None):
        output = src
        for layer in self.layers:
            output = layer(output, freqs, src_id, attn_mask=attn_mask)
        return output


class TimeSeriesEncoder(nn.Module):
    """
    Time Series Encoder with PatchTST-like patching, RoPE.

    Args:
        d_model (int): Model dimension
        d_proj (int): Projection dimension
        patch_size (int): Size of each patch
        num_layers (int): Number of encoder layers
        num_heads (int): Number of attention heads
        d_ff_dropout (float): Dropout rate
        max_total_tokens (int): Maximum sequence length
        use_rope (bool): Use RoPE if True
        num_features (int): Number of features in the time series
        activation (str): "relu" or "gelu"

    Inputs:
        time_series (Tensor): Shape (batch_size, seq_len, num_features)
        mask (Tensor): Shape (batch_size, seq_len)

    Outputs:
        local_embeddings (Tensor): Shape (batch_size, seq_len, num_features, d_proj)
    """

    def __init__(self, d_model=2048, d_proj=512, patch_size=32, num_layers=6, num_heads=8,
                 d_ff_dropout=0.1, max_total_tokens=8192, use_rope=True, num_features=1,
                 activation="relu"):
        super().__init__()
        self.patch_size = patch_size
        self.d_model = d_model
        self.d_proj = d_proj
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.d_ff_dropout = d_ff_dropout
        self.max_total_tokens = max_total_tokens
        self.use_rope = use_rope
        self.num_features = num_features
        self.activation = activation

        # Patch embedding layer
        self.embedding_layer = nn.Linear(patch_size, d_model)

        if use_rope:
            # Initialize RoPE and custom encoder
            self.rope_embedder = RotaryEmbedding(d_model)
            self.transformer_encoder = CustomTransformerEncoder(
                d_model=d_model,
                nhead=num_heads,
                dim_feedforward=d_model * 4,
                dropout=d_ff_dropout,
                activation=activation,
                num_layers=num_layers,
                num_features=num_features
            )
        else:
            # Standard encoder without RoPE
            encoder_layer = nn.TransformerEncoderLayer(
                d_model=d_model,
                nhead=num_heads,
                dim_feedforward=d_model * 4,
                dropout=d_ff_dropout,
                batch_first=True,
                activation=activation
            )
            self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers)

        # Output projection layers
        self.projection_layer = nn.Linear(d_model, patch_size * d_proj)
        self._init_parameters()

    def _init_parameters(self):
        for name, param in self.named_parameters():
            if 'weight' in name and 'linear' in name:
                if self.activation == "relu":
                    nn.init.kaiming_uniform_(param, nonlinearity='relu')
                elif self.activation == "gelu":
                    nn.init.kaiming_uniform_(param, nonlinearity='gelu')
            elif 'bias' in name:
                nn.init.constant_(param, 0.0)

    def forward(self, time_series, mask):
        """Forward pass to generate local embeddings."""
        if time_series.dim() == 2:
            time_series = time_series.unsqueeze(-1)
        device = time_series.device
        B, seq_len, num_features = time_series.size()
        assert num_features == self.num_features, f"Number of features mismatch with data: {num_features} vs param: {self.num_features}"
        assert mask.size() == (B, seq_len), "Mask shape mismatch"

        # Pad sequence to be divisible by patch_size
        padded_length = math.ceil(seq_len / self.patch_size) * self.patch_size
        if padded_length > seq_len:
            pad_amount = padded_length - seq_len
            time_series = F.pad(time_series, (0, 0, 0, pad_amount), value=0)
            mask = F.pad(mask, (0, pad_amount), value=0)

        # Convert to patches
        num_patches = padded_length // self.patch_size
        total_length = num_patches * num_features
        patches = time_series.view(B, num_patches, self.patch_size, num_features)
        patches = patches.permute(0, 3, 1, 2).contiguous()  # (B, num_features, num_patches, patch_size)
        patches = patches.view(B, num_features * num_patches, self.patch_size)  # (B, L, patch_size)
        # Create feature IDs for patches
        feature_id = torch.arange(num_features, device=device).repeat_interleave(
            num_patches)  # (num_features * num_patches = L,)
        feature_id = feature_id.unsqueeze(0).expand(B, -1)  # (B, L)

        # Embed patches
        embedded_patches = self.embedding_layer(patches)  # (B, L, d_model)

        # Create patch-level mask
        mask = mask.view(B, num_patches, self.patch_size)
        patch_mask = mask.sum(dim=-1) > 0  # (B, num_patches)
        full_mask = patch_mask.unsqueeze(1).expand(-1, num_features, -1)  # (B, num_features, num_patches)
        full_mask = full_mask.reshape(B, num_features * num_patches)  # (B, L)

        # Generate RoPE frequencies if applicable
        if self.use_rope:
            freqs = self.rope_embedder(total_length).to(device)
        else:
            freqs = None

        # Encode sequence
        if num_features > 1:
            output = self.transformer_encoder(
                embedded_patches,
                freqs=freqs,
                src_id=feature_id,
                attn_mask=full_mask
            )
        else:
            output = self.transformer_encoder(
                embedded_patches,
                freqs=freqs,
                attn_mask=full_mask
            )

        # Extract and project local embeddings
        patch_embeddings = output  # (B, L, d_model)
        patch_proj = self.projection_layer(patch_embeddings)  # (B, L, patch_size * d_proj)
        local_embeddings = patch_proj.view(B, num_features, num_patches, self.patch_size, self.d_proj)
        local_embeddings = local_embeddings.permute(0, 2, 3, 1, 4)  # (B, num_patches, patch_size, num_features, d_proj)
        local_embeddings = local_embeddings.view(B, -1, num_features, self.d_proj)[:, :seq_len, :,
                           :]  # (B, seq_len, num_features, d_proj)

        return local_embeddings