File size: 14,203 Bytes
d3b9aa2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 | """
model.py β TriChronos-50M
Encoder-only Transformer for probabilistic time-series forecasting.
Architecture
------------
d_model = 768
n_layers = 6
n_heads = 12
patch_size = 8
Each encoder block alternates between:
1. Temporal self-attention (within the sequence dimension)
2. Group attention (across the batch dimension β cross-series)
3. BitLinear FFN
All Q/K/V/O projections inside attention blocks use BitLinear.
Patch embedding and quantile output head stay in FP16.
Output: 21 quantile predictions (Ο = 0.05, 0.10, β¦, 0.95 + 0.50 median)
"""
from __future__ import annotations
import math
from typing import List, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from bitlinear import BitLinear
# ---------------------------------------------------------------------------
# Quantile levels (21 values)
# ---------------------------------------------------------------------------
QUANTILE_LEVELS: List[float] = [0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4,
0.45, 0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8,
0.85, 0.9, 0.95, 0.025, 0.975]
N_QUANTILES: int = len(QUANTILE_LEVELS) # 21
# ---------------------------------------------------------------------------
# Sinusoidal positional encoding
# ---------------------------------------------------------------------------
class SinusoidalPE(nn.Module):
"""Fixed sinusoidal positional encoding (Vaswani et al., 2017)."""
def __init__(self, d_model: int, max_len: int = 4096):
super().__init__()
pe = torch.zeros(max_len, d_model)
position = torch.arange(max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2, dtype=torch.float) * (-math.log(10000.0) / d_model)
)
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
self.register_buffer("pe", pe.unsqueeze(0)) # (1, max_len, d_model)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x : (B, T, d_model)
return x + self.pe[:, : x.size(1)]
# ---------------------------------------------------------------------------
# Patch embedding (FP16, not quantized)
# ---------------------------------------------------------------------------
class PatchEmbedding(nn.Module):
"""
Linear projection of each patch into the model dimension.
Stays in FP16 β this is the entry-point to the model and the
quantization boundary.
"""
def __init__(self, patch_size: int, d_model: int):
super().__init__()
self.proj = nn.Linear(patch_size, d_model, bias=True)
self.pos_enc = SinusoidalPE(d_model)
self.norm = nn.LayerNorm(d_model)
def forward(self, patches: torch.Tensor) -> torch.Tensor:
# patches : (B, n_patches, patch_size)
x = self.proj(patches) # (B, n_patches, d_model)
x = self.pos_enc(x)
x = self.norm(x)
return x
# ---------------------------------------------------------------------------
# BitLinear multi-head attention (shared by temporal + group variants)
# ---------------------------------------------------------------------------
class BitMHA(nn.Module):
"""
Multi-head attention using BitLinear projections for Q, K, V, and O.
"""
def __init__(self, d_model: int, n_heads: int, dropout: float = 0.0):
super().__init__()
assert d_model % n_heads == 0
self.n_heads = n_heads
self.d_head = d_model // n_heads
self.scale = self.d_head ** -0.5
self.q_proj = BitLinear(d_model, d_model, bias=False)
self.k_proj = BitLinear(d_model, d_model, bias=False)
self.v_proj = BitLinear(d_model, d_model, bias=False)
self.o_proj = BitLinear(d_model, d_model, bias=False)
self.dropout = nn.Dropout(dropout)
def _split_heads(self, x: torch.Tensor) -> torch.Tensor:
B, T, D = x.shape
return x.view(B, T, self.n_heads, self.d_head).transpose(1, 2)
# β (B, n_heads, T, d_head)
def _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
B, H, T, D = x.shape
return x.transpose(1, 2).contiguous().view(B, T, H * D)
def forward(
self,
q_in: torch.Tensor, # query source (B, Tq, d_model)
kv_in: torch.Tensor, # key/value src (B, Tk, d_model)
attn_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
Q = self._split_heads(self.q_proj(q_in))
K = self._split_heads(self.k_proj(kv_in))
V = self._split_heads(self.v_proj(kv_in))
attn = torch.matmul(Q, K.transpose(-2, -1)) * self.scale
if attn_mask is not None:
attn = attn + attn_mask
attn = F.softmax(attn, dim=-1)
attn = self.dropout(attn)
out = torch.matmul(attn, V) # (B, n_heads, Tq, d_head)
out = self._merge_heads(out) # (B, Tq, d_model)
return self.o_proj(out)
# ---------------------------------------------------------------------------
# Group (cross-series) attention
# ---------------------------------------------------------------------------
class GroupAttention(nn.Module):
"""
Cross-series attention: treats the batch dimension as the sequence
dimension and performs attention across different series in the batch.
For each patch position t, we gather the representation of that position
across all B series in the batch and run attention over the B "tokens".
Assumption: the batch is composed of related series (same dataset/subset).
The data pipeline sorts samples by subset to honour this assumption.
"""
def __init__(self, d_model: int, n_heads: int, dropout: float = 0.0):
super().__init__()
self.mha = BitMHA(d_model, n_heads, dropout)
self.norm = nn.LayerNorm(d_model)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x : (B, T, d_model)
B, T, D = x.shape
# Swap B and T: treat each patch position as an independent "sequence"
# of B "tokens" (one per series in the batch).
x_t = x.permute(1, 0, 2) # (T, B, d_model)
# Run attention over the B dimension for each of the T positions.
# We process all T positions at once by temporarily reshaping:
# (T, B, D) β viewed as (T*1, B, D) so that the batch dim = T
# and the sequence dim = B.
x_flat = x_t.reshape(T, B, D) # (T, B, D) β B is seq len here
out = self.mha(x_flat, x_flat) # (T, B, D)
out = self.norm(out + x_flat) # residual + norm
# Restore original layout
return out.permute(1, 0, 2) # (B, T, D)
# ---------------------------------------------------------------------------
# BitLinear FFN
# ---------------------------------------------------------------------------
class BitFFN(nn.Module):
"""Two-layer FFN with BitLinear and GELU activation."""
def __init__(self, d_model: int, ffn_dim: int, dropout: float = 0.0):
super().__init__()
self.fc1 = BitLinear(d_model, ffn_dim, bias=False)
self.fc2 = BitLinear(ffn_dim, d_model, bias=False)
self.dropout = nn.Dropout(dropout)
self.norm = nn.LayerNorm(d_model)
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
x = self.fc1(x)
x = F.gelu(x)
x = self.dropout(x)
x = self.fc2(x)
x = self.dropout(x)
return self.norm(x + residual)
# ---------------------------------------------------------------------------
# Encoder block
# ---------------------------------------------------------------------------
class EncoderBlock(nn.Module):
"""
Single encoder block with:
1. Temporal self-attention (within sequence) + residual + norm
2. Group attention (across batch) + residual + norm
3. BitLinear FFN + residual + norm
"""
def __init__(self, d_model: int, n_heads: int, ffn_dim: int, dropout: float = 0.0):
super().__init__()
# Temporal attention
self.temporal_attn = BitMHA(d_model, n_heads, dropout)
self.temporal_norm = nn.LayerNorm(d_model)
# Group (cross-series) attention
self.group_attn = GroupAttention(d_model, n_heads, dropout)
# FFN
self.ffn = BitFFN(d_model, ffn_dim, dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# 1. Temporal self-attention
residual = x
x = self.temporal_attn(x, x)
x = self.temporal_norm(x + residual)
# 2. Group attention (operates on batch axis β no residual shape conflict)
x = self.group_attn(x)
# 3. FFN
x = self.ffn(x)
return x
# ---------------------------------------------------------------------------
# Quantile output head (FP16)
# ---------------------------------------------------------------------------
class QuantileHead(nn.Module):
"""
FP16 head that maps the pooled encoder representation to 21 quantile
predictions over the forecast horizon.
Input : (B, d_model) β mean-pooled encoder output
Output : (B, horizon, 21) β quantile predictions per future timestep
"""
def __init__(self, d_model: int, horizon: int, n_quantiles: int = N_QUANTILES):
super().__init__()
self.horizon = horizon
self.n_quantiles = n_quantiles
# Residual projection β stays FP16
self.proj = nn.Linear(d_model, horizon * n_quantiles, bias=True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# x : (B, d_model)
out = self.proj(x) # (B, horizon*n_quantiles)
return out.view(x.size(0), self.horizon, self.n_quantiles)
# ---------------------------------------------------------------------------
# TriChronos β full model
# ---------------------------------------------------------------------------
class TriChronos(nn.Module):
"""
TriChronos-0.1B: encoder-only Transformer for time-series forecasting.
Config (hand-verified ~49.9M params)
--------------------------------------
d_model = 768
n_layers = 6
n_heads = 12
patch_size = 8
ffn_dim = 2304 (~3 Γ d_model; tuned to hit ~50M target)
"""
def __init__(
self,
patch_size: int = 8,
d_model: int = 768,
n_layers: int = 6,
n_heads: int = 12,
ffn_dim: int = 2304,
horizon: int = 24,
dropout: float = 0.1,
n_quantiles: int = N_QUANTILES,
):
super().__init__()
self.patch_size = patch_size
self.d_model = d_model
self.n_layers = n_layers
self.horizon = horizon
# Patch embedding (FP16)
self.patch_embed = PatchEmbedding(patch_size, d_model)
# Encoder stack
self.encoder = nn.ModuleList([
EncoderBlock(d_model, n_heads, ffn_dim, dropout)
for _ in range(n_layers)
])
self.encoder_norm = nn.LayerNorm(d_model)
# Output head (FP16)
self.quantile_head = QuantileHead(d_model, horizon, n_quantiles)
# Initialise weights
self._init_weights()
# ------------------------------------------------------------------
def _init_weights(self):
for module in self.modules():
if isinstance(module, (nn.Linear, BitLinear)):
nn.init.trunc_normal_(module.weight, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
nn.init.ones_(module.weight)
nn.init.zeros_(module.bias)
# ------------------------------------------------------------------
def forward(self, patches: torch.Tensor) -> torch.Tensor:
"""
Parameters
----------
patches : (B, n_patches, patch_size) β Gold-stage output
Returns
-------
quantiles : (B, horizon, n_quantiles) β probabilistic forecast
"""
# Patch embedding
x = self.patch_embed(patches) # (B, n_patches, d_model)
# Encoder stack
for block in self.encoder:
x = block(x)
x = self.encoder_norm(x) # (B, n_patches, d_model)
# Global mean pooling over patch dimension
x = x.mean(dim=1) # (B, d_model)
# Quantile predictions
return self.quantile_head(x) # (B, horizon, n_quantiles)
# ------------------------------------------------------------------
def count_params(self, trainable_only: bool = True) -> int:
"""Return total (or trainable-only) parameter count."""
params = (
self.parameters() if not trainable_only
else filter(lambda p: p.requires_grad, self.parameters())
)
return sum(p.numel() for p in params)
# ------------------------------------------------------------------
@property
def quantile_levels(self) -> List[float]:
return QUANTILE_LEVELS
# ---------------------------------------------------------------------------
# Quick smoke test (python model.py)
# ---------------------------------------------------------------------------
if __name__ == "__main__":
torch.manual_seed(42)
model = TriChronos()
n_params = model.count_params()
print(f"TriChronos-50M - {n_params:,} trainable parameters")
print(f" Target: ~49,900,000")
print(f" Delta : {abs(n_params - 49_900_000):,}")
# Synthetic batch
B, T = 4, 64 # batch=4, 64 patches
patches = torch.randn(B, T, model.patch_size)
quantiles = model(patches)
print(f"\nInput shape : {patches.shape}")
print(f"Output shape : {quantiles.shape} (expect ({B}, {model.horizon}, {N_QUANTILES}))")
assert quantiles.shape == (B, model.horizon, N_QUANTILES)
# Backward pass
loss = quantiles.sum()
loss.backward()
print(f"\nBackward pass β OK")
print("model.py - all checks passed OK")
|