File size: 11,437 Bytes
64fca1b | 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 | from __future__ import annotations
import contextlib
import torch
import torch.nn as nn
from enum import Enum
class InputMode(str, Enum):
TOKENS_ONLY = "tokens_only" # discrete tokens from KronosTokenizer only
FEATURES_ONLY = "features_only" # 21 engineered features only (original mode)
COMBINED = "combined" # tokens + features concatenated then projected
class InputStem(nn.Module):
"""
Converts raw inputs into a unified (B, L, d_model) tensor
regardless of input_mode. The rest of the model never needs
to know which mode is active.
"""
def __init__(self, input_mode: InputMode, d_model: int,
n_tokens: int, # vocab size for token embedding (hierarchical sum)
n_features: int, # number of engineered features (e.g. 21)
s1_bits: int = 6,
s2_bits: int = 6):
super().__init__()
self.mode = InputMode(input_mode)
self.d_model = d_model
self.s1_bits = s1_bits
self.s2_bits = s2_bits
if self.mode in (InputMode.TOKENS_ONLY, InputMode.COMBINED):
self.embed_coarse = nn.Embedding(2 ** s1_bits, d_model)
self.embed_fine = nn.Embedding(2 ** s2_bits, d_model)
self.tok_dropout = nn.Dropout(0.05) # Reduce from 0.50 → 0.05: less aggressive token regularization
if self.mode in (InputMode.FEATURES_ONLY, InputMode.COMBINED):
self.feature_proj = nn.Linear(n_features, d_model)
if self.mode == InputMode.COMBINED:
# After summing token_emb + feature_proj, project back to d_model
# Use a gating mechanism so model learns how much to trust each source
self.gate = nn.Sequential(
nn.Linear(d_model * 2, d_model * 2),
nn.SiLU(),
nn.Linear(d_model * 2, d_model)
)
def forward(self, tokens=None, features=None):
"""
tokens : tuple (idx_coarse, idx_fine) each (B, L) — from tokenizer.encode(half=True)
OR None if mode is features_only
features : (B, L, n_features) float tensor
OR None if mode is tokens_only
Returns : (B, L, d_model)
"""
if self.mode == InputMode.TOKENS_ONLY:
assert tokens is not None, "tokens required for tokens_only mode"
idx_c, idx_f = tokens
emb = self.embed_coarse(idx_c) + self.embed_fine(idx_f) # (B, L, d_model)
return self.tok_dropout(emb) # Apply dropout here
elif self.mode == InputMode.FEATURES_ONLY:
assert features is not None, "features required for features_only mode"
return self.feature_proj(features) # (B, L, d_model)
elif self.mode == InputMode.COMBINED:
assert tokens is not None and features is not None
idx_c, idx_f = tokens
tok_emb = self.embed_coarse(idx_c) + self.embed_fine(idx_f) # (B, L, d_model)
feat_emb = self.feature_proj(features) # (B, L, d_model)
# Gated fusion — model learns weighting between discrete and continuous
fused = self.gate(torch.cat([tok_emb, feat_emb], dim=-1)) # (B, L, d_model)
return fused
class PatchTST(nn.Module):
"""
Channel-independent PatchTST with support for multi-modal input stems.
Architecture:
1. InputStem: (tokens, features) -> (B, L, d_model)
2. Patching: Unfold -> (B, num_patches, patch_len * d_model) -> Linear(d_model)
3. Positional Embedding
4. LSTM (Optional): Temporal smoothing/context
5. Transformer Encoder: Patch-to-Patch attention
6. Aggregation Head: Global pooling -> Projection
"""
def __init__(
self,
seq_len: int = 400,
num_features: int = 21,
patch_len: int = 16,
stride: int = 8,
d_model: int = 128,
n_heads: int = 4,
n_layers: int = 2,
lstm_layers: int = 0,
dropout: float = 0.2,
aggregation: str = "mixing",
input_mode: str = "features_only",
vocab_size: int = 4096,
s1_bits: int = 6,
s2_bits: int = 6,
**legacy_kwargs,
):
super().__init__()
self.seq_len = int(seq_len)
self.num_features = int(num_features)
self.patch_len = int(patch_len)
self.stride = int(stride)
self.d_model = int(d_model)
self.n_heads = int(n_heads)
self.n_layers = int(n_layers)
self.lstm_layers = int(lstm_layers)
self.dropout_rate = float(dropout)
self.aggregation = aggregation.lower().strip()
self.input_mode = InputMode(input_mode)
# ── 1. Input Stem ────────────────────────────────────────────────────
self.input_stem = InputStem(
input_mode=self.input_mode,
d_model=self.d_model,
n_tokens=vocab_size,
n_features=self.num_features,
s1_bits=s1_bits,
s2_bits=s2_bits
)
# ── 2. Patching ──────────────────────────────────────────────────────
self.patch_embed = nn.Linear(self.patch_len * self.d_model, self.d_model)
self.num_patches = (self.seq_len - self.patch_len) // self.stride + 1
self.register_buffer(
"pos_embedding_base",
torch.randn(1, self.num_patches, self.d_model) * 0.02
)
self.dropout = nn.Dropout(dropout)
# ── 4. LSTM (Temporal Context) ───────────────────────────────────────
if self.lstm_layers > 0:
self.lstm = nn.LSTM(
input_size=self.d_model,
hidden_size=self.d_model,
num_layers=self.lstm_layers,
batch_first=True,
dropout=dropout if self.lstm_layers > 1 else 0,
)
else:
self.lstm = None
# ── 5. Transformer Encoder ───────────────────────────────────────────
encoder_layer = nn.TransformerEncoderLayer(
d_model=self.d_model,
nhead=n_heads,
dim_feedforward=self.d_model * 4,
dropout=dropout,
batch_first=True,
norm_first=True,
)
self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=n_layers)
self.enc_dropout = nn.Dropout(dropout)
# ── 6. Head ──────────────────────────────────────────────────────────
if self.aggregation == "mean":
self.head = nn.Linear(self.num_patches * self.d_model, 1)
else: # "mixing"
self.feature_head = nn.Sequential(
nn.Linear(self.d_model, self.d_model // 2), # 128 → 64
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(self.d_model // 2, 1), # 64 → 1
)
self.apply(self._init_weights)
for proj in filter(None, [getattr(self, "head", None), getattr(self, "feature_head", None)]):
if isinstance(proj, nn.Linear):
nn.init.trunc_normal_(proj.weight, std=0.02)
nn.init.zeros_(proj.bias)
def _init_weights(self, m: nn.Module) -> None:
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.Embedding):
# Standard init (0.02): balances expressivity vs OOV noise
nn.init.normal_(m.weight, std=0.02)
elif isinstance(m, nn.LSTM):
for name, param in m.named_parameters():
if 'weight_ih' in name:
nn.init.xavier_uniform_(param.data)
elif 'weight_hh' in name:
nn.init.orthogonal_(param.data)
elif 'bias' in name:
nn.init.constant_(param.data, 0)
def forward(self, tokens=None, features=None) -> torch.Tensor:
"""
Args:
tokens : tuple (s1, s2) each (B, L) or None
features : (B, L, F) float or None
"""
# Step 1: Unified embedding via stem
x = self.input_stem(tokens=tokens, features=features) # (B, L, d_model)
# Step 2: Patching
# unfold: (B, L, d_model) -> (B, num_patches, patch_len, d_model)
x = x.unfold(1, self.patch_len, self.stride)
# flatten: (B, num_patches, patch_len * d_model)
x = x.reshape(x.shape[0], x.shape[1], -1)
# project: (B, num_patches, d_model)
x = self.patch_embed(x)
# Step 3: Positional Embedding — interpolate if seq len changed
num_patches_actual = x.shape[1]
if num_patches_actual == self.num_patches:
pos = self.pos_embedding_base
else:
# Linear interpolation to handle variable-length sequences at val/test
pos = torch.nn.functional.interpolate(
self.pos_embedding_base.transpose(1, 2), # (1, d_model, num_patches)
size=num_patches_actual,
# Linear interpolation for 1D sequence data
mode='linear',
align_corners=False
).transpose(1, 2) # (1, num_patches_actual, d_model)
x = x + pos
x = self.dropout(x)
# Step 4: LSTM (if present)
if self.lstm is not None:
x, _ = self.lstm(x)
# Step 5: Transformer Encoder
x = self.encoder(x)
x = self.enc_dropout(x)
# Step 6: Aggregation
if self.aggregation == "mean":
x_flat = x.reshape(x.shape[0], -1)
x = self.head(x_flat)
return x / (1.0 + x.abs()) # softsign: same range as tanh, gradient never vanishes
else:
# Global average pooling over patches
pooled = torch.mean(x, dim=1)
x = self.feature_head(pooled)
return x / (1.0 + x.abs()) # softsign: same range as tanh, gradient never vanishes
class LPatchTST(PatchTST):
"""
Refined LPatchTST that uses InputStem and follows the
Patch -> LSTM -> Transformer -> Head pipeline.
"""
def __init__(self,
input_mode: str = "combined",
n_features: int = 21,
vocab_size: int = 4096,
s1_bits: int = 6,
s2_bits: int = 6,
d_model: int = 128,
patch_len: int = 8,
stride: int = 4,
**kwargs):
super().__init__(
input_mode=input_mode,
num_features=n_features,
vocab_size=vocab_size,
s1_bits=s1_bits,
s2_bits=s2_bits,
d_model=d_model,
patch_len=patch_len,
stride=stride,
**kwargs
) |