Spaces:
Sleeping
Sleeping
File size: 20,899 Bytes
dc138e1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 | """
transformer.py
Full Transformer implementation for English → Bengali translation
with complete calculation tracking at every step.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
from typing import Optional, Tuple, Dict, List
# ─────────────────────────────────────────────
# Calculation Logger
# ─────────────────────────────────────────────
class CalcLog:
"""Captures every intermediate tensor for visualization."""
def __init__(self):
self.steps: List[Dict] = []
def log(self, name: str, data, formula: str = "", note: str = ""):
entry = {
"name": name,
"formula": formula,
"note": note,
"shape": None,
"value": None,
}
if isinstance(data, torch.Tensor):
entry["shape"] = list(data.shape)
entry["value"] = data.detach().cpu().numpy().tolist()
elif isinstance(data, np.ndarray):
entry["shape"] = list(data.shape)
entry["value"] = data.tolist()
else:
entry["value"] = data
self.steps.append(entry)
return data
def clear(self):
self.steps = []
def to_dict(self):
return self.steps
# ─────────────────────────────────────────────
# Positional Encoding
# ─────────────────────────────────────────────
class PositionalEncoding(nn.Module):
def __init__(self, d_model: int, max_len: int = 512, dropout: float = 0.1):
super().__init__()
self.d_model = d_model
self.dropout = nn.Dropout(dropout)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len).unsqueeze(1).float()
div_term = torch.exp(
torch.arange(0, d_model, 2).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, log: Optional[CalcLog] = None) -> torch.Tensor:
seq_len = x.size(1)
pe_slice = self.pe[:, :seq_len, :]
if log:
log.log("PE_matrix", pe_slice[0, :seq_len, :8],
formula="PE(pos,2i)=sin(pos/10000^(2i/d)), PE(pos,2i+1)=cos(...)",
note=f"Showing first 8 dims for {seq_len} positions")
log.log("Embedding_before_PE", x[0, :, :8],
note="Token embeddings (first 8 dims)")
x = x + pe_slice
if log:
log.log("Embedding_after_PE", x[0, :, :8],
formula="X = Embedding + PE",
note="After adding positional encoding")
return self.dropout(x)
# ─────────────────────────────────────────────
# Scaled Dot-Product Attention
# ─────────────────────────────────────────────
def scaled_dot_product_attention(
Q: torch.Tensor,
K: torch.Tensor,
V: torch.Tensor,
mask: Optional[torch.Tensor] = None,
log: Optional[CalcLog] = None,
head_idx: int = 0,
layer_idx: int = 0,
attn_type: str = "self",
) -> Tuple[torch.Tensor, torch.Tensor]:
d_k = Q.size(-1)
prefix = f"L{layer_idx}_H{head_idx}_{attn_type}"
# Raw scores
scores = torch.matmul(Q, K.transpose(-2, -1))
if log:
log.log(f"{prefix}_Q", Q[0],
formula="Q = X · Wq",
note=f"Query matrix head {head_idx}")
log.log(f"{prefix}_K", K[0],
formula="K = X · Wk",
note=f"Key matrix head {head_idx}")
log.log(f"{prefix}_V", V[0],
formula="V = X · Wv",
note=f"Value matrix head {head_idx}")
log.log(f"{prefix}_QKt", scores[0],
formula="scores = Q · Kᵀ",
note=f"Raw attention scores (before scaling)")
# Scale
scale = math.sqrt(d_k)
scores = scores / scale
if log:
log.log(f"{prefix}_QKt_scaled", scores[0],
formula=f"scores = Q·Kᵀ / √{d_k} = Q·Kᵀ / {scale:.3f}",
note="Scaled scores — prevents vanishing gradients")
# Mask
# masks arrive as (B,1,1,T) or (B,1,T,T) from make_src/tgt_mask;
# scores here are 3-D (B,T_q,T_k) because we loop per-head,
# so squeeze the head dim to avoid (B,B,...) broadcasting.
if mask is not None:
if mask.dim() == 4:
mask = mask.squeeze(1) # (B,1,T,T) or (B,1,1,T) → (B,T,T) or (B,1,T)
scores = scores.masked_fill(mask == 0, float("-inf"))
if log:
log.log(f"{prefix}_mask", mask[0].float(),
formula="mask[i,j]=0 → score=-inf (future token blocked)",
note="Causal mask (training decoder) or padding mask")
log.log(f"{prefix}_scores_masked", scores[0],
note="Scores after masking (-inf will become 0 after softmax)")
# Softmax
attn_weights = F.softmax(scores, dim=-1)
# replace nan from -inf rows with 0 (edge case)
attn_weights = torch.nan_to_num(attn_weights, nan=0.0)
if log:
log.log(f"{prefix}_softmax", attn_weights[0],
formula="α = softmax(scores, dim=-1)",
note="Attention weights — each row sums to 1.0")
# Weighted sum
output = torch.matmul(attn_weights, V)
if log:
log.log(f"{prefix}_output", output[0],
formula="Attention = α · V",
note="Weighted sum of values")
return output, attn_weights
# ─────────────────────────────────────────────
# Multi-Head Attention
# ─────────────────────────────────────────────
class MultiHeadAttention(nn.Module):
def __init__(self, d_model: int, num_heads: int):
super().__init__()
assert d_model % num_heads == 0
self.d_model = d_model
self.num_heads = num_heads
self.d_k = d_model // num_heads
self.W_q = nn.Linear(d_model, d_model, bias=False)
self.W_k = nn.Linear(d_model, d_model, bias=False)
self.W_v = nn.Linear(d_model, d_model, bias=False)
self.W_o = nn.Linear(d_model, d_model, bias=False)
def split_heads(self, x: torch.Tensor) -> torch.Tensor:
B, T, D = x.shape
return x.view(B, T, self.num_heads, self.d_k).transpose(1, 2)
# → (B, num_heads, T, d_k)
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
mask: Optional[torch.Tensor] = None,
log: Optional[CalcLog] = None,
layer_idx: int = 0,
attn_type: str = "self",
) -> Tuple[torch.Tensor, torch.Tensor]:
B = query.size(0)
prefix = f"L{layer_idx}_{attn_type}_MHA"
# Linear projections
Q = self.W_q(query)
K = self.W_k(key)
V = self.W_v(value)
if log:
log.log(f"{prefix}_Wq", self.W_q.weight[:4, :4],
formula="Wq shape: (d_model, d_model)",
note=f"Query weight matrix (first 4×4 shown)")
log.log(f"{prefix}_Q_full", Q[0, :, :8],
formula="Q = input · Wq",
note=f"Full Q projection (first 8 dims shown)")
# Split into heads
Q = self.split_heads(Q) # (B, h, T, d_k)
K = self.split_heads(K)
V = self.split_heads(V)
if log:
log.log(f"{prefix}_Q_head0", Q[0, 0, :, :],
formula=f"Split: (B,T,D) → (B,{self.num_heads},T,{self.d_k})",
note=f"Head 0 queries — d_k={self.d_k}")
# Per-head attention (log only first 2 heads to avoid bloat)
all_attn = []
all_weights = []
for h in range(self.num_heads):
h_log = log if h < 2 else None
out_h, w_h = scaled_dot_product_attention(
Q[:, h], K[:, h], V[:, h],
mask=mask,
log=h_log,
head_idx=h,
layer_idx=layer_idx,
attn_type=attn_type,
)
all_attn.append(out_h)
all_weights.append(w_h)
# Concat heads
concat = torch.stack(all_attn, dim=1) # (B, h, T, d_k)
concat = concat.transpose(1, 2).contiguous() # (B, T, h, d_k)
concat = concat.view(B, -1, self.d_model) # (B, T, D)
if log:
log.log(f"{prefix}_concat", concat[0, :, :8],
formula="concat = [head_1; head_2; ...; head_h]",
note=f"Concatenated heads (first 8 dims)")
# Final projection
output = self.W_o(concat)
if log:
log.log(f"{prefix}_output", output[0, :, :8],
formula="MHA_out = concat · Wo",
note="Final multi-head attention output")
# Stack all attention weights: (B, h, T_q, T_k)
attn_weights = torch.stack(all_weights, dim=1)
return output, attn_weights
# ─────────────────────────────────────────────
# Feed-Forward Network
# ─────────────────────────────────────────────
class FeedForward(nn.Module):
def __init__(self, d_model: int, d_ff: int, dropout: float = 0.1):
super().__init__()
self.linear1 = nn.Linear(d_model, d_ff)
self.linear2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, log: Optional[CalcLog] = None,
layer_idx: int = 0, loc: str = "enc") -> torch.Tensor:
prefix = f"L{layer_idx}_{loc}_FFN"
h = self.linear1(x)
if log:
log.log(f"{prefix}_linear1", h[0, :, :8],
formula="h = X · W1 + b1",
note=f"First linear (d_model→d_ff), showing first 8 dims")
h = F.relu(h)
if log:
log.log(f"{prefix}_relu", h[0, :, :8],
formula="h = ReLU(h) = max(0, h)",
note="Negative values zeroed out")
h = self.dropout(h)
out = self.linear2(h)
if log:
log.log(f"{prefix}_linear2", out[0, :, :8],
formula="out = h · W2 + b2",
note=f"Second linear (d_ff→d_model)")
return out
# ─────────────────────────────────────────────
# Layer Norm + Residual
# ─────────────────────────────────────────────
class AddNorm(nn.Module):
def __init__(self, d_model: int, eps: float = 1e-6):
super().__init__()
self.norm = nn.LayerNorm(d_model, eps=eps)
def forward(self, x: torch.Tensor, sublayer_out: torch.Tensor,
log: Optional[CalcLog] = None, tag: str = "") -> torch.Tensor:
residual = x + sublayer_out
out = self.norm(residual)
if log:
log.log(f"{tag}_residual", residual[0, :, :8],
formula="residual = x + sublayer(x)",
note="Residual (skip) connection")
log.log(f"{tag}_layernorm", out[0, :, :8],
formula="LayerNorm(x) = γ·(x−μ)/σ + β",
note="Layer normalization output")
return out
# ─────────────────────────────────────────────
# Encoder Layer
# ─────────────────────────────────────────────
class EncoderLayer(nn.Module):
def __init__(self, d_model: int, num_heads: int, d_ff: int, dropout: float = 0.1):
super().__init__()
self.self_attn = MultiHeadAttention(d_model, num_heads)
self.ffn = FeedForward(d_model, d_ff, dropout)
self.add_norm1 = AddNorm(d_model)
self.add_norm2 = AddNorm(d_model)
def forward(self, x: torch.Tensor, src_mask: Optional[torch.Tensor] = None,
log: Optional[CalcLog] = None, layer_idx: int = 0):
attn_out, attn_w = self.self_attn(
x, x, x, mask=src_mask, log=log,
layer_idx=layer_idx, attn_type="enc_self"
)
x = self.add_norm1(x, attn_out, log=log, tag=f"L{layer_idx}_enc_self")
ffn_out = self.ffn(x, log=log, layer_idx=layer_idx, loc="enc")
x = self.add_norm2(x, ffn_out, log=log, tag=f"L{layer_idx}_enc_ffn")
return x, attn_w
# ─────────────────────────────────────────────
# Decoder Layer
# ─────────────────────────────────────────────
class DecoderLayer(nn.Module):
def __init__(self, d_model: int, num_heads: int, d_ff: int, dropout: float = 0.1):
super().__init__()
self.masked_self_attn = MultiHeadAttention(d_model, num_heads)
self.cross_attn = MultiHeadAttention(d_model, num_heads)
self.ffn = FeedForward(d_model, d_ff, dropout)
self.add_norm1 = AddNorm(d_model)
self.add_norm2 = AddNorm(d_model)
self.add_norm3 = AddNorm(d_model)
def forward(
self,
x: torch.Tensor,
enc_out: torch.Tensor,
tgt_mask: Optional[torch.Tensor] = None,
src_mask: Optional[torch.Tensor] = None,
log: Optional[CalcLog] = None,
layer_idx: int = 0,
):
# 1. Masked self-attention
m_attn_out, m_attn_w = self.masked_self_attn(
x, x, x, mask=tgt_mask, log=log,
layer_idx=layer_idx, attn_type="dec_masked"
)
x = self.add_norm1(x, m_attn_out, log=log, tag=f"L{layer_idx}_dec_masked")
# 2. Cross-attention: Q from decoder, K/V from encoder
if log:
log.log(f"L{layer_idx}_cross_Q_source", x[0, :, :8],
note="Cross-attn Q comes from DECODER (Bengali context)")
log.log(f"L{layer_idx}_cross_KV_source", enc_out[0, :, :8],
note="Cross-attn K,V come from ENCODER (English context)")
c_attn_out, c_attn_w = self.cross_attn(
query=x, key=enc_out, value=enc_out,
mask=src_mask, log=log,
layer_idx=layer_idx, attn_type="dec_cross"
)
x = self.add_norm2(x, c_attn_out, log=log, tag=f"L{layer_idx}_dec_cross")
# 3. FFN
ffn_out = self.ffn(x, log=log, layer_idx=layer_idx, loc="dec")
x = self.add_norm3(x, ffn_out, log=log, tag=f"L{layer_idx}_dec_ffn")
return x, m_attn_w, c_attn_w
# ─────────────────────────────────────────────
# Full Transformer
# ─────────────────────────────────────────────
class Transformer(nn.Module):
def __init__(
self,
src_vocab_size: int,
tgt_vocab_size: int,
d_model: int = 128,
num_heads: int = 4,
num_layers: int = 2,
d_ff: int = 256,
max_len: int = 64,
dropout: float = 0.1,
pad_idx: int = 0,
):
super().__init__()
self.d_model = d_model
self.pad_idx = pad_idx
self.num_layers = num_layers
self.src_embed = nn.Embedding(src_vocab_size, d_model, padding_idx=pad_idx)
self.tgt_embed = nn.Embedding(tgt_vocab_size, d_model, padding_idx=pad_idx)
self.src_pe = PositionalEncoding(d_model, max_len, dropout)
self.tgt_pe = PositionalEncoding(d_model, max_len, dropout)
self.encoder_layers = nn.ModuleList(
[EncoderLayer(d_model, num_heads, d_ff, dropout) for _ in range(num_layers)]
)
self.decoder_layers = nn.ModuleList(
[DecoderLayer(d_model, num_heads, d_ff, dropout) for _ in range(num_layers)]
)
self.output_linear = nn.Linear(d_model, tgt_vocab_size)
self._init_weights()
def _init_weights(self):
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
# ── mask helpers ──────────────────────────
def make_src_mask(self, src: torch.Tensor) -> torch.Tensor:
# (B, 1, 1, T_src) — 1 where not pad
return (src != self.pad_idx).unsqueeze(1).unsqueeze(2)
def make_tgt_mask(self, tgt: torch.Tensor) -> torch.Tensor:
T = tgt.size(1)
pad_mask = (tgt != self.pad_idx).unsqueeze(1).unsqueeze(2) # (B,1,1,T)
causal = torch.tril(torch.ones(T, T, device=tgt.device)).bool() # (T,T)
return pad_mask & causal # (B,1,T,T)
# ── forward ───────────────────────────────
def forward(
self,
src: torch.Tensor,
tgt: torch.Tensor,
log: Optional[CalcLog] = None,
) -> Tuple[torch.Tensor, Dict]:
src_mask = self.make_src_mask(src)
tgt_mask = self.make_tgt_mask(tgt)
# ── Encoder ──────────────────────────
src_emb = self.src_embed(src) * math.sqrt(self.d_model)
if log:
log.log("SRC_tokens", src[0],
note="Source token IDs (English)")
log.log("SRC_embedding_raw", src_emb[0, :, :8],
formula=f"emb = Embedding(token_id) × √{self.d_model}",
note="Token embeddings (first 8 dims)")
enc_x = self.src_pe(src_emb, log=log)
enc_attn_weights = []
for i, layer in enumerate(self.encoder_layers):
enc_x, ew = layer(enc_x, src_mask=src_mask, log=log, layer_idx=i)
enc_attn_weights.append(ew.detach().cpu().numpy())
if log:
log.log("ENCODER_output", enc_x[0, :, :8],
note="Final encoder output — passed as K,V to every decoder cross-attention")
# ── Decoder ──────────────────────────
tgt_emb = self.tgt_embed(tgt) * math.sqrt(self.d_model)
if log:
log.log("TGT_tokens", tgt[0],
note="Target token IDs (Bengali, teacher-forced in training)")
log.log("TGT_embedding_raw", tgt_emb[0, :, :8],
formula=f"emb = Embedding(token_id) × √{self.d_model}",
note="Bengali token embeddings")
dec_x = self.tgt_pe(tgt_emb, log=log)
dec_self_attn_w = []
dec_cross_attn_w = []
for i, layer in enumerate(self.decoder_layers):
dec_x, mw, cw = layer(
dec_x, enc_x,
tgt_mask=tgt_mask, src_mask=src_mask,
log=log, layer_idx=i,
)
dec_self_attn_w.append(mw.detach().cpu().numpy())
dec_cross_attn_w.append(cw.detach().cpu().numpy())
# ── Output projection ─────────────────
logits = self.output_linear(dec_x) # (B, T, vocab)
if log:
log.log("LOGITS", logits[0, :, :16],
formula="logits = dec_out · W_out (first 16 vocab entries shown)",
note=f"Raw scores over vocab of {logits.size(-1)} Bengali tokens")
probs = F.softmax(logits[0], dim=-1)
log.log("SOFTMAX_probs", probs[:, :16],
formula="P(token) = exp(logit) / Σ exp(logits)",
note="Probability distribution over Bengali vocabulary")
meta = {
"enc_attn": enc_attn_weights,
"dec_self_attn": dec_self_attn_w,
"dec_cross_attn": dec_cross_attn_w,
"src_mask": src_mask.cpu().numpy(),
"tgt_mask": tgt_mask.cpu().numpy(),
}
return logits, meta
|