Spaces:
Sleeping
Sleeping
File size: 11,215 Bytes
67ded87 | 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 | """
Training & Knowledge Distillation Pipelines for Q-TensorFormer (`src/training.py`).
Provides:
- `LanguageModelTrainer`: Deterministic causal LM trainer with AdamW, cosine LR schedule,
gradient clipping, validation tracking, and full training history logging.
- `DistillationTrainer`: Teacher-student knowledge distillation trainer for `scripts/distill.py`.
"""
import math
import time
from typing import Any, Dict, List, Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from .config import TrainingConfig
from .metrics import evaluate_model
class LanguageModelTrainer:
"""
Canonical causal language modeling trainer for Q-TensorFormer and DenseBaseline.
"""
def __init__(
self,
model: nn.Module,
train_config: Optional[TrainingConfig] = None,
device: str = "cpu",
pad_token_id: int = 0,
):
self.model = model.to(device)
self.cfg = train_config or TrainingConfig()
self.device = device
self.pad_token_id = pad_token_id
self.optimizer = torch.optim.AdamW(
self.model.parameters(),
lr=self.cfg.learning_rate,
weight_decay=self.cfg.weight_decay,
)
def _get_lr(self, step: int, total_steps: int) -> float:
base_lr = self.cfg.learning_rate
min_lr = base_lr * self.cfg.lr_min_factor
warmup = min(self.cfg.warmup_steps, max(1, total_steps // 5))
if step < warmup:
return base_lr * float(step + 1) / float(max(1, warmup))
if self.cfg.lr_scheduler == "cosine":
progress = float(step - warmup) / float(max(1, total_steps - warmup))
progress = min(1.0, max(0.0, progress))
return min_lr + 0.5 * (base_lr - min_lr) * (1.0 + math.cos(math.pi * progress))
if self.cfg.lr_scheduler == "linear":
progress = float(step - warmup) / float(max(1, total_steps - warmup))
progress = min(1.0, max(0.0, progress))
return base_lr - progress * (base_lr - min_lr)
return base_lr
def train(
self,
train_loader: DataLoader,
val_loader: Optional[DataLoader] = None,
verbose: bool = True,
) -> Dict[str, Any]:
torch.manual_seed(self.cfg.seed)
total_steps = max(1, self.cfg.max_epochs * len(train_loader))
initial_val = (
evaluate_model(self.model, val_loader, device=self.device, pad_token_id=self.pad_token_id)
if val_loader is not None
else evaluate_model(self.model, train_loader, device=self.device, pad_token_id=self.pad_token_id)
)
epoch_history: List[Dict[str, Any]] = []
global_step = 0
t_start = time.perf_counter()
for epoch in range(1, self.cfg.max_epochs + 1):
self.model.train()
epoch_nll = 0.0
epoch_tokens = 0
t_ep = time.perf_counter()
self.optimizer.zero_grad()
for b_idx, batch in enumerate(train_loader):
if len(batch) == 3:
x, y, attn_mask = batch
attn_mask = attn_mask.to(self.device)
else:
x, y = batch
attn_mask = (x != self.pad_token_id).long().to(self.device)
x = x.to(self.device)
y = y.to(self.device)
lr = self._get_lr(global_step, total_steps)
for pg in self.optimizer.param_groups:
pg["lr"] = lr
try:
out = self.model(x, attention_mask=attn_mask)
except TypeError:
out = self.model(x)
logits = out[0] if isinstance(out, tuple) else (out.logits if hasattr(out, "logits") else out)
vocab_size = logits.size(-1)
valid_mask = y != self.pad_token_id
n_valid = int(valid_mask.sum().item())
if n_valid == 0:
continue
loss = F.cross_entropy(
logits.reshape(-1, vocab_size),
y.reshape(-1),
ignore_index=self.pad_token_id,
reduction="mean",
)
scaled_loss = loss / max(1, self.cfg.gradient_accumulation_steps)
scaled_loss.backward()
if (b_idx + 1) % max(1, self.cfg.gradient_accumulation_steps) == 0 or (b_idx + 1) == len(train_loader):
if self.cfg.max_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.cfg.max_grad_norm)
self.optimizer.step()
self.optimizer.zero_grad()
epoch_nll += float(loss.item()) * n_valid
epoch_tokens += n_valid
global_step += 1
train_loss = epoch_nll / max(1, epoch_tokens)
train_ppl = math.exp(min(train_loss, 20.0))
val_metrics = (
evaluate_model(self.model, val_loader, device=self.device, pad_token_id=self.pad_token_id)
if val_loader is not None
else {"loss": train_loss, "perplexity": train_ppl, "accuracy": 0.0}
)
ep_time = time.perf_counter() - t_ep
rec = {
"epoch": epoch,
"train_loss": round(train_loss, 6),
"train_perplexity": round(train_ppl, 4),
"val_loss": val_metrics["loss"],
"val_perplexity": val_metrics["perplexity"],
"val_accuracy": val_metrics["accuracy"],
"lr": round(self.optimizer.param_groups[0]["lr"], 8),
"epoch_time_s": round(ep_time, 3),
}
epoch_history.append(rec)
if verbose:
print(
f" Epoch {epoch:02d}/{self.cfg.max_epochs:02d} | "
f"train_loss={rec['train_loss']:.4f} (ppl={rec['train_perplexity']:.2f}) | "
f"val_loss={rec['val_loss']:.4f} (ppl={rec['val_perplexity']:.2f}, acc={rec['val_accuracy']:.3f}) | "
f"{rec['epoch_time_s']:.2f}s"
)
total_time = time.perf_counter() - t_start
final_val = (
evaluate_model(self.model, val_loader, device=self.device, pad_token_id=self.pad_token_id)
if val_loader is not None
else evaluate_model(self.model, train_loader, device=self.device, pad_token_id=self.pad_token_id)
)
return {
"initial_val": initial_val,
"final_val": final_val,
"epochs": epoch_history,
"total_steps": global_step,
"total_training_time_s": round(total_time, 3),
"loss_improved": final_val["loss"] < initial_val["loss"],
"perplexity_reduction_ratio": round(
initial_val["perplexity"] / max(1e-6, final_val["perplexity"]), 4
),
}
class DistillationTrainer:
"""
Knowledge distillation trainer for `scripts/distill.py`.
Minimizes `alpha * KL(student || teacher) * T^2 + (1 - alpha) * CE(student, y)`.
"""
def __init__(
self,
student: nn.Module,
teacher: nn.Module,
train_config: Optional[TrainingConfig] = None,
config: Optional[Any] = None,
train_loader: Optional[DataLoader] = None,
val_loader: Optional[DataLoader] = None,
test_loader: Optional[DataLoader] = None,
output_dir: Optional[str] = None,
temperature: float = 2.0,
alpha: float = 0.7,
device: str = "cpu",
pad_token_id: int = 0,
**kwargs,
):
self.student = student.to(device)
self.teacher = teacher.to(device)
self.teacher.eval()
self.train_loader = train_loader
self.val_loader = val_loader
self.test_loader = test_loader
self.output_dir = output_dir
if train_config is not None:
self.cfg = train_config
elif config is not None and hasattr(config, "training"):
self.cfg = config.training
elif config is not None and isinstance(config, TrainingConfig):
self.cfg = config
else:
self.cfg = TrainingConfig()
self.temperature = float(temperature)
self.alpha = float(alpha)
self.device = device
self.pad_token_id = pad_token_id
self.optimizer = torch.optim.AdamW(
self.student.parameters(),
lr=self.cfg.learning_rate,
weight_decay=self.cfg.weight_decay,
)
def train(
self,
train_loader: Optional[DataLoader] = None,
val_loader: Optional[DataLoader] = None,
) -> Dict[str, Any]:
tl = train_loader if train_loader is not None else self.train_loader
vl = val_loader if val_loader is not None else self.val_loader
if tl is None:
raise ValueError("train_loader must be provided either in __init__ or train().")
torch.manual_seed(self.cfg.seed)
history: List[Dict[str, Any]] = []
for epoch in range(1, self.cfg.max_epochs + 1):
self.student.train()
total_loss = 0.0
n_batches = 0
for batch in tl:
x, y = batch[0].to(self.device), batch[1].to(self.device)
self.optimizer.zero_grad()
with torch.no_grad():
t_out = self.teacher(x)
t_logits = t_out[0] if isinstance(t_out, tuple) else t_out
s_out = self.student(x)
s_logits = s_out[0] if isinstance(s_out, tuple) else s_out
vocab_size = s_logits.size(-1)
ce_loss = F.cross_entropy(
s_logits.reshape(-1, vocab_size),
y.reshape(-1),
ignore_index=self.pad_token_id,
)
T = self.temperature
kd_loss = F.kl_div(
F.log_softmax(s_logits / T, dim=-1),
F.softmax(t_logits / T, dim=-1),
reduction="batchmean",
) * (T * T)
loss = self.alpha * kd_loss + (1.0 - self.alpha) * ce_loss
loss.backward()
if self.cfg.max_grad_norm > 0:
torch.nn.utils.clip_grad_norm_(self.student.parameters(), self.cfg.max_grad_norm)
self.optimizer.step()
total_loss += float(loss.item())
n_batches += 1
val_metrics = (
evaluate_model(self.student, vl, device=self.device, pad_token_id=self.pad_token_id)
if vl is not None
else {"loss": total_loss / max(1, n_batches), "perplexity": 0.0}
)
history.append({
"epoch": epoch,
"train_loss": round(total_loss / max(1, n_batches), 4),
"val_loss": val_metrics["loss"],
"val_perplexity": val_metrics["perplexity"],
})
return {"epochs": history}
|