File size: 11,617 Bytes
8b187bb | 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 | """
Knowledge Distillation for MiniMind
Train smaller models using larger teacher models.
"""
import math
from typing import Optional, Dict, Any, Callable
from pathlib import Path
from dataclasses import dataclass
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch.cuda.amp import GradScaler, autocast
@dataclass
class DistillationConfig:
"""Configuration for knowledge distillation."""
# Distillation parameters
temperature: float = 2.0
alpha_ce: float = 0.5 # Weight for hard label loss
alpha_kd: float = 0.5 # Weight for distillation loss
alpha_hidden: float = 0.0 # Weight for hidden state matching
# Optimization
learning_rate: float = 1e-4
min_learning_rate: float = 1e-5
weight_decay: float = 0.1
warmup_steps: int = 500
grad_clip: float = 1.0
# Training
num_epochs: int = 5
batch_size: int = 4
gradient_accumulation_steps: int = 8
max_steps: Optional[int] = None
# Mixed precision
use_amp: bool = True
# Checkpointing
save_steps: int = 500
output_dir: str = "./distill_outputs"
log_steps: int = 10
class DistillationTrainer:
"""
Knowledge Distillation Trainer.
Supports:
- Soft label distillation (KL divergence)
- Hard label training (CE loss)
- Hidden state matching (optional)
- Online and offline distillation
"""
def __init__(
self,
student_model: nn.Module,
teacher_model: Optional[nn.Module] = None,
train_dataloader: DataLoader = None,
config: Optional[DistillationConfig] = None,
projection_layer: Optional[nn.Module] = None,
):
self.student = student_model
self.teacher = teacher_model
self.train_dataloader = train_dataloader
self.config = config or DistillationConfig()
self.projection_layer = projection_layer # For hidden state matching
self.device = next(student_model.parameters()).device
if self.teacher is not None:
self.teacher.eval()
for param in self.teacher.parameters():
param.requires_grad = False
self.optimizer = self._create_optimizer()
self.scheduler = self._create_scheduler()
self.scaler = GradScaler() if self.config.use_amp else None
self.global_step = 0
Path(self.config.output_dir).mkdir(parents=True, exist_ok=True)
def _create_optimizer(self) -> torch.optim.Optimizer:
params = list(self.student.parameters())
if self.projection_layer is not None:
params += list(self.projection_layer.parameters())
return torch.optim.AdamW(
params,
lr=self.config.learning_rate,
weight_decay=self.config.weight_decay,
)
def _create_scheduler(self):
total_steps = self._get_total_steps()
def lr_lambda(step):
if step < self.config.warmup_steps:
return step / max(1, self.config.warmup_steps)
progress = (step - self.config.warmup_steps) / max(1, total_steps - self.config.warmup_steps)
return max(
self.config.min_learning_rate / self.config.learning_rate,
0.5 * (1.0 + math.cos(math.pi * progress))
)
return torch.optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda)
def _get_total_steps(self) -> int:
if self.config.max_steps:
return self.config.max_steps
steps_per_epoch = len(self.train_dataloader) // self.config.gradient_accumulation_steps
return steps_per_epoch * self.config.num_epochs
def distillation_loss(
self,
student_logits: torch.Tensor,
teacher_logits: torch.Tensor,
labels: torch.Tensor,
student_hidden: Optional[torch.Tensor] = None,
teacher_hidden: Optional[torch.Tensor] = None,
) -> Dict[str, torch.Tensor]:
"""
Compute combined distillation loss.
Args:
student_logits: Student model output logits [B, T, V]
teacher_logits: Teacher model output logits [B, T, V]
labels: Ground truth labels [B, T]
student_hidden: Student hidden states (optional)
teacher_hidden: Teacher hidden states (optional)
Returns:
Dictionary with loss components and total loss
"""
# Temperature-scaled soft labels
T = self.config.temperature
# Soft label loss (KL divergence)
student_log_probs = F.log_softmax(student_logits / T, dim=-1)
teacher_probs = F.softmax(teacher_logits / T, dim=-1)
kd_loss = F.kl_div(
student_log_probs,
teacher_probs,
reduction="batchmean"
) * (T ** 2)
# Hard label loss (Cross entropy)
shift_logits = student_logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
ce_loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100,
)
# Hidden state matching (optional)
hidden_loss = torch.tensor(0.0, device=self.device)
if student_hidden is not None and teacher_hidden is not None and self.projection_layer is not None:
projected_student = self.projection_layer(student_hidden)
hidden_loss = F.mse_loss(projected_student, teacher_hidden)
# Combined loss
total_loss = (
self.config.alpha_ce * ce_loss +
self.config.alpha_kd * kd_loss +
self.config.alpha_hidden * hidden_loss
)
return {
"total_loss": total_loss,
"ce_loss": ce_loss,
"kd_loss": kd_loss,
"hidden_loss": hidden_loss,
}
def train(self) -> Dict[str, float]:
"""Main distillation training loop."""
self.student.train()
total_steps = self._get_total_steps()
print(f"Starting knowledge distillation for {total_steps} steps")
print(f" Temperature: {self.config.temperature}")
print(f" Alpha CE: {self.config.alpha_ce}, Alpha KD: {self.config.alpha_kd}")
running_losses = {"total": 0.0, "ce": 0.0, "kd": 0.0}
for epoch in range(self.config.num_epochs):
for step, batch in enumerate(self.train_dataloader):
losses = self._training_step(batch)
for key in running_losses:
running_losses[key] += losses.get(f"{key}_loss", losses.get("total_loss", 0.0)).item() if isinstance(losses.get(f"{key}_loss", losses.get("total_loss")), torch.Tensor) else 0.0
if (step + 1) % self.config.gradient_accumulation_steps == 0:
self._optimizer_step()
self.global_step += 1
if self.global_step % self.config.log_steps == 0:
avg_losses = {k: v / self.config.log_steps for k, v in running_losses.items()}
print(
f"Step {self.global_step}/{total_steps} | "
f"Total: {avg_losses['total']:.4f} | "
f"CE: {avg_losses['ce']:.4f} | "
f"KD: {avg_losses['kd']:.4f}"
)
running_losses = {k: 0.0 for k in running_losses}
if self.global_step % self.config.save_steps == 0:
self._save_checkpoint()
if self.config.max_steps and self.global_step >= self.config.max_steps:
break
if self.config.max_steps and self.global_step >= self.config.max_steps:
break
self._save_checkpoint(final=True)
return {"final_step": self.global_step}
def _training_step(self, batch: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
"""Single distillation training step."""
input_ids = batch["input_ids"].to(self.device)
attention_mask = batch.get("attention_mask")
if attention_mask is not None:
attention_mask = attention_mask.to(self.device)
labels = batch["labels"].to(self.device)
# Check for pre-computed teacher logits
teacher_logits = batch.get("teacher_logits")
if teacher_logits is not None:
teacher_logits = teacher_logits.to(self.device)
elif self.teacher is not None:
with torch.no_grad():
_, teacher_logits, _, _ = self.teacher(input_ids, attention_mask)
if self.config.use_amp:
with autocast(dtype=torch.float16):
_, student_logits, _, _ = self.student(input_ids, attention_mask)
losses = self.distillation_loss(student_logits, teacher_logits, labels)
loss = losses["total_loss"] / self.config.gradient_accumulation_steps
self.scaler.scale(loss).backward()
else:
_, student_logits, _, _ = self.student(input_ids, attention_mask)
losses = self.distillation_loss(student_logits, teacher_logits, labels)
loss = losses["total_loss"] / self.config.gradient_accumulation_steps
loss.backward()
return losses
def _optimizer_step(self):
if self.config.use_amp:
self.scaler.unscale_(self.optimizer)
torch.nn.utils.clip_grad_norm_(self.student.parameters(), self.config.grad_clip)
if self.config.use_amp:
self.scaler.step(self.optimizer)
self.scaler.update()
else:
self.optimizer.step()
self.scheduler.step()
self.optimizer.zero_grad()
def _save_checkpoint(self, final: bool = False):
name = "final" if final else f"step_{self.global_step}"
path = Path(self.config.output_dir) / name
path.mkdir(parents=True, exist_ok=True)
torch.save(self.student.state_dict(), path / "student_model.pt")
if self.projection_layer is not None:
torch.save(self.projection_layer.state_dict(), path / "projection.pt")
print(f"Checkpoint saved to {path}")
def generate_teacher_logits(
teacher_model: nn.Module,
dataloader: DataLoader,
output_path: str,
device: str = "cuda",
top_k: int = 100, # Only save top-k logits to reduce storage
):
"""
Pre-generate teacher logits for offline distillation.
Saves storage by only keeping top-k logits per position.
"""
teacher_model.eval()
teacher_model.to(device)
all_logits = []
print(f"Generating teacher logits for {len(dataloader)} batches...")
with torch.no_grad():
for batch in dataloader:
input_ids = batch["input_ids"].to(device)
attention_mask = batch.get("attention_mask")
if attention_mask is not None:
attention_mask = attention_mask.to(device)
_, logits, _, _ = teacher_model(input_ids, attention_mask)
# Keep only top-k logits
if top_k > 0 and top_k < logits.shape[-1]:
topk_values, topk_indices = torch.topk(logits, k=top_k, dim=-1)
sparse_logits = {
"values": topk_values.cpu(),
"indices": topk_indices.cpu(),
}
all_logits.append(sparse_logits)
else:
all_logits.append(logits.cpu())
torch.save(all_logits, output_path)
print(f"Teacher logits saved to {output_path}")
|