File size: 7,782 Bytes
c8b77b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Refactored LoRA Knowledge Distillation Trainer using modular architecture.

This module implements a clean, testable trainer that follows the interface contracts
and provides better separation of concerns.
"""

import logging
from pathlib import Path
from typing import Any, Dict, Optional

import torch
import torch.nn as nn
from torch.utils.data import DataLoader

from ..core.base_components import BaseTrainer
from ..core.exceptions import TrainingError
from ..core.interfaces import TrainingConfig

logger = logging.getLogger(__name__)


class ModularLoRATrainer(BaseTrainer):
    """Modular LoRA trainer with clean separation of concerns."""

    def __init__(
        self,
        model: nn.Module,
        optimizer: torch.optim.Optimizer,
        loss_function,
        device: str = "cpu",
        teacher_model: Optional[nn.Module] = None,
    ):
        """
        Initialize the modular LoRA trainer.

        Args:
            model: Student model to train
            optimizer: Optimizer for training
            loss_function: Loss function implementing ILossFunction
            device: Training device
            teacher_model: Optional teacher model for distillation
        """
        super().__init__(model, optimizer, device)
        self.loss_function = loss_function
        self.teacher_model = teacher_model
        if self.teacher_model:
            self.teacher_model.to(self.device)
            self.teacher_model.eval()

        self.custom_loss_fn = None

    def set_custom_loss_fn(self, loss_fn):
        """Set custom loss function for specialized training."""
        self.custom_loss_fn = loss_fn

    def compute_distillation_loss(self, student_outputs, teacher_outputs, batch):
        """Compute standard distillation loss."""
        return self.loss_function.compute(
            student_outputs.logits,
            (
                teacher_outputs.logits
                if hasattr(teacher_outputs, "logits")
                else teacher_outputs
            ),
            labels=batch.get("labels"),
        )

    def train(self, dataloader: DataLoader, config: TrainingConfig) -> Dict[str, Any]:
        """
        Train the model with the given configuration.

        Args:
            dataloader: Training data loader
            config: Training configuration

        Returns:
            Training results and metrics
        """
        try:
            self.model.train()
            total_loss = 0.0
            num_batches = 0
            training_metrics = {}

            for epoch in range(config.num_epochs):
                epoch_loss = 0.0
                epoch_batches = 0

                for batch_idx, batch in enumerate(dataloader):
                    # Move batch to device
                    batch = self._move_batch_to_device(batch)

                    # Forward pass
                    self.optimizer.zero_grad()

                    # Student model forward pass
                    student_outputs = self.model(**batch)

                    # Teacher model forward pass (if available)
                    teacher_outputs = None
                    if self.teacher_model:
                        with torch.no_grad():
                            teacher_outputs = self.teacher_model(**batch)

                    # Compute loss
                    if self.custom_loss_fn:
                        loss = self.custom_loss_fn(
                            student_outputs, teacher_outputs, batch
                        )
                    else:
                        loss = self.loss_function.compute(
                            (
                                student_outputs.logits
                                if hasattr(student_outputs, "logits")
                                else student_outputs
                            ),
                            batch.get("labels", batch.get("input_ids")),
                        )

                    # Backward pass
                    loss.backward()
                    self.optimizer.step()

                    # Track metrics
                    epoch_loss += loss.item()
                    epoch_batches += 1

                    # Log training step
                    if batch_idx % config.save_steps == 0:
                        step_metrics = self.loss_function.get_metrics()
                        self._log_training_step(
                            epoch, batch_idx, loss.item(), step_metrics
                        )

                        logger.info(
                            f"Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item():.4f}"
                        )

                # End of epoch
                avg_epoch_loss = (
                    epoch_loss / epoch_batches if epoch_batches > 0 else 0.0
                )
                total_loss += epoch_loss
                num_batches += epoch_batches

                training_metrics[f"epoch_{epoch}_loss"] = avg_epoch_loss

                # Save checkpoint
                if epoch % config.save_steps == 0:
                    checkpoint_path = (
                        Path(config.output_dir) / f"checkpoint_epoch_{epoch}.pt"
                    )
                    self.save_checkpoint(checkpoint_path, epoch)

            # Final results
            avg_loss = total_loss / num_batches if num_batches > 0 else 0.0

            results = {
                "average_loss": avg_loss,
                "total_epochs": config.num_epochs,
                "total_batches": num_batches,
                "training_metrics": training_metrics,
                "loss_function_metrics": self.loss_function.get_metrics(),
            }

            logger.info(f"Training completed. Average loss: {avg_loss:.4f}")
            return results

        except Exception as e:
            raise TrainingError(
                f"Training failed: {str(e)}",
                "TRAINING_FAILED",
                {"epoch": getattr(self, "current_epoch", 0)},
            )

    def evaluate(self, dataloader: DataLoader) -> Dict[str, float]:
        """
        Evaluate the model on the given dataset.

        Args:
            dataloader: Evaluation data loader

        Returns:
            Evaluation metrics
        """
        try:
            self.model.eval()
            total_loss = 0.0
            num_batches = 0

            with torch.no_grad():
                for batch in dataloader:
                    batch = self._move_batch_to_device(batch)

                    # Forward pass
                    outputs = self.model(**batch)

                    # Compute loss
                    loss = self.loss_function.compute(
                        outputs.logits if hasattr(outputs, "logits") else outputs,
                        batch.get("labels", batch.get("input_ids")),
                    )

                    total_loss += loss.item()
                    num_batches += 1

            avg_loss = total_loss / num_batches if num_batches > 0 else 0.0

            results = {"eval_loss": avg_loss, "eval_batches": num_batches}
            results.update(self.loss_function.get_metrics())

            logger.info(f"Evaluation completed. Average loss: {avg_loss:.4f}")
            return results

        except Exception as e:
            raise TrainingError(f"Evaluation failed: {str(e)}", "EVALUATION_FAILED")

    def _move_batch_to_device(
        self, batch: Dict[str, torch.Tensor]
    ) -> Dict[str, torch.Tensor]:
        """Move batch tensors to the training device."""
        device_batch = {}
        for key, value in batch.items():
            if isinstance(value, torch.Tensor):
                device_batch[key] = value.to(self.device)
            else:
                device_batch[key] = value
        return device_batch