File size: 4,976 Bytes
d7013e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Seq2Seq Training Service.

Handles teacher-forced training for the Grassmann encoder-decoder model.
Uses cross-entropy loss instead of CTC, with proper token shifting
and padding masks.
"""

from __future__ import annotations

from dataclasses import dataclass

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

from src.models.config import GrassmannConfig
from src.models.grassmann_seq2seq import GrassmannSeq2SeqModel


@dataclass
class TrainMetrics:
    epoch: int
    train_loss: float
    val_loss: float | None = None


class Seq2SeqTrainingService:
    """Training loop for the fully attention-free encoder-decoder."""

    def __init__(
        self,
        model: GrassmannSeq2SeqModel,
        config: GrassmannConfig,
        device: torch.device | None = None,
    ):
        self.model = model
        self.config = config

        if device is not None:
            self.device = device
        elif torch.cuda.is_available():
            self.device = torch.device("cuda")
        else:
            self.device = torch.device("cpu")

        self.model.to(self.device)

        self.optimizer = torch.optim.AdamW(
            model.parameters(),
            lr=config.learning_rate,
            weight_decay=config.weight_decay,
        )
        self.scheduler: torch.optim.lr_scheduler.LRScheduler | None = None

        # Cross-entropy loss, ignoring padding (index 0)
        self.criterion = nn.CrossEntropyLoss(ignore_index=0, label_smoothing=0.1)
        self.history: list[TrainMetrics] = []

    def set_cosine_schedule(self, epochs: int, warmup_epochs: int = 2) -> None:
        from torch.optim.lr_scheduler import CosineAnnealingLR, LinearLR, SequentialLR
        warmup = LinearLR(self.optimizer, start_factor=0.1, total_iters=warmup_epochs)
        cosine = CosineAnnealingLR(self.optimizer, T_max=epochs - warmup_epochs, eta_min=1e-6)
        self.scheduler = SequentialLR(self.optimizer, [warmup, cosine], milestones=[warmup_epochs])

    def train_epoch(self, train_loader: DataLoader) -> float:
        self.model.train()
        total_loss = 0.0
        n_batches = 0
        total_batches = len(train_loader)

        for batch_idx, (mels, tokens, mel_lengths, token_lengths) in enumerate(train_loader):
            mels = mels.to(self.device)
            tokens = tokens.to(self.device)

            self.optimizer.zero_grad()

            # Teacher-forced forward pass
            logits = self.model(mels, tokens)  # (B, T_dec, vocab)

            # Reshape for cross-entropy: (B*T, vocab) vs (B*T,)
            B, T, V = logits.shape
            loss = self.criterion(logits.reshape(B * T, V), tokens.reshape(B * T))

            if torch.isfinite(loss):
                loss.backward()
                torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=5.0)
                self.optimizer.step()
                total_loss += loss.item()
                n_batches += 1

            # Log every 10 batches for CI visibility
            if (batch_idx + 1) % 10 == 0 or (batch_idx + 1) == total_batches:
                avg = total_loss / max(n_batches, 1)
                print(f"    batch {batch_idx+1}/{total_batches}  loss: {avg:.4f}", flush=True)

        return total_loss / max(n_batches, 1)

    @torch.no_grad()
    def validate(self, val_loader: DataLoader) -> float:
        self.model.eval()
        total_loss = 0.0
        n_batches = 0

        for mels, tokens, mel_lengths, token_lengths in val_loader:
            mels = mels.to(self.device)
            tokens = tokens.to(self.device)

            logits = self.model(mels, tokens)

            B, T, V = logits.shape
            loss = self.criterion(logits.reshape(B * T, V), tokens.reshape(B * T))

            if torch.isfinite(loss):
                total_loss += loss.item()
                n_batches += 1

        return total_loss / max(n_batches, 1)

    def train_loop(
        self,
        train_loader: DataLoader,
        val_loader: DataLoader | None,
        epochs: int,
        callback: object | None = None,
    ) -> list[TrainMetrics]:
        for epoch in range(1, epochs + 1):
            train_loss = self.train_epoch(train_loader)
            val_loss = self.validate(val_loader) if val_loader else None

            metrics = TrainMetrics(epoch=epoch, train_loss=train_loss, val_loss=val_loss)
            self.history.append(metrics)

            if callback and hasattr(callback, "on_epoch_end"):
                callback.on_epoch_end(metrics)

            if self.scheduler is not None:
                self.scheduler.step()

        return self.history

    @torch.no_grad()
    def greedy_decode(self, mel: torch.Tensor) -> torch.Tensor:
        """Autoregressive greedy decoding for a single mel spectrogram."""
        self.model.eval()
        if mel.dim() == 2:
            mel = mel.unsqueeze(0)
        mel = mel.to(self.device)
        return self.model.greedy_decode(mel, max_len=200)