dataloader.py

import json
from pathlib import Path
import math

import torch
import torch.distributed as dist
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import Dataset, DataLoader, Sampler, WeightedRandomSampler
from torch.utils.data.distributed import DistributedSampler

from torchcodec.decoders import AudioDecoder

MELD_LABELS = ["anger", "disgust", "fear", "joy", "neutral", "sadness", "surprise"]
lab2id = {l: i for i, l in enumerate(MELD_LABELS)}


def load_audio_with_torchcodec(path: Path):
    """
    TorchCodec-native audio load: returns mono [T] and sample rate.
    """
    dec = AudioDecoder(str(path))
    samples = dec.get_all_samples()

    # TorchCodec AudioSamples object
    if hasattr(samples, "data") and hasattr(samples, "sample_rate"):
        waveform = samples.data
        sr = int(samples.sample_rate)
    # Fallback: (waveform, sr)
    elif isinstance(samples, (tuple, list)) and len(samples) == 2:
        waveform, sr = samples
    else:
        waveform = samples
        sr = getattr(dec, "sample_rate", None)
        if sr is None and hasattr(dec, "metadata"):
            sr = getattr(dec.metadata, "sample_rate", None)
        if sr is None:
            raise RuntimeError("Could not determine sample rate from TorchCodec AudioDecoder.")

    if not isinstance(waveform, torch.Tensor):
        waveform = torch.as_tensor(waveform)
    waveform = waveform.to(torch.float32).cpu()

    # [C, T]
    if waveform.ndim == 1:
        waveform = waveform.unsqueeze(0)

    # mono [T]
    mono = waveform.mean(dim=0).contiguous()
    return mono, int(sr)


class AudioDataset(Dataset):
    def __init__(self, root, split: str = "train"):
        self.root = Path(root)
        self.split = split

        candidate_dirs = [
            self.root / "data" / split,
            self.root / split,
        ]

        self.split_dir = None
        for cand in candidate_dirs:
            meta_candidate = cand / "meta.jsonl"
            if meta_candidate.exists():
                self.split_dir = cand
                break

        if self.split_dir is None:
            locations = ", ".join(str(c / "meta.jsonl") for c in candidate_dirs)
            raise FileNotFoundError(
                f"Could not locate meta.jsonl for split '{split}'. Checked: {locations}."
            )

        meta_file = self.split_dir / "meta.jsonl"
        with open(meta_file, "r", encoding="utf-8") as f:
            self.meta = [json.loads(line) for line in f if line.strip()]

        # Per-sample weights for class balancing: inverse frequency
        label_counts = {l: 0 for l in MELD_LABELS}
        for item in self.meta:
            label = item["Emotion"].lower()
            if label in lab2id:
                label_counts[label] += 1

        total = len(self.meta)
        weights = []
        for item in self.meta:
            label = item["Emotion"].lower()
            if label in lab2id and label_counts[label] > 0:
                weights.append(float(total) / float(label_counts[label]))
            else:
                weights.append(1.0)
        self.weights = torch.as_tensor(weights, dtype=torch.float)

    def __len__(self):
        return len(self.meta)

    def __getitem__(self, idx):
        item = self.meta[idx]

        audio_path = self.split_dir / item["audio"]
        mono, sr = load_audio_with_torchcodec(audio_path)

        text = item["Utterance"]
        label = item["Emotion"].lower()
        if label not in lab2id:
            raise ValueError(f"Unknown label {label}; expected one of {MELD_LABELS}")

        return {
            "audio": mono,
            "sr": int(sr),
            "text": text,
            "label_id": lab2id[label],
            "id": item.get("id"),
            "dialogue_id": item.get("Dialogue_ID"),
            "utterance_id": item.get("Utterance_ID"),
        }


class DistributedWeightedSampler(Sampler):
    """
    Weighted sampler that works with DDP.

    - Uses global `weights` to sample `total_size = num_replicas * num_samples_per_replica`.
    - Then shards indices by rank (same pattern as DistributedSampler).
    - Call `set_epoch(epoch)` in your training loop so each epoch changes.
    """

    def __init__(
        self,
        dataset,
        weights,
        num_replicas=None,
        rank=None,
        replacement=True,
        drop_last=False,
    ):
        if num_replicas is None:
            if not (dist.is_available() and dist.is_initialized()):
                raise RuntimeError(
                    "DistributedWeightedSampler requires DDP initialized or explicit num_replicas."
                )
            num_replicas = dist.get_world_size()

        if rank is None:
            if not (dist.is_available() and dist.is_initialized()):
                raise RuntimeError(
                    "DistributedWeightedSampler requires DDP initialized or explicit rank."
                )
            rank = dist.get_rank()

        self.dataset = dataset
        self.weights = torch.as_tensor(weights, dtype=torch.float)
        if self.weights.numel() != len(self.dataset):
            raise ValueError("weights length must match dataset length")

        self.num_replicas = int(num_replicas)
        self.rank = int(rank)
        self.replacement = bool(replacement)
        self.drop_last = bool(drop_last)

        if self.drop_last:
            self.num_samples = len(self.dataset) // self.num_replicas
        else:
            self.num_samples = int(math.ceil(len(self.dataset) / self.num_replicas))

        self.total_size = self.num_samples * self.num_replicas
        self.epoch = 0

    def __iter__(self):
        # Same seed on all ranks, different shard per rank.
        g = torch.Generator()
        g.manual_seed(self.epoch)

        if self.replacement:
            # Sample with replacement (recommended for rebalancing)
            indices = torch.multinomial(
                self.weights,
                self.total_size,
                replacement=True,
                generator=g,
            ).tolist()
        else:
            # Strict no-replacement: ensure we have enough data.
            if self.total_size > len(self.dataset):
                raise ValueError(
                    "DistributedWeightedSampler: total_size > dataset size with "
                    "replacement=False. Use replacement=True or drop_last=True."
                )
            indices = torch.multinomial(
                self.weights,
                self.total_size,
                replacement=False,
                generator=g,
            ).tolist()

        # Shard like DistributedSampler
        indices = indices[self.rank:self.total_size:self.num_replicas]
        assert len(indices) == self.num_samples

        return iter(indices)

    def __len__(self):
        return self.num_samples

    def set_epoch(self, epoch: int):
        self.epoch = int(epoch)


def meld_collate(batch):
    if not batch:
        raise ValueError("meld_collate received an empty batch")

    audio_tensors = [sample["audio"].to(dtype=torch.float32) for sample in batch]
    padded_audio = pad_sequence(audio_tensors, batch_first=True)
    audio_lengths = torch.tensor([tensor.size(-1) for tensor in audio_tensors], dtype=torch.long)

    sr_values = torch.tensor([int(sample["sr"]) for sample in batch], dtype=torch.long)
    labels = torch.tensor([int(sample["label_id"]) for sample in batch], dtype=torch.long)
    texts = [str(sample["text"]) for sample in batch]

    return {
        "audio": padded_audio,
        "audio_lengths": audio_lengths,
        "sr": sr_values,
        "text": texts,
        "label_id": labels,
        "id": [sample.get("id") for sample in batch],
        "dialogue_id": [sample.get("dialogue_id") for sample in batch],
        "utterance_id": [sample.get("utterance_id") for sample in batch],
    }


def build_dataloader(
    root,
    split: str = "train",
    batch_size: int = 8,
    num_workers: int = 0,
    collate_fn=None,
    shuffle=None,
    prefetch_factor: int = 4,
):
    ds = AudioDataset(root, split)

    if shuffle is None:
        shuffle = (split == "train")

    drop_last = (split == "train")
    is_distributed = dist.is_available() and dist.is_initialized()

    sampler = None

    if split == "train":
        if is_distributed:
            # Global class-balanced sampling across ranks
            sampler = DistributedWeightedSampler(
                dataset=ds,
                weights=ds.weights,
                replacement=True,
                drop_last=drop_last,
            )
            shuffle = False
        else:
            # Single-process balanced sampling
            sampler = WeightedRandomSampler(
                weights=ds.weights.tolist(),
                num_samples=len(ds),
                replacement=False,
            )
            shuffle = False
    else:
        if is_distributed:
            sampler = DistributedSampler(
                ds,
                shuffle=False,
                drop_last=drop_last,
            )
            shuffle = False
        else:
            sampler = None

    loader_kwargs = {
        "dataset": ds,
        "batch_size": batch_size,
        "num_workers": num_workers,
        "collate_fn": collate_fn,
        "pin_memory": True,
        "drop_last": drop_last,
    }

    if sampler is not None:
        loader_kwargs["sampler"] = sampler
        loader_kwargs["shuffle"] = False  # sampler controls ordering
    else:
        loader_kwargs["shuffle"] = shuffle

    if num_workers > 0:
        loader_kwargs["persistent_workers"] = True
        loader_kwargs["prefetch_factor"] = prefetch_factor

    if loader_kwargs.get("collate_fn") is None:
        loader_kwargs["collate_fn"] = meld_collate

    loader = DataLoader(**loader_kwargs)
    return loader, sampler