lightning.py

import itertools
import math

from collections import namedtuple
from typing import List, Tuple

import sentencepiece as spm

import torch
import torchaudio
from models.conformer_rnnt import conformer_rnnt
from models.emformer_rnnt import emformer_rnnt
from models.resnet import video_resnet
from models.resnet1d import audio_resnet
from pytorch_lightning import LightningModule
from schedulers import WarmupCosineScheduler
from torchaudio.models import Hypothesis, RNNTBeamSearch

_expected_spm_vocab_size = 1023

Batch = namedtuple("Batch", ["inputs", "input_lengths", "targets", "target_lengths"])


def post_process_hypos(
    hypos: List[Hypothesis], sp_model: spm.SentencePieceProcessor
) -> List[Tuple[str, float, List[int], List[int]]]:
    tokens_idx = 0
    score_idx = 3
    post_process_remove_list = [
        sp_model.unk_id(),
        sp_model.eos_id(),
        sp_model.pad_id(),
    ]
    filtered_hypo_tokens = [
        [token_index for token_index in h[tokens_idx][1:] if token_index not in post_process_remove_list] for h in hypos
    ]
    hypos_str = [sp_model.decode(s) for s in filtered_hypo_tokens]
    hypos_ids = [h[tokens_idx][1:] for h in hypos]
    hypos_score = [[math.exp(h[score_idx])] for h in hypos]

    nbest_batch = list(zip(hypos_str, hypos_score, hypos_ids))

    return nbest_batch


class ConformerRNNTModule(LightningModule):
    def __init__(self, args=None, sp_model=None, pretrained_model_path=None):
        super().__init__()
        self.save_hyperparameters(args)
        self.args = args
        self.sp_model = sp_model
        spm_vocab_size = self.sp_model.get_piece_size()
        assert spm_vocab_size == _expected_spm_vocab_size, (
            "The model returned by conformer_rnnt_base expects a SentencePiece model of "
            f"vocabulary size {_expected_spm_vocab_size}, but the given SentencePiece model has a vocabulary size "
            f"of {spm_vocab_size}. Please provide a correctly configured SentencePiece model."
        )
        self.blank_idx = spm_vocab_size

        if args.md == "v":
            self.frontend = video_resnet()
        if args.md == "a":
            self.frontend = audio_resnet()

        if args.mode == "online":
            self.model = emformer_rnnt()
        if args.mode == "offline":
            self.model = conformer_rnnt()

        # -- initialise
        if args.pretrained_model_path:
            ckpt = torch.load(args.pretrained_model_path, map_location=lambda storage, loc: storage)
            tmp_ckpt = {
                k.replace("encoder.frontend.", ""): v for k, v in ckpt.items() if k.startswith("encoder.frontend.")
            }
            self.frontend.load_state_dict(tmp_ckpt)

        self.loss = torchaudio.transforms.RNNTLoss(reduction="sum")

        self.optimizer = torch.optim.AdamW(
            itertools.chain(*([self.frontend.parameters(), self.model.parameters()])),
            lr=8e-4,
            weight_decay=0.06,
            betas=(0.9, 0.98),
        )

        self.automatic_optimization = False

    def _step(self, batch, _, step_type):
        if batch is None:
            return None

        prepended_targets = batch.targets.new_empty([batch.targets.size(0), batch.targets.size(1) + 1])
        prepended_targets[:, 1:] = batch.targets
        prepended_targets[:, 0] = self.blank_idx
        prepended_target_lengths = batch.target_lengths + 1
        features = self.frontend(batch.inputs)
        output, src_lengths, _, _ = self.model(
            features, batch.input_lengths, prepended_targets, prepended_target_lengths
        )
        loss = self.loss(output, batch.targets, src_lengths, batch.target_lengths)
        self.log(f"Losses/{step_type}_loss", loss, on_step=True, on_epoch=True)

        return loss

    def configure_optimizers(self):
        self.warmup_lr_scheduler = WarmupCosineScheduler(
            self.optimizer,
            10,
            self.args.epochs,
            len(self.trainer.datamodule.train_dataloader()) / self.trainer.num_devices / self.trainer.num_nodes,
        )
        self.lr_scheduler_interval = "step"
        return (
            [self.optimizer],
            [{"scheduler": self.warmup_lr_scheduler, "interval": self.lr_scheduler_interval}],
        )

    def forward(self, batch: Batch):
        decoder = RNNTBeamSearch(self.model, self.blank_idx)
        x = self.frontend(batch.inputs.to(self.device))
        hypotheses = decoder(x, batch.input_lengths.to(self.device), beam_width=20)
        return post_process_hypos(hypotheses, self.sp_model)[0][0]

    def training_step(self, batch: Batch, batch_idx):
        """Custom training step.

        By default, DDP does the following on each train step:
        - For each GPU, compute loss and gradient on shard of training data.
        - Sync and average gradients across all GPUs. The final gradient
          is (sum of gradients across all GPUs) / N, where N is the world
          size (total number of GPUs).
        - Update parameters on each GPU.

        Here, we do the following:
        - For k-th GPU, compute loss and scale it by (N / B_total), where B_total is
          the sum of batch sizes across all GPUs. Compute gradient from scaled loss.
        - Sync and average gradients across all GPUs. The final gradient
          is (sum of gradients across all GPUs) / B_total.
        - Update parameters on each GPU.

        Doing so allows us to account for the variability in batch sizes that
        variable-length sequential data commonly yields.
        """

        opt = self.optimizers()
        opt.zero_grad()
        loss = self._step(batch, batch_idx, "train")
        batch_size = batch.inputs.size(0)
        batch_sizes = self.all_gather(batch_size)

        loss *= batch_sizes.size(0) / batch_sizes.sum()  # world size / batch size
        self.manual_backward(loss)
        torch.nn.utils.clip_grad_norm_(self.model.parameters(), 10)
        opt.step()

        sch = self.lr_schedulers()
        sch.step()

        self.log("monitoring_step", self.global_step)

        return loss

    def validation_step(self, batch, batch_idx):
        return self._step(batch, batch_idx, "val")

    def test_step(self, batch, batch_idx):
        return self._step(batch, batch_idx, "test")