train.py

#!/usr/bin/env python3
# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
#                                                  Wei Kang,
#                                                  Mingshuang Luo,)
#                                                  Zengwei Yao)
#              2023  Johns Hopkins University (author: Desh Raj)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Usage:

export CUDA_VISIBLE_DEVICES="0,1,2,3"

cd egs/libricss/SURT
./prepare.sh

./dprnn_zipformer/train.py \
  --world-size 4 \
  --num-epochs 30 \
  --start-epoch 1 \
  --exp-dir dprnn_zipformer/exp \
  --max-duration 300

# For mix precision training:

./dprnn_zipformer/train.py \
  --world-size 4 \
  --num-epochs 30 \
  --start-epoch 1 \
  --use-fp16 1 \
  --exp-dir dprnn_zipformer/exp \
  --max-duration 550
"""

import argparse
import copy
import logging
import warnings
from pathlib import Path
from shutil import copyfile
from typing import Any, Dict, Optional, Tuple, Union

import k2
import optim
import sentencepiece as spm
import torch
import torch.multiprocessing as mp
import torch.nn as nn
from asr_datamodule import LibriCssAsrDataModule
from decoder import Decoder
from dprnn import DPRNN
from einops.layers.torch import Rearrange
from joiner import Joiner
from lhotse.dataset.sampling.base import CutSampler
from lhotse.utils import LOG_EPSILON, fix_random_seed
from model import SURT
from optim import Eden, ScaledAdam
from scaling import ScaledLSTM
from torch import Tensor
from torch.cuda.amp import GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.tensorboard import SummaryWriter
from zipformer import Zipformer

from icefall import diagnostics
from icefall.checkpoint import load_checkpoint, remove_checkpoints
from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
from icefall.checkpoint import (
    save_checkpoint_with_global_batch_idx,
    update_averaged_model,
)
from icefall.dist import cleanup_dist, setup_dist
from icefall.env import get_env_info
from icefall.err import raise_grad_scale_is_too_small_error
from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool

LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler]


def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None:
    if isinstance(model, DDP):
        # get underlying nn.Module
        model = model.module
    for module in model.modules():
        if hasattr(module, "batch_count"):
            module.batch_count = batch_count


def add_model_arguments(parser: argparse.ArgumentParser):
    parser.add_argument(
        "--num-mask-encoder-layers",
        type=int,
        default=4,
        help="Number of layers in the DPRNN based mask encoder.",
    )

    parser.add_argument(
        "--mask-encoder-dim",
        type=int,
        default=256,
        help="Hidden dimension of the LSTM blocks in DPRNN.",
    )

    parser.add_argument(
        "--mask-encoder-segment-size",
        type=int,
        default=32,
        help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the "
        "decode-chunk-length of the zipformer encoder.",
    )

    parser.add_argument(
        "--chunk-width-randomization",
        type=bool,
        default=False,
        help="Whether to randomize the chunk width in DPRNN.",
    )

    # Zipformer config is based on:
    # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740
    parser.add_argument(
        "--num-encoder-layers",
        type=str,
        default="2,2,2,2,2",
        help="Number of zipformer encoder layers, comma separated.",
    )

    parser.add_argument(
        "--feedforward-dims",
        type=str,
        default="768,768,768,768,768",
        help="Feedforward dimension of the zipformer encoder layers, comma separated.",
    )

    parser.add_argument(
        "--nhead",
        type=str,
        default="8,8,8,8,8",
        help="Number of attention heads in the zipformer encoder layers.",
    )

    parser.add_argument(
        "--encoder-dims",
        type=str,
        default="256,256,256,256,256",
        help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated",
    )

    parser.add_argument(
        "--attention-dims",
        type=str,
        default="192,192,192,192,192",
        help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated;
        not the same as embedding dimension.""",
    )

    parser.add_argument(
        "--encoder-unmasked-dims",
        type=str,
        default="192,192,192,192,192",
        help="Unmasked dimensions in the encoders, relates to augmentation during training.  "
        "Must be <= each of encoder_dims.  Empirically, less than 256 seems to make performance "
        " worse.",
    )

    parser.add_argument(
        "--zipformer-downsampling-factors",
        type=str,
        default="1,2,4,8,2",
        help="Downsampling factor for each stack of encoder layers.",
    )

    parser.add_argument(
        "--cnn-module-kernels",
        type=str,
        default="31,31,31,31,31",
        help="Sizes of kernels in convolution modules",
    )

    parser.add_argument(
        "--use-joint-encoder-layer",
        type=str,
        default="lstm",
        choices=["linear", "lstm", "none"],
        help="Whether to use a joint layer to combine all branches.",
    )

    parser.add_argument(
        "--decoder-dim",
        type=int,
        default=512,
        help="Embedding dimension in the decoder model.",
    )

    parser.add_argument(
        "--joiner-dim",
        type=int,
        default=512,
        help="""Dimension used in the joiner model.
        Outputs from the encoder and decoder model are projected
        to this dimension before adding.
        """,
    )

    parser.add_argument(
        "--short-chunk-size",
        type=int,
        default=50,
        help="""Chunk length of dynamic training, the chunk size would be either
        max sequence length of current batch or uniformly sampled from (1, short_chunk_size).
        """,
    )

    parser.add_argument(
        "--num-left-chunks",
        type=int,
        default=4,
        help="How many left context can be seen in chunks when calculating attention.",
    )

    parser.add_argument(
        "--decode-chunk-len",
        type=int,
        default=32,
        help="The chunk size for decoding (in frames before subsampling)",
    )


def get_parser():
    parser = argparse.ArgumentParser(
        formatter_class=argparse.ArgumentDefaultsHelpFormatter
    )

    parser.add_argument(
        "--world-size",
        type=int,
        default=1,
        help="Number of GPUs for DDP training.",
    )

    parser.add_argument(
        "--master-port",
        type=int,
        default=12354,
        help="Master port to use for DDP training.",
    )

    parser.add_argument(
        "--tensorboard",
        type=str2bool,
        default=True,
        help="Should various information be logged in tensorboard.",
    )

    parser.add_argument(
        "--num-epochs",
        type=int,
        default=30,
        help="Number of epochs to train.",
    )

    parser.add_argument(
        "--start-epoch",
        type=int,
        default=1,
        help="""Resume training from this epoch. It should be positive.
        If larger than 1, it will load checkpoint from
        exp-dir/epoch-{start_epoch-1}.pt
        """,
    )

    parser.add_argument(
        "--start-batch",
        type=int,
        default=0,
        help="""If positive, --start-epoch is ignored and
        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
        """,
    )

    parser.add_argument(
        "--exp-dir",
        type=str,
        default="conv_lstm_transducer_stateless_ctc/exp",
        help="""The experiment dir.
        It specifies the directory where all training related
        files, e.g., checkpoints, log, etc, are saved
        """,
    )

    parser.add_argument(
        "--model-init-ckpt",
        type=str,
        default=None,
        help="""The model checkpoint to initialize the model (either full or part).
        If not specified, the model is randomly initialized.
        """,
    )

    parser.add_argument(
        "--bpe-model",
        type=str,
        default="data/lang_bpe_500/bpe.model",
        help="Path to the BPE model",
    )

    parser.add_argument(
        "--base-lr", type=float, default=0.004, help="The base learning rate."
    )

    parser.add_argument(
        "--lr-batches",
        type=float,
        default=5000,
        help="""Number of steps that affects how rapidly the learning rate
        decreases. We suggest not to change this.""",
    )

    parser.add_argument(
        "--lr-epochs",
        type=float,
        default=6,
        help="""Number of epochs that affects how rapidly the learning rate decreases.
        """,
    )

    parser.add_argument(
        "--context-size",
        type=int,
        default=2,
        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
    )

    parser.add_argument(
        "--prune-range",
        type=int,
        default=5,
        help="The prune range for rnnt loss, it means how many symbols(context)"
        "we are using to compute the loss",
    )

    parser.add_argument(
        "--lm-scale",
        type=float,
        default=0.25,
        help="The scale to smooth the loss with lm "
        "(output of prediction network) part.",
    )

    parser.add_argument(
        "--am-scale",
        type=float,
        default=0.0,
        help="The scale to smooth the loss with am (output of encoder network) part.",
    )

    parser.add_argument(
        "--simple-loss-scale",
        type=float,
        default=0.5,
        help="To get pruning ranges, we will calculate a simple version"
        "loss(joiner is just addition), this simple loss also uses for"
        "training (as a regularization item). We will scale the simple loss"
        "with this parameter before adding to the final loss.",
    )

    parser.add_argument(
        "--ctc-loss-scale",
        type=float,
        default=0.2,
        help="Scale for CTC loss.",
    )

    parser.add_argument(
        "--heat-loss-scale",
        type=float,
        default=0.0,
        help="Scale for HEAT loss on separated sources.",
    )

    parser.add_argument(
        "--seed",
        type=int,
        default=42,
        help="The seed for random generators intended for reproducibility",
    )

    parser.add_argument(
        "--print-diagnostics",
        type=str2bool,
        default=False,
        help="Accumulate stats on activations, print them and exit.",
    )

    parser.add_argument(
        "--save-every-n",
        type=int,
        default=2000,
        help="""Save checkpoint after processing this number of batches"
        periodically. We save checkpoint to exp-dir/ whenever
        params.batch_idx_train % save_every_n == 0. The checkpoint filename
        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
        end of each epoch where `xxx` is the epoch number counting from 0.
        """,
    )

    parser.add_argument(
        "--keep-last-k",
        type=int,
        default=1,
        help="""Only keep this number of checkpoints on disk.
        For instance, if it is 3, there are only 3 checkpoints
        in the exp-dir with filenames `checkpoint-xxx.pt`.
        It does not affect checkpoints with name `epoch-xxx.pt`.
        """,
    )

    parser.add_argument(
        "--average-period",
        type=int,
        default=100,
        help="""Update the averaged model, namely `model_avg`, after processing
        this number of batches. `model_avg` is a separate version of model,
        in which each floating-point parameter is the average of all the
        parameters from the start of training. Each time we take the average,
        we do: `model_avg = model * (average_period / batch_idx_train) +
            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
        """,
    )

    parser.add_argument(
        "--use-fp16",
        type=str2bool,
        default=False,
        help="Whether to use half precision training.",
    )

    add_model_arguments(parser)

    return parser


def get_params() -> AttributeDict:
    """Return a dict containing training parameters.

    All training related parameters that are not passed from the commandline
    are saved in the variable `params`.

    Commandline options are merged into `params` after they are parsed, so
    you can also access them via `params`.

    Explanation of options saved in `params`:

        - best_train_loss: Best training loss so far. It is used to select
                           the model that has the lowest training loss. It is
                           updated during the training.

        - best_valid_loss: Best validation loss so far. It is used to select
                           the model that has the lowest validation loss. It is
                           updated during the training.

        - best_train_epoch: It is the epoch that has the best training loss.

        - best_valid_epoch: It is the epoch that has the best validation loss.

        - batch_idx_train: Used to writing statistics to tensorboard. It
                           contains number of batches trained so far across
                           epochs.

        - log_interval:  Print training loss if batch_idx % log_interval` is 0

        - reset_interval: Reset statistics if batch_idx % reset_interval is 0

        - valid_interval:  Run validation if batch_idx % valid_interval is 0

        - feature_dim: The model input dim. It has to match the one used
                       in computing features.

        - subsampling_factor:  The subsampling factor for the model.

        - num_decoder_layers: Number of decoder layer of transformer decoder.

        - warm_step: The warm_step for Noam optimizer.
    """
    params = AttributeDict(
        {
            "best_train_loss": float("inf"),
            "best_valid_loss": float("inf"),
            "best_train_epoch": -1,
            "best_valid_epoch": -1,
            "batch_idx_train": 0,
            "log_interval": 50,
            "reset_interval": 200,
            "valid_interval": 2000,
            # parameters for SURT
            "num_channels": 2,
            "feature_dim": 80,
            "subsampling_factor": 4,  # not passed in, this is fixed
            # parameters for Noam
            "model_warm_step": 5000,  # arg given to model, not for lrate
            # parameters for ctc loss
            "beam_size": 10,
            "use_double_scores": True,
            "env_info": get_env_info(),
        }
    )

    return params


def get_mask_encoder_model(params: AttributeDict) -> nn.Module:
    mask_encoder = DPRNN(
        feature_dim=params.feature_dim,
        input_size=params.mask_encoder_dim,
        hidden_size=params.mask_encoder_dim,
        output_size=params.feature_dim * params.num_channels,
        segment_size=params.mask_encoder_segment_size,
        num_blocks=params.num_mask_encoder_layers,
        chunk_width_randomization=params.chunk_width_randomization,
    )
    return mask_encoder


def get_encoder_model(params: AttributeDict) -> nn.Module:
    # TODO: We can add an option to switch between Zipformer and Transformer
    def to_int_tuple(s: str):
        return tuple(map(int, s.split(",")))

    encoder = Zipformer(
        num_features=params.feature_dim,
        output_downsampling_factor=2,
        zipformer_downsampling_factors=to_int_tuple(
            params.zipformer_downsampling_factors
        ),
        encoder_dims=to_int_tuple(params.encoder_dims),
        attention_dim=to_int_tuple(params.attention_dims),
        encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims),
        nhead=to_int_tuple(params.nhead),
        feedforward_dim=to_int_tuple(params.feedforward_dims),
        cnn_module_kernels=to_int_tuple(params.cnn_module_kernels),
        num_encoder_layers=to_int_tuple(params.num_encoder_layers),
        num_left_chunks=params.num_left_chunks,
        short_chunk_size=params.short_chunk_size,
        decode_chunk_size=params.decode_chunk_len // 2,
    )
    return encoder


def get_joint_encoder_layer(params: AttributeDict) -> nn.Module:
    class TakeFirst(nn.Module):
        def forward(self, x):
            return x[0]

    if params.use_joint_encoder_layer == "linear":
        encoder_dim = int(params.encoder_dims.split(",")[-1])
        joint_layer = nn.Sequential(
            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
            nn.Linear(
                params.num_channels * encoder_dim, params.num_channels * encoder_dim
            ),
            nn.ReLU(),
            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
        )
    elif params.use_joint_encoder_layer == "lstm":
        encoder_dim = int(params.encoder_dims.split(",")[-1])
        joint_layer = nn.Sequential(
            Rearrange("(c b) t d -> b t (c d)", c=params.num_channels),
            ScaledLSTM(
                input_size=params.num_channels * encoder_dim,
                hidden_size=params.num_channels * encoder_dim,
                num_layers=1,
                bias=True,
                batch_first=True,
                dropout=0.0,
                bidirectional=False,
            ),
            TakeFirst(),
            nn.ReLU(),
            Rearrange("b t (c d) -> (c b) t d", c=params.num_channels),
        )
    elif params.use_joint_encoder_layer == "none":
        joint_layer = None
    else:
        raise ValueError(
            f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}"
        )
    return joint_layer


def get_decoder_model(params: AttributeDict) -> nn.Module:
    decoder = Decoder(
        vocab_size=params.vocab_size,
        decoder_dim=params.decoder_dim,
        blank_id=params.blank_id,
        context_size=params.context_size,
    )
    return decoder


def get_joiner_model(params: AttributeDict) -> nn.Module:
    joiner = Joiner(
        encoder_dim=int(params.encoder_dims.split(",")[-1]),
        decoder_dim=params.decoder_dim,
        joiner_dim=params.joiner_dim,
        vocab_size=params.vocab_size,
    )
    return joiner


def get_surt_model(
    params: AttributeDict,
) -> nn.Module:
    mask_encoder = get_mask_encoder_model(params)
    encoder = get_encoder_model(params)
    joint_layer = get_joint_encoder_layer(params)
    decoder = get_decoder_model(params)
    joiner = get_joiner_model(params)

    model = SURT(
        mask_encoder=mask_encoder,
        encoder=encoder,
        joint_encoder_layer=joint_layer,
        decoder=decoder,
        joiner=joiner,
        num_channels=params.num_channels,
        encoder_dim=int(params.encoder_dims.split(",")[-1]),
        decoder_dim=params.decoder_dim,
        joiner_dim=params.joiner_dim,
        vocab_size=params.vocab_size,
    )
    return model


def load_checkpoint_if_available(
    params: AttributeDict,
    model: nn.Module,
    model_avg: nn.Module = None,
    optimizer: Optional[torch.optim.Optimizer] = None,
    scheduler: Optional[LRSchedulerType] = None,
) -> Optional[Dict[str, Any]]:
    """Load checkpoint from file.

    If params.start_batch is positive, it will load the checkpoint from
    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
    params.start_epoch is larger than 1, it will load the checkpoint from
    `params.start_epoch - 1`.

    Apart from loading state dict for `model` and `optimizer` it also updates
    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
    and `best_valid_loss` in `params`.

    Args:
      params:
        The return value of :func:`get_params`.
      model:
        The training model.
      model_avg:
        The stored model averaged from the start of training.
      optimizer:
        The optimizer that we are using.
      scheduler:
        The scheduler that we are using.
    Returns:
      Return a dict containing previously saved training info.
    """
    if params.start_batch > 0:
        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
    elif params.start_epoch > 1:
        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
    else:
        return None

    assert filename.is_file(), f"{filename} does not exist!"

    saved_params = load_checkpoint(
        filename,
        model=model,
        model_avg=model_avg,
        optimizer=optimizer,
        scheduler=scheduler,
    )

    keys = [
        "best_train_epoch",
        "best_valid_epoch",
        "batch_idx_train",
        "best_train_loss",
        "best_valid_loss",
    ]
    for k in keys:
        params[k] = saved_params[k]

    if params.start_batch > 0:
        if "cur_epoch" in saved_params:
            params["start_epoch"] = saved_params["cur_epoch"]

    return saved_params


def save_checkpoint(
    params: AttributeDict,
    model: Union[nn.Module, DDP],
    model_avg: Optional[nn.Module] = None,
    optimizer: Optional[torch.optim.Optimizer] = None,
    scheduler: Optional[LRSchedulerType] = None,
    sampler: Optional[CutSampler] = None,
    scaler: Optional[GradScaler] = None,
    rank: int = 0,
) -> None:
    """Save model, optimizer, scheduler and training stats to file.

    Args:
      params:
        It is returned by :func:`get_params`.
      model:
        The training model.
      model_avg:
        The stored model averaged from the start of training.
      optimizer:
        The optimizer used in the training.
      sampler:
       The sampler for the training dataset.
      scaler:
        The scaler used for mix precision training.
    """
    if rank != 0:
        return
    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
    save_checkpoint_impl(
        filename=filename,
        model=model,
        model_avg=model_avg,
        params=params,
        optimizer=optimizer,
        scheduler=scheduler,
        sampler=sampler,
        scaler=scaler,
        rank=rank,
    )

    if params.best_train_epoch == params.cur_epoch:
        best_train_filename = params.exp_dir / "best-train-loss.pt"
        copyfile(src=filename, dst=best_train_filename)

    if params.best_valid_epoch == params.cur_epoch:
        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
        copyfile(src=filename, dst=best_valid_filename)


def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor:
    """
    Compute HEAT loss for separated sources using the output of mask encoder.
    Args:
      x_masked:
        The output of mask encoder. It is a tensor of shape (B, T, C).
      batch:
        A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()`
        for the content in it.
      num_channels:
        The number of output branches in the SURT model.
    """
    B, T, D = x_masked[0].shape
    device = x_masked[0].device

    # Create training targets for each channel.
    targets = []
    for i in range(num_channels):
        target = torch.ones_like(x_masked[i]) * LOG_EPSILON
        targets.append(target)

    source_feats = batch["source_feats"]
    source_boundaries = batch["source_boundaries"]
    input_lens = batch["input_lens"].to(device)
    # Assign sources to channels based on the HEAT criteria
    for b in range(B):
        cut_source_feats = source_feats[b]
        cut_source_boundaries = source_boundaries[b]
        last_seg_end = [0 for _ in range(num_channels)]
        for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries):
            assigned = False
            for i in range(num_channels):
                if start >= last_seg_end[i]:
                    targets[i][b, start:end, :] += source_feat.to(device)
                    last_seg_end[i] = max(end, last_seg_end[i])
                    assigned = True
                    break
            if not assigned:
                min_end_channel = last_seg_end.index(min(last_seg_end))
                targets[min_end_channel][b, start:end, :] += source_feat
                last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel])

    # Get padding mask based on input lengths
    pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1)
    pad_mask = pad_mask.unsqueeze(-1)

    # Compute masked loss for each channel
    losses = torch.zeros((num_channels, B, T, D), device=device)
    for i in range(num_channels):
        loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none")
        # Apply padding mask to loss
        loss.masked_fill_(pad_mask, 0)
        losses[i] = loss

    # loss: C x B x T x D. pad_mask: B x T x 1
    # We want to compute loss for each item in the batch. Each item has loss given
    # by the sum over C, and average over T and D. For T, we need to use the padding.
    loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device)
    return loss


def compute_loss(
    params: AttributeDict,
    model: Union[nn.Module, DDP],
    sp: spm.SentencePieceProcessor,
    batch: dict,
    is_training: bool,
) -> Tuple[Tensor, MetricsTracker]:
    """
    Compute RNN-T loss given the model and its inputs.

    Args:
      params:
        Parameters for training. See :func:`get_params`.
      model:
        The model for training. It is an instance of Conformer in our case.
      batch:
        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
        for the content in it.
      is_training:
        True for training. False for validation. When it is True, this
        function enables autograd during computation; when it is False, it
        disables autograd.
    """
    device = model.device if isinstance(model, DDP) else next(model.parameters()).device
    feature = batch["inputs"].to(device)
    feature_lens = batch["input_lens"].to(device)

    # at entry, feature is (N, T, C)
    assert feature.ndim == 3

    # The dataloader returns text as a list of cuts, each of which is a list of channel
    # text. We flatten this to a list where all channels are together, i.e., it looks like
    # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2].
    text = [val for tup in zip(*batch["text"]) for val in tup]
    assert len(text) == len(feature) * params.num_channels

    # Convert all channel texts to token IDs and create a ragged tensor.
    y = sp.encode(text, out_type=int)
    y = k2.RaggedTensor(y).to(device)

    batch_idx_train = params.batch_idx_train
    warm_step = params.model_warm_step

    with torch.set_grad_enabled(is_training):
        (simple_loss, pruned_loss, ctc_loss, x_masked) = model(
            x=feature,
            x_lens=feature_lens,
            y=y,
            prune_range=params.prune_range,
            am_scale=params.am_scale,
            lm_scale=params.lm_scale,
            reduction="none",
            subsampling_factor=params.subsampling_factor,
        )
        simple_loss_is_finite = torch.isfinite(simple_loss)
        pruned_loss_is_finite = torch.isfinite(pruned_loss)
        ctc_loss_is_finite = torch.isfinite(ctc_loss)

        # Compute HEAT loss
        if is_training and params.heat_loss_scale > 0.0:
            heat_loss = compute_heat_loss(
                x_masked, batch, num_channels=params.num_channels
            )
        else:
            heat_loss = torch.tensor(0.0, device=device)

        heat_loss_is_finite = torch.isfinite(heat_loss)
        is_finite = (
            simple_loss_is_finite
            & pruned_loss_is_finite
            & ctc_loss_is_finite
            & heat_loss_is_finite
        )
        if not torch.all(is_finite):
            logging.info(
                "Not all losses are finite!\n"
                f"simple_losses: {simple_loss}\n"
                f"pruned_losses: {pruned_loss}\n"
                f"ctc_losses: {ctc_loss}\n"
                f"heat_losses: {heat_loss}\n"
            )
            display_and_save_batch(batch, params=params, sp=sp)
            simple_loss = simple_loss[simple_loss_is_finite]
            pruned_loss = pruned_loss[pruned_loss_is_finite]
            ctc_loss = ctc_loss[ctc_loss_is_finite]
            heat_loss = heat_loss[heat_loss_is_finite]

            # If either all simple_loss or pruned_loss is inf or nan,
            # we stop the training process by raising an exception
            if (
                torch.all(~simple_loss_is_finite)
                or torch.all(~pruned_loss_is_finite)
                or torch.all(~ctc_loss_is_finite)
                or torch.all(~heat_loss_is_finite)
            ):
                raise ValueError(
                    "There are too many utterances in this batch "
                    "leading to inf or nan losses."
                )

        simple_loss_sum = simple_loss.sum()
        pruned_loss_sum = pruned_loss.sum()
        ctc_loss_sum = ctc_loss.sum()
        heat_loss_sum = heat_loss.sum()

        s = params.simple_loss_scale
        # take down the scale on the simple loss from 1.0 at the start
        # to params.simple_loss scale by warm_step.
        simple_loss_scale = (
            s
            if batch_idx_train >= warm_step
            else 1.0 - (batch_idx_train / warm_step) * (1.0 - s)
        )
        pruned_loss_scale = (
            1.0
            if batch_idx_train >= warm_step
            else 0.1 + 0.9 * (batch_idx_train / warm_step)
        )
        loss = (
            simple_loss_scale * simple_loss_sum
            + pruned_loss_scale * pruned_loss_sum
            + params.ctc_loss_scale * ctc_loss_sum
            + params.heat_loss_scale * heat_loss_sum
        )

    assert loss.requires_grad == is_training

    info = MetricsTracker()
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        # info["frames"] is an approximate number for two reasons:
        # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2
        # (2) If some utterances in the batch lead to inf/nan loss, they
        #     are filtered out.
        info["frames"] = (feature_lens // params.subsampling_factor).sum().item()

    # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances`  # noqa
    info["utterances"] = feature.size(0)
    # averaged input duration in frames over utterances
    info["utt_duration"] = feature_lens.sum().item()
    # averaged padding proportion over utterances
    info["utt_pad_proportion"] = (
        ((feature.size(1) - feature_lens) / feature.size(1)).sum().item()
    )

    # Note: We use reduction=sum while computing the loss.
    info["loss"] = loss.detach().cpu().item()
    info["simple_loss"] = simple_loss_sum.detach().cpu().item()
    info["pruned_loss"] = pruned_loss_sum.detach().cpu().item()
    if params.ctc_loss_scale > 0.0:
        info["ctc_loss"] = ctc_loss_sum.detach().cpu().item()
    if params.heat_loss_scale > 0.0:
        info["heat_loss"] = heat_loss_sum.detach().cpu().item()

    return loss, info


def compute_validation_loss(
    params: AttributeDict,
    model: Union[nn.Module, DDP],
    sp: spm.SentencePieceProcessor,
    valid_dl: torch.utils.data.DataLoader,
    world_size: int = 1,
) -> MetricsTracker:
    """Run the validation process."""
    model.eval()

    tot_loss = MetricsTracker()

    for batch_idx, batch in enumerate(valid_dl):
        loss, loss_info = compute_loss(
            params=params,
            model=model,
            sp=sp,
            batch=batch,
            is_training=False,
        )
        assert loss.requires_grad is False
        tot_loss = tot_loss + loss_info

    if world_size > 1:
        tot_loss.reduce(loss.device)

    loss_value = tot_loss["loss"] / tot_loss["frames"]
    if loss_value < params.best_valid_loss:
        params.best_valid_epoch = params.cur_epoch
        params.best_valid_loss = loss_value

    return tot_loss


def train_one_epoch(
    params: AttributeDict,
    model: Union[nn.Module, DDP],
    optimizer: torch.optim.Optimizer,
    scheduler: LRSchedulerType,
    sp: spm.SentencePieceProcessor,
    train_dl: torch.utils.data.DataLoader,
    train_dl_warmup: Optional[torch.utils.data.DataLoader],
    valid_dl: torch.utils.data.DataLoader,
    scaler: GradScaler,
    model_avg: Optional[nn.Module] = None,
    tb_writer: Optional[SummaryWriter] = None,
    world_size: int = 1,
    rank: int = 0,
) -> None:
    """Train the model for one epoch.

    The training loss from the mean of all frames is saved in
    `params.train_loss`. It runs the validation process every
    `params.valid_interval` batches.

    Args:
      params:
        It is returned by :func:`get_params`.
      model:
        The model for training.
      optimizer:
        The optimizer we are using.
      scheduler:
        The learning rate scheduler, we call step() every step.
      train_dl:
        Dataloader for the training dataset.
      train_dl_warmup:
        Dataloader for the training dataset with 2 speakers. This is used during the
        warmup stage.
      valid_dl:
        Dataloader for the validation dataset.
      scaler:
        The scaler used for mix precision training.
      model_avg:
        The stored model averaged from the start of training.
      tb_writer:
        Writer to write log messages to tensorboard.
      world_size:
        Number of nodes in DDP training. If it is 1, DDP is disabled.
      rank:
        The rank of the node in DDP training. If no DDP is used, it should
        be set to 0.
    """
    torch.cuda.empty_cache()
    model.train()

    tot_loss = MetricsTracker()

    iter_train = iter(train_dl)
    iter_train_warmup = iter(train_dl_warmup) if train_dl_warmup is not None else None

    batch_idx = 0

    while True:
        # We first sample a batch from the main dataset. This is because we want to
        # make sure all epochs have the same number of batches.
        try:
            batch = next(iter_train)
        except StopIteration:
            break

        # If we are in warmup stage, get the batch from the warmup dataset.
        if (
            params.batch_idx_train <= params.model_warm_step
            and iter_train_warmup is not None
        ):
            try:
                batch = next(iter_train_warmup)
            except StopIteration:
                iter_train_warmup = iter(train_dl_warmup)
                batch = next(iter_train_warmup)

        batch_idx += 1

        params.batch_idx_train += 1
        batch_size = batch["inputs"].shape[0]

        try:
            with torch.cuda.amp.autocast(enabled=params.use_fp16):
                loss, loss_info = compute_loss(
                    params=params,
                    model=model,
                    sp=sp,
                    batch=batch,
                    is_training=True,
                )
            # summary stats
            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info

            # NOTE: We use reduction==sum and loss is computed over utterances
            # in the batch and there is no normalization to it so far.
            scaler.scale(loss).backward()
            set_batch_count(model, params.batch_idx_train)
            scheduler.step_batch(params.batch_idx_train)

            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad()
        except:  # noqa
            display_and_save_batch(batch, params=params, sp=sp)
            raise

        if params.print_diagnostics and batch_idx == 5:
            return

        if (
            rank == 0
            and params.batch_idx_train > 0
            and params.batch_idx_train % params.average_period == 0
        ):
            update_averaged_model(
                params=params,
                model_cur=model,
                model_avg=model_avg,
            )

        if (
            params.batch_idx_train > 0
            and params.batch_idx_train % params.save_every_n == 0
        ):
            params.cur_batch_idx = batch_idx
            save_checkpoint_with_global_batch_idx(
                out_dir=params.exp_dir,
                global_batch_idx=params.batch_idx_train,
                model=model,
                model_avg=model_avg,
                params=params,
                optimizer=optimizer,
                scheduler=scheduler,
                sampler=train_dl.sampler,
                scaler=scaler,
                rank=rank,
            )
            del params.cur_batch_idx
            remove_checkpoints(
                out_dir=params.exp_dir,
                topk=params.keep_last_k,
                rank=rank,
            )

        if batch_idx % 100 == 0 and params.use_fp16:
            # If the grad scale was less than 1, try increasing it.    The _growth_interval
            # of the grad scaler is configurable, but we can't configure it to have different
            # behavior depending on the current grad scale.
            cur_grad_scale = scaler._scale.item()
            if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0):
                scaler.update(cur_grad_scale * 2.0)
            if cur_grad_scale < 0.01:
                logging.warning(f"Grad scale is small: {cur_grad_scale}")
            if cur_grad_scale < 1.0e-05:
                raise_grad_scale_is_too_small_error(cur_grad_scale)

        if batch_idx % params.log_interval == 0:
            cur_lr = scheduler.get_last_lr()[0]
            cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0

            logging.info(
                f"Epoch {params.cur_epoch}, "
                f"batch {batch_idx}, loss[{loss_info}], "
                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
                f"lr: {cur_lr:.2e}, "
                + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "")
            )

            if tb_writer is not None:
                tb_writer.add_scalar(
                    "train/learning_rate", cur_lr, params.batch_idx_train
                )

                loss_info.write_summary(
                    tb_writer, "train/current_", params.batch_idx_train
                )
                tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train)
                if params.use_fp16:
                    tb_writer.add_scalar(
                        "train/grad_scale", cur_grad_scale, params.batch_idx_train
                    )

        if batch_idx % params.valid_interval == 0 and not params.print_diagnostics:
            logging.info("Computing validation loss")
            valid_info = compute_validation_loss(
                params=params,
                model=model,
                sp=sp,
                valid_dl=valid_dl,
                world_size=world_size,
            )
            model.train()
            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
            logging.info(
                f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB"
            )
            if tb_writer is not None:
                valid_info.write_summary(
                    tb_writer, "train/valid_", params.batch_idx_train
                )

    loss_value = tot_loss["loss"] / tot_loss["frames"]
    params.train_loss = loss_value
    if params.train_loss < params.best_train_loss:
        params.best_train_epoch = params.cur_epoch
        params.best_train_loss = params.train_loss


def run(rank, world_size, args):
    """
    Args:
      rank:
        It is a value between 0 and `world_size-1`, which is
        passed automatically by `mp.spawn()` in :func:`main`.
        The node with rank 0 is responsible for saving checkpoint.
      world_size:
        Number of GPUs for DDP training.
      args:
        The return value of get_parser().parse_args()
    """
    params = get_params()
    params.update(vars(args))

    fix_random_seed(params.seed)
    if world_size > 1:
        setup_dist(rank, world_size, params.master_port)

    setup_logger(f"{params.exp_dir}/log/log-train")
    logging.info("Training started")

    if args.tensorboard and rank == 0:
        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
    else:
        tb_writer = None

    device = torch.device("cpu")
    if torch.cuda.is_available():
        device = torch.device("cuda", rank)
    logging.info(f"Device: {device}")

    sp = spm.SentencePieceProcessor()
    sp.load(params.bpe_model)

    # <blk> is defined in local/train_bpe_model.py
    params.blank_id = sp.piece_to_id("<blk>")
    params.vocab_size = sp.get_piece_size()

    logging.info(params)

    logging.info("About to create model")
    model = get_surt_model(params)

    num_param = sum([p.numel() for p in model.parameters()])
    logging.info(f"Number of model parameters: {num_param}")

    assert params.save_every_n >= params.average_period
    model_avg: Optional[nn.Module] = None
    if rank == 0:
        # model_avg is only used with rank 0
        model_avg = copy.deepcopy(model)

    assert params.start_epoch > 0, params.start_epoch
    checkpoints = load_checkpoint_if_available(
        params=params, model=model, model_avg=model_avg
    )

    model.to(device)

    if checkpoints is None and params.model_init_ckpt is not None:
        logging.info(
            f"Initializing model with checkpoint from {params.model_init_ckpt}"
        )
        init_ckpt = torch.load(params.model_init_ckpt, map_location=device)
        model.load_state_dict(init_ckpt["model"], strict=False)

    if world_size > 1:
        logging.info("Using DDP")
        model = DDP(model, device_ids=[rank], find_unused_parameters=True)

    parameters_names = []
    parameters_names.append(
        [name_param_pair[0] for name_param_pair in model.named_parameters()]
    )
    optimizer = ScaledAdam(
        model.parameters(),
        lr=params.base_lr,
        clipping_scale=2.0,
        parameters_names=parameters_names,
    )

    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)

    if checkpoints and "optimizer" in checkpoints:
        logging.info("Loading optimizer state dict")
        optimizer.load_state_dict(checkpoints["optimizer"])

    if (
        checkpoints
        and "scheduler" in checkpoints
        and checkpoints["scheduler"] is not None
    ):
        logging.info("Loading scheduler state dict")
        scheduler.load_state_dict(checkpoints["scheduler"])

    if params.print_diagnostics:
        diagnostic = diagnostics.attach_diagnostics(model)

    libricss = LibriCssAsrDataModule(args)

    train_cuts = libricss.lsmix_cuts(rvb_affix="comb", type_affix="full", sources=True)
    train_cuts_ov40 = libricss.lsmix_cuts(
        rvb_affix="comb", type_affix="ov40", sources=True
    )
    dev_cuts = libricss.libricss_cuts(split="dev", type="sdm")

    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
        # We only load the sampler's state dict when it loads a checkpoint
        # saved in the middle of an epoch
        sampler_state_dict = checkpoints["sampler"]
    else:
        sampler_state_dict = None

    train_dl = libricss.train_dataloaders(
        train_cuts,
        sampler_state_dict=sampler_state_dict,
    )
    train_dl_ov40 = libricss.train_dataloaders(train_cuts_ov40)
    valid_dl = libricss.valid_dataloaders(dev_cuts)

    scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0)
    if checkpoints and "grad_scaler" in checkpoints:
        logging.info("Loading grad scaler state dict")
        scaler.load_state_dict(checkpoints["grad_scaler"])

    for epoch in range(params.start_epoch, params.num_epochs + 1):
        scheduler.step_epoch(epoch - 1)
        fix_random_seed(params.seed + epoch - 1)
        train_dl.sampler.set_epoch(epoch - 1)

        if tb_writer is not None:
            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)

        params.cur_epoch = epoch

        train_one_epoch(
            params=params,
            model=model,
            model_avg=model_avg,
            optimizer=optimizer,
            scheduler=scheduler,
            sp=sp,
            train_dl=train_dl,
            train_dl_warmup=train_dl_ov40,
            valid_dl=valid_dl,
            scaler=scaler,
            tb_writer=tb_writer,
            world_size=world_size,
            rank=rank,
        )

        if params.print_diagnostics:
            diagnostic.print_diagnostics()
            break

        save_checkpoint(
            params=params,
            model=model,
            model_avg=model_avg,
            optimizer=optimizer,
            scheduler=scheduler,
            sampler=train_dl.sampler,
            scaler=scaler,
            rank=rank,
        )

    logging.info("Done!")

    if world_size > 1:
        torch.distributed.barrier()
        cleanup_dist()


def display_and_save_batch(
    batch: dict,
    params: AttributeDict,
    sp: spm.SentencePieceProcessor,
) -> None:
    """Display the batch statistics and save the batch into disk.

    Args:
      batch:
        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
        for the content in it.
      params:
        Parameters for training. See :func:`get_params`.
      sp:
        The BPE model.
    """
    from lhotse.utils import uuid4

    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
    logging.info(f"Saving batch to {filename}")
    torch.save(batch, filename)

    features = batch["inputs"]

    logging.info(f"features shape: {features.shape}")

    y = [sp.encode(text_ch) for text_ch in batch["text"]]
    num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y]
    logging.info(f"num tokens: {num_tokens}")


def main():
    parser = get_parser()
    LibriCssAsrDataModule.add_arguments(parser)
    args = parser.parse_args()
    args.exp_dir = Path(args.exp_dir)

    world_size = args.world_size
    assert world_size >= 1
    if world_size > 1:
        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
    else:
        run(rank=0, world_size=1, args=args)


torch.set_num_threads(1)
torch.set_num_interop_threads(1)
torch.multiprocessing.set_sharing_strategy("file_system")

if __name__ == "__main__":
    main()