train.py

#!/usr/bin/env python3

import os
import sys
import torch
import logging
from pathlib import Path
import speechbrain as sb
from hyperpyyaml import load_hyperpyyaml
from speechbrain.utils.distributed import run_on_main, if_main_process

logger = logging.getLogger(__name__)


# Define training procedure
class ASR(sb.core.Brain):
    def compute_forward(self, batch, stage):
        """Forward computations from the waveform batches to the output probabilities."""
        batch = batch.to(self.device)
        wavs, wav_lens = batch.sig
        tokens_bos, _ = batch.tokens_bos

        # Add augmentation if specified
        if stage == sb.Stage.TRAIN:
            if hasattr(self.modules, "env_corrupt"):
                wavs_noise = self.modules.env_corrupt(wavs, wav_lens)
                wavs = torch.cat([wavs, wavs_noise], dim=0)
                wav_lens = torch.cat([wav_lens, wav_lens])
                tokens_bos = torch.cat([tokens_bos, tokens_bos], dim=0)

        # compute features
        feats = self.hparams.compute_features(wavs)
        current_epoch = self.hparams.epoch_counter.current
        feats = self.modules.normalize(feats, wav_lens, epoch=current_epoch)

        if stage == sb.Stage.TRAIN:
            if hasattr(self.hparams, "augmentation"):
                feats = self.hparams.augmentation(feats)

        # forward modules
        src = self.modules.CNN(feats)

        enc_out, pred = self.modules.Transformer(
            src, tokens_bos, wav_lens, pad_idx=self.hparams.pad_index,
        )

        # output layer for ctc log-probabilities
        logits = self.modules.ctc_lin(enc_out)
        p_ctc = self.hparams.log_softmax(logits)

        # output layer for seq2seq log-probabilities
        pred = self.modules.seq_lin(pred)
        p_seq = self.hparams.log_softmax(pred)

        # Compute outputs
        hyps = None
        if stage == sb.Stage.TRAIN:
            hyps = None
        elif stage == sb.Stage.VALID:
            hyps = None
            current_epoch = self.hparams.epoch_counter.current
            if current_epoch % self.hparams.valid_search_interval == 0:
                # for the sake of efficiency, we only perform beamsearch with limited capacity
                # and no LM to give user some idea of how the AM is doing
                hyps, _ = self.hparams.valid_search(enc_out.detach(), wav_lens)
        elif stage == sb.Stage.TEST:
            hyps, _ = self.hparams.test_search(enc_out.detach(), wav_lens)

        return p_ctc, p_seq, wav_lens, hyps

    def compute_objectives(self, predictions, batch, stage):
        """Computes the loss (CTC+NLL) given predictions and targets."""

        (p_ctc, p_seq, wav_lens, hyps,) = predictions

        ids = batch.id
        tokens_eos, tokens_eos_lens = batch.tokens_eos
        tokens, tokens_lens = batch.tokens

        if hasattr(self.modules, "env_corrupt") and stage == sb.Stage.TRAIN:
            tokens_eos = torch.cat([tokens_eos, tokens_eos], dim=0)
            tokens_eos_lens = torch.cat(
                [tokens_eos_lens, tokens_eos_lens], dim=0
            )
            tokens = torch.cat([tokens, tokens], dim=0)
            tokens_lens = torch.cat([tokens_lens, tokens_lens], dim=0)

        loss_seq = self.hparams.seq_cost(
            p_seq, tokens_eos, length=tokens_eos_lens
        ).sum()

        loss_ctc = self.hparams.ctc_cost(
            p_ctc, tokens, wav_lens, tokens_lens
        ).sum()

        loss = (
            self.hparams.ctc_weight * loss_ctc
            + (1 - self.hparams.ctc_weight) * loss_seq
        )

        if stage != sb.Stage.TRAIN:
            current_epoch = self.hparams.epoch_counter.current
            valid_search_interval = self.hparams.valid_search_interval
            if current_epoch % valid_search_interval == 0 or (
                stage == sb.Stage.TEST
            ):
                # Decode token terms to words
                predicted_words = [
                    tokenizer.decode_ids(utt_seq).split(" ") for utt_seq in hyps
                ]
                target_words = [wrd.split(" ") for wrd in batch.wrd]
                self.wer_metric.append(ids, predicted_words, target_words)

            # compute the accuracy of the one-step-forward prediction
            self.acc_metric.append(p_seq, tokens_eos, tokens_eos_lens)
        return loss

    def on_evaluate_start(self, max_key=None, min_key=None):
        """perform checkpoint averge if needed"""
        super().on_evaluate_start()

        ckpts = self.checkpointer.find_checkpoints(
            max_key=max_key, min_key=min_key
        )
        ckpt = sb.utils.checkpoints.average_checkpoints(
            ckpts, recoverable_name="model", device=self.device
        )

        self.hparams.model.load_state_dict(ckpt, strict=True)
        self.hparams.model.eval()
        print("Loaded the average")

    def evaluate_batch(self, batch, stage):
        """Computations needed for validation/test batches"""
        with torch.no_grad():
            predictions = self.compute_forward(batch, stage=stage)
            loss = self.compute_objectives(predictions, batch, stage=stage)
        return loss.detach()

    def on_stage_start(self, stage, epoch):
        """Gets called at the beginning of each epoch"""
        if stage != sb.Stage.TRAIN:
            self.acc_metric = self.hparams.acc_computer()
            self.wer_metric = self.hparams.error_rate_computer()

    def on_stage_end(self, stage, stage_loss, epoch):
        """Gets called at the end of a epoch."""
        # Compute/store important stats
        stage_stats = {"loss": stage_loss}
        if stage == sb.Stage.TRAIN:
            self.train_stats = stage_stats
        else:
            stage_stats["ACC"] = self.acc_metric.summarize()
            current_epoch = self.hparams.epoch_counter.current
            valid_search_interval = self.hparams.valid_search_interval
            if (
                current_epoch % valid_search_interval == 0
                or stage == sb.Stage.TEST
            ):
                stage_stats["WER"] = self.wer_metric.summarize("error_rate")

        # log stats and save checkpoint at end-of-epoch
        if stage == sb.Stage.VALID and sb.utils.distributed.if_main_process():

            lr = self.hparams.noam_annealing.current_lr
            steps = self.optimizer_step
            optimizer = self.optimizer.__class__.__name__

            epoch_stats = {
                "epoch": epoch,
                "lr": lr,
                "steps": steps,
                "optimizer": optimizer,
            }
            self.hparams.train_logger.log_stats(
                stats_meta=epoch_stats,
                train_stats=self.train_stats,
                valid_stats=stage_stats,
            )
            self.checkpointer.save_and_keep_only(
                meta={"ACC": stage_stats["ACC"], "epoch": epoch},
                max_keys=["ACC"],
                num_to_keep=10,
            )

        elif stage == sb.Stage.TEST:
            self.hparams.train_logger.log_stats(
                stats_meta={"Epoch loaded": self.hparams.epoch_counter.current},
                test_stats=stage_stats,
            )
            if if_main_process():
                with open(self.hparams.test_wer_file, "w") as w:
                    self.wer_metric.write_stats(w)

            # save the averaged checkpoint at the end of the evaluation stage
            # delete the rest of the intermediate checkpoints
            # ACC is set to 1.1 so checkpointer only keeps the averaged checkpoint
            self.checkpointer.save_and_keep_only(
                meta={"ACC": 1.1, "epoch": epoch},
                max_keys=["ACC"],
                num_to_keep=1,
            )

    def fit_batch(self, batch):

        should_step = self.step % self.grad_accumulation_factor == 0
        # Managing automatic mixed precision
        if self.auto_mix_prec:
            with torch.autocast(torch.device(self.device).type):
                outputs = self.compute_forward(batch, sb.Stage.TRAIN)

            # Losses are excluded from mixed precision to avoid instabilities
            loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
            with self.no_sync(not should_step):
                self.scaler.scale(
                    loss / self.grad_accumulation_factor
                ).backward()
            if should_step:
                self.scaler.unscale_(self.optimizer)
                if self.check_gradients(loss):
                    self.scaler.step(self.optimizer)
                self.scaler.update()
                self.zero_grad()
                self.optimizer_step += 1
                self.hparams.noam_annealing(self.optimizer)
        else:
            if self.bfloat16_mix_prec:
                with torch.autocast(
                    device_type=torch.device(self.device).type,
                    dtype=torch.bfloat16,
                ):
                    outputs = self.compute_forward(batch, sb.Stage.TRAIN)
                    loss = self.compute_objectives(
                        outputs, batch, sb.Stage.TRAIN
                    )
            else:
                outputs = self.compute_forward(batch, sb.Stage.TRAIN)
                loss = self.compute_objectives(outputs, batch, sb.Stage.TRAIN)
            with self.no_sync(not should_step):
                (loss / self.grad_accumulation_factor).backward()
            if should_step:
                if self.check_gradients(loss):
                    self.optimizer.step()
                self.zero_grad()
                self.optimizer_step += 1
                self.hparams.noam_annealing(self.optimizer)

        self.on_fit_batch_end(batch, outputs, loss, should_step)
        return loss.detach().cpu()


def dataio_prepare(hparams):
    """This function prepares the datasets to be used in the brain class.
    It also defines the data processing pipeline through user-defined functions."""
    data_folder = hparams["data_folder"]

    train_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
        csv_path=hparams["train_csv"], replacements={"data_root": data_folder},
    )

    if hparams["sorting"] == "ascending":
        # we sort training data to speed up training and get better results.
        train_data = train_data.filtered_sorted(sort_key="duration")
        # when sorting do not shuffle in dataloader ! otherwise is pointless
        hparams["train_dataloader_opts"]["shuffle"] = False

    elif hparams["sorting"] == "descending":
        train_data = train_data.filtered_sorted(
            sort_key="duration", reverse=True
        )
        # when sorting do not shuffle in dataloader ! otherwise is pointless
        hparams["train_dataloader_opts"]["shuffle"] = False

    elif hparams["sorting"] == "random":
        pass

    else:
        raise NotImplementedError(
            "sorting must be random, ascending or descending"
        )
    valid_data = sb.dataio.dataset.DynamicItemDataset.from_csv(
        csv_path=hparams["valid_csv"], replacements={"data_root": data_folder},
    )
    valid_data = valid_data.filtered_sorted(sort_key="duration")

    # test is separate
    test_datasets = {}
    for csv_file in hparams["test_csv"]:
        name = Path(csv_file).stem
        test_datasets[name] = sb.dataio.dataset.DynamicItemDataset.from_csv(
            csv_path=csv_file, replacements={"data_root": data_folder}
        )
        test_datasets[name] = test_datasets[name].filtered_sorted(
            sort_key="duration"
        )

    datasets = [train_data, valid_data] + [i for k, i in test_datasets.items()]
    valtest_datasets = [valid_data] + [i for k, i in test_datasets.items()]

    # We get the tokenizer as we need it to encode the labels when creating
    # mini-batches.
    tokenizer = hparams["tokenizer"]

    # 2. Define audio pipeline:
    @sb.utils.data_pipeline.takes("wav")
    @sb.utils.data_pipeline.provides("sig")
    def audio_pipeline(wav):
        sig = sb.dataio.dataio.read_audio(wav)
        return sig

    sb.dataio.dataset.add_dynamic_item(valtest_datasets, audio_pipeline)

    @sb.utils.data_pipeline.takes("wav")
    @sb.utils.data_pipeline.provides("sig")
    def audio_pipeline_train(wav):
        # Speed Perturb is done here so it is multi-threaded with the
        # workers of the dataloader (faster).
        if "speed_perturb" in hparams:
            sig = sb.dataio.dataio.read_audio(wav)

            sig = hparams["speed_perturb"](sig.unsqueeze(0)).squeeze(0)
        else:
            sig = sb.dataio.dataio.read_audio(wav)
        return sig

    sb.dataio.dataset.add_dynamic_item([train_data], audio_pipeline_train)

    # 3. Define text pipeline:
    @sb.utils.data_pipeline.takes("wrd")
    @sb.utils.data_pipeline.provides(
        "wrd", "tokens_list", "tokens_bos", "tokens_eos", "tokens"
    )
    def text_pipeline(wrd):
        yield wrd
        tokens_list = tokenizer.encode_as_ids(wrd)
        yield tokens_list
        tokens_bos = torch.LongTensor([hparams["bos_index"]] + (tokens_list))
        yield tokens_bos
        tokens_eos = torch.LongTensor(tokens_list + [hparams["eos_index"]])
        yield tokens_eos
        tokens = torch.LongTensor(tokens_list)
        yield tokens

    sb.dataio.dataset.add_dynamic_item(datasets, text_pipeline)

    # 4. Set output:
    sb.dataio.dataset.set_output_keys(
        datasets, ["id", "sig", "wrd", "tokens_bos", "tokens_eos", "tokens"],
    )

    # 5. If Dynamic Batching is used, we instantiate the needed samplers.
    train_batch_sampler = None
    valid_batch_sampler = None
    if hparams["dynamic_batching"]:
        from speechbrain.dataio.sampler import DynamicBatchSampler  # noqa

        dynamic_hparams = hparams["dynamic_batch_sampler"]
        num_buckets = dynamic_hparams["num_buckets"]

        train_batch_sampler = DynamicBatchSampler(
            train_data,
            dynamic_hparams["max_batch_len"],
            num_buckets=num_buckets,
            length_func=lambda x: x["duration"],
            shuffle=dynamic_hparams["shuffle_ex"],
            batch_ordering=dynamic_hparams["batch_ordering"],
            max_batch_ex=dynamic_hparams["max_batch_ex"],
        )

        valid_batch_sampler = DynamicBatchSampler(
            valid_data,
            dynamic_hparams["max_batch_len_val"],
            num_buckets=num_buckets,
            length_func=lambda x: x["duration"],
            shuffle=dynamic_hparams["shuffle_ex"],
            batch_ordering=dynamic_hparams["batch_ordering"],
        )

    return (
        train_data,
        valid_data,
        test_datasets,
        tokenizer,
        train_batch_sampler,
        valid_batch_sampler,
    )


if __name__ == "__main__":
    # CLI:
    hparams_file, run_opts, overrides = sb.parse_arguments(sys.argv[1:])
    with open(hparams_file) as fin:
        hparams = load_hyperpyyaml(fin, overrides)

    # If --distributed_launch then
    # create ddp_group with the right communication protocol
    sb.utils.distributed.ddp_init_group(run_opts)

    # 1.  # Dataset prep (parsing Librispeech)
    from librispeech_prepare import prepare_librispeech  # noqa

    # Create experiment directory
    sb.create_experiment_directory(
        experiment_directory=hparams["output_folder"],
        hyperparams_to_save=hparams_file,
        overrides=overrides,
    )

    # multi-gpu (ddp) save data preparation
    run_on_main(
        prepare_librispeech,
        kwargs={
            "data_folder": hparams["data_folder"],
            "tr_splits": hparams["train_splits"],
            "dev_splits": hparams["dev_splits"],
            "te_splits": hparams["test_splits"],
            "save_folder": hparams["output_folder"],
            "merge_lst": hparams["train_splits"],
            "merge_name": "train.csv",
            "skip_prep": hparams["skip_prep"],
        },
    )

    # here we create the datasets objects as well as tokenization and encoding
    (
        train_data,
        valid_data,
        test_datasets,
        tokenizer,
        train_bsampler,
        valid_bsampler,
    ) = dataio_prepare(hparams)

    # We download the pretrained LM from HuggingFace (or elsewhere depending on
    # the path given in the YAML file). The tokenizer is loaded at the same time.
    run_on_main(hparams["pretrainer"].collect_files)
    hparams["pretrainer"].load_collected(device=run_opts["device"])

    # Trainer initialization
    asr_brain = ASR(
        modules=hparams["modules"],
        opt_class=hparams["Adam"],
        hparams=hparams,
        run_opts=run_opts,
        checkpointer=hparams["checkpointer"],
    )

    # adding objects to trainer:
    asr_brain.tokenizer = hparams["tokenizer"]
    train_dataloader_opts = hparams["train_dataloader_opts"]
    valid_dataloader_opts = hparams["valid_dataloader_opts"]

    if train_bsampler is not None:
        collate_fn = None
        if "collate_fn" in train_dataloader_opts:
            collate_fn = train_dataloader_opts["collate_fn"]

        train_dataloader_opts = {
            "batch_sampler": train_bsampler,
            "num_workers": hparams["num_workers"],
        }

        if collate_fn is not None:
            train_dataloader_opts["collate_fn"] = collate_fn

    if valid_bsampler is not None:
        collate_fn = None
        if "collate_fn" in valid_dataloader_opts:
            collate_fn = valid_dataloader_opts["collate_fn"]

        valid_dataloader_opts = {"batch_sampler": valid_bsampler}

        if collate_fn is not None:
            valid_dataloader_opts["collate_fn"] = collate_fn

    # Training
    asr_brain.fit(
        asr_brain.hparams.epoch_counter,
        train_data,
        valid_data,
        train_loader_kwargs=train_dataloader_opts,
        valid_loader_kwargs=valid_dataloader_opts,
    )

    # Testing
    if not os.path.exists(hparams["output_wer_folder"]):
        os.makedirs(hparams["output_wer_folder"])

    for k in test_datasets.keys():  # keys are test_clean, test_other etc
        asr_brain.hparams.test_wer_file = os.path.join(
            hparams["output_wer_folder"], f"wer_{k}.txt"
        )
        asr_brain.evaluate(
            test_datasets[k],
            max_key="ACC",
            test_loader_kwargs=hparams["test_dataloader_opts"],
        )