Dev gzf exp (#1670)

zhifu gao

2024-04-28 93ef505e2d426b6aa1e58c0b4721999de789ff8e

Dev gzf exp (#1670)

* resume from step

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

* batch

 docs/images/wechat.png

 funasr/bin/train.py

@@ -13,7 +13,7 @@

from contextlib import nullcontext
import torch.distributed as dist
from collections.abc import Sequence

from omegaconf import DictConfig, OmegaConf
from torch.cuda.amp import autocast, GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
@@ -99,7 +99,7 @@
    if freeze_param is not None:
        if "," in freeze_param:
            freeze_param = eval(freeze_param)
        if not isinstance(freeze_param, Sequence):
        if not isinstance(freeze_param, (list, tuple)):
            freeze_param = (freeze_param,)
        logging.info("freeze_param is not None: %s", freeze_param)
        for t in freeze_param:
@@ -193,7 +193,7 @@
    try:
        from tensorboardX import SummaryWriter

        writer = SummaryWriter(tensorboard_dir) if trainer.rank == 0 else None
        writer = SummaryWriter(tensorboard_dir)  # if trainer.rank == 0 else None
    except:
        writer = None

@@ -206,6 +206,7 @@
                epoch, data_split_i=data_split_i, start_step=trainer.start_step
            )
            trainer.start_step = 0

            trainer.train_epoch(
                model=model,
                optim=optim,
@@ -225,7 +226,7 @@
            model=model, dataloader_val=dataloader_val, epoch=epoch, writer=writer
        )
        scheduler.step()

        trainer.step_cur_in_epoch = 0
        trainer.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler)

        time2 = time.perf_counter()

 funasr/datasets/sense_voice_datasets/datasets.py

@@ -51,6 +51,7 @@
        self.batch_size = kwargs.get("batch_size")
        self.batch_type = kwargs.get("batch_type")
        self.prompt_ids_len = 0
        self.retry = kwargs.get("retry", 5)

    def get_source_len(self, index):
        item = self.index_ds[index]
@@ -64,9 +65,21 @@
        return len(self.index_ds)

    def __getitem__(self, index):
        item = self.index_ds[index]
        # import pdb;
        # pdb.set_trace()

        output = None
        for idx in range(self.retry):
            if idx == 0:
                index_cur = index
            else:
                if index <= self.retry:
                    index_cur = index + idx
                else:
                    index_cur = torch.randint(0, index, ()).item()

            item = self.index_ds[index_cur]

        source = item["source"]
        data_src = load_audio_text_image_video(source, fs=self.fs)
        if self.preprocessor_speech:
@@ -76,7 +89,7 @@
        )  # speech: [b, T, d]

        if speech_lengths > self.batch_size:
            return None
                continue
        speech = speech.permute(0, 2, 1)
        target = item["target"]
        if self.preprocessor_text:
@@ -93,7 +106,7 @@
        target_ids = self.tokenizer.encode(target, allowed_special="all")
        target_ids_len = len(target_ids) + 1  # [lid, text]
        if target_ids_len > 200:
            return None
                continue

        eos = self.tokenizer.encode(self.eos, allowed_special="all")  # [eos]

@@ -109,7 +122,8 @@
        target_mask_lengths = len(target_mask)
        target_mask = torch.tensor(target_mask, dtype=torch.float32)
        target_mask_lengths = torch.tensor([target_mask_lengths], dtype=torch.int32)
        return {

            output = {
            "speech": speech[0, :, :],
            "speech_lengths": speech_lengths,
            "text": text,
@@ -117,6 +131,9 @@
            "target_mask": target_mask,
            "target_mask_lengths": target_mask_lengths,
        }
            break

        return output

    def collator(self, samples: list = None):
        outputs = {}
@@ -129,13 +146,30 @@
                outputs[key].append(sample[key])

        if len(outputs) < 1:
            logging.info(f"ERROR: data is empty!")
            logging.error(f"ERROR: data is empty!")
            outputs = {
                "speech": torch.rand((10, 128), dtype=torch.float32),
                "speech_lengths": torch.tensor([10], dtype=torch.int32),
                "text": torch.tensor([58836], dtype=torch.int32),
                "text_lengths": torch.tensor([1], dtype=torch.int32),
                "target_mask": torch.tensor([[0] * (self.prompt_ids_len) + [1] * (1) + [1]]),
                "speech": torch.rand((10, 128), dtype=torch.float32)[None, :, :],
                "speech_lengths": torch.tensor(
                    [
                        10,
                    ],
                    dtype=torch.int32,
                )[:, None],
                "text": torch.tensor(
                    [
                        58836,
                    ],
                    dtype=torch.int32,
                )[None, :],
                "text_lengths": torch.tensor(
                    [
                        1,
                    ],
                    dtype=torch.int32,
                )[:, None],
                "target_mask": torch.tensor([[0] * (self.prompt_ids_len) + [1] * (1) + [1]])[
                    None, :
                ],
            }
            return outputs

@@ -169,7 +203,6 @@
            )
            for key, data_list in outputs.items():
                outputs[key] = outputs[key][beg : beg + b : 2]

                
            speech_lengths_max = outputs["speech_lengths"].max().item()
            outputs["speech"] = outputs["speech"][:, :speech_lengths_max, :]

 funasr/train_utils/trainer.py

@@ -116,6 +116,7 @@
        self.reset_gpu_cache = kwargs.get("reset_gpu_cache", False)
        self.start_data_split_i = 0
        self.start_step = 0
        self.step_cur_in_epoch = 0
        self.use_wandb = kwargs.get("use_wandb", False)
        if self.use_wandb:
            wandb.login(key=kwargs.get("wandb_token"))
@@ -137,6 +138,8 @@
        optim=None,
        scheduler=None,
        scaler=None,
        step_cur_in_epoch=None,
        **kwargs,
    ):
        """
        Saves a checkpoint containing the model's state, the optimizer's state,
@@ -147,6 +150,7 @@
            epoch (int): The epoch number at which the checkpoint is being saved.
        """

        step_cur_in_epoch = None if step is None else step_cur_in_epoch
        if self.rank == 0:
            logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
            # self.step_or_epoch += 1
@@ -161,7 +165,12 @@
                "best_step_or_epoch": self.best_step_or_epoch,
                "avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
                "step": step,
                "step_cur_in_epoch": step_cur_in_epoch,
                "data_split_i": kwargs.get("data_split_i", 0),
                "data_split_num": kwargs.get("data_split_num", 1),
                "batch_total": self.batch_total,
            }
            step = step_cur_in_epoch
            if hasattr(model, "module"):
                state["state_dict"] = model.module.state_dict()

@@ -293,6 +302,12 @@
                self.batch_total = checkpoint["batch_total"] if "batch_total" in checkpoint else 0
                self.start_step = checkpoint["step"] if "step" in checkpoint else 0
                self.start_step = 0 if self.start_step is None else self.start_step
                self.step_cur_in_epoch = (
                    checkpoint["step_cur_in_epoch"] if "step_cur_in_epoch" in checkpoint else 0
                )
                self.step_cur_in_epoch = (
                    0 if self.step_cur_in_epoch is None else self.step_cur_in_epoch
                )

                model.to(self.device)
                print(f"Checkpoint loaded successfully from '{ckpt}'")
@@ -321,7 +336,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Train epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Train epoch: {epoch}, rank: {self.rank}\n")
        model.train()

        # Set the number of steps for gradient accumulation
@@ -341,6 +356,7 @@
                if iterator_stop > 0:
                    break
            self.batch_total += 1
            self.step_cur_in_epoch += 1
            time1 = time.perf_counter()
            speed_stats["data_load"] = f"{time1-time_beg:0.3f}"

@@ -443,6 +459,7 @@
                self.log(
                    epoch,
                    batch_idx,
                    step_cur_in_epoch=self.step_cur_in_epoch,
                    batch_num_epoch=batch_num_epoch,
                    lr=lr,
                    loss=loss.detach().cpu().item(),
@@ -461,6 +478,7 @@
                    epoch=epoch,
                    writer=writer,
                    step=batch_idx + 1,
                    step_cur_in_epoch=self.step_cur_in_epoch,
                )

            if (batch_idx + 1) % self.save_checkpoint_interval == 0:
@@ -471,6 +489,9 @@
                    scheduler=scheduler,
                    scaler=scaler,
                    step=batch_idx + 1,
                    step_cur_in_epoch=self.step_cur_in_epoch,
                    data_split_i=kwargs.get("data_split_i", 0),
                    data_split_num=kwargs.get("data_split_num", 1),
                )

            time_beg = time.perf_counter()
@@ -500,7 +521,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Validate epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Validate epoch: {epoch}, rank: {self.rank}\n")
        model.eval()

        with torch.no_grad():
@@ -578,10 +599,10 @@
                    iterator_stop.fill_(1)
                    dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)

        if kwargs.get("step", None) is None:
        if kwargs.get("step_cur_in_epoch", None) is None:
            ckpt_name = f"model.pt.ep{epoch}"
        else:
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step")}'
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step_cur_in_epoch")}'
        self.val_acc_step_or_eoch[ckpt_name] = self.val_acc_avg
        self.val_loss_step_or_eoch[ckpt_name] = self.val_loss_avg
        model.train()
@@ -594,6 +615,7 @@
        self,
        epoch=0,
        batch_idx=0,
        step_cur_in_epoch=0,
        batch_num_epoch=-1,
        lr=0.0,
        loss=0.0,
@@ -626,6 +648,7 @@
                f"{tag}, "
                f"rank: {self.rank}, "
                f"epoch: {epoch}/{self.max_epoch}, "
                f"step_cur_in_epoch: {step_cur_in_epoch}, "
                f"data_slice: {data_split_i}/{data_split_num}, "
                f"step: {batch_idx + 1}/{batch_num_epoch}, total step: {self.batch_total}, "
                f"(loss_avg_rank: {loss:.3f}), "