From 5052ca8bf03c178d9ae73cb68785cd4afb0144d2 Mon Sep 17 00:00:00 2001
From: 辰冢 <49506152+BruceLee569@users.noreply.github.com>
Date: 星期三, 12 二月 2025 16:13:08 +0800
Subject: [PATCH] Hotwords file needs to specify default utf-8 encoding. (#2379)
---
funasr/bin/train.py | 44 +++++++++++++++++++++++++++++++-------------
1 files changed, 31 insertions(+), 13 deletions(-)
diff --git a/funasr/bin/train.py b/funasr/bin/train.py
index 448e464..45e23a8 100644
--- a/funasr/bin/train.py
+++ b/funasr/bin/train.py
@@ -13,13 +13,14 @@
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
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.algorithms.join import Join
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
+from tensorboardX import SummaryWriter
from funasr.train_utils.average_nbest_models import average_checkpoints
from funasr.register import tables
@@ -27,7 +28,7 @@
from funasr.train_utils.trainer import Trainer
from funasr.schedulers import scheduler_classes
from funasr.train_utils.initialize import initialize
-from funasr.download.download_from_hub import download_model
+from funasr.download.download_model_from_hub import download_model
from funasr.models.lora.utils import mark_only_lora_as_trainable
from funasr.train_utils.set_all_random_seed import set_all_random_seed
from funasr.train_utils.load_pretrained_model import load_pretrained_model
@@ -99,7 +100,7 @@
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:
@@ -191,21 +192,20 @@
tensorboard_dir = os.path.join(kwargs.get("output_dir"), "tensorboard")
os.makedirs(tensorboard_dir, exist_ok=True)
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
dataloader_tr, dataloader_val = None, None
- for epoch in range(trainer.start_epoch, trainer.max_epoch + 1):
+ for epoch in range(trainer.start_epoch, trainer.max_epoch):
time1 = time.perf_counter()
for data_split_i in range(trainer.start_data_split_i, dataloader.data_split_num):
+ time_slice_i = time.perf_counter()
dataloader_tr, dataloader_val = dataloader.build_iter(
epoch, data_split_i=data_split_i, start_step=trainer.start_step
)
- trainer.start_step = 0
+
trainer.train_epoch(
model=model,
optim=optim,
@@ -217,16 +217,32 @@
writer=writer,
data_split_i=data_split_i,
data_split_num=dataloader.data_split_num,
+ start_step=trainer.start_step,
+ )
+ trainer.start_step = 0
+
+ # device = next(model.parameters()).device
+ # if device.type == 'cuda':
+ # with torch.cuda.device():
+ # torch.cuda.empty_cache()
+
+ time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
+ logging.info(
+ f"rank: {local_rank}, "
+ f"time_escaped_epoch: {time_escaped:.3f} hours, "
+ f"estimated to finish {dataloader.data_split_num} data_slices, remaining: {dataloader.data_split_num-data_split_i} slices, {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours, "
+ f"epoch: {trainer.max_epoch - epoch} epochs, {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
)
- torch.cuda.empty_cache()
-
+ trainer.start_data_split_i = 0
trainer.validate_epoch(
- model=model, dataloader_val=dataloader_val, epoch=epoch, writer=writer
+ model=model, dataloader_val=dataloader_val, epoch=epoch + 1, writer=writer
)
scheduler.step()
-
- trainer.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler)
+ trainer.step_in_epoch = 0
+ trainer.save_checkpoint(
+ epoch + 1, model=model, optim=optim, scheduler=scheduler, scaler=scaler
+ )
time2 = time.perf_counter()
time_escaped = (time2 - time1) / 3600.0
@@ -236,6 +252,8 @@
f"estimated to finish {trainer.max_epoch} "
f"epoch: {(trainer.max_epoch - epoch) * time_escaped:.3f} hours\n"
)
+ trainer.train_acc_avg = 0.0
+ trainer.train_loss_avg = 0.0
if trainer.rank == 0:
average_checkpoints(trainer.output_dir, trainer.avg_nbest_model)
--
Gitblit v1.9.1