From 0a4a1d5257dace9561d95b38a9386539908dcd5e Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 23 四月 2024 12:48:52 +0800
Subject: [PATCH] Dev gzf exp (#1645)
---
funasr/bin/train.py | 31 ++++++++++++++++++-------------
1 files changed, 18 insertions(+), 13 deletions(-)
diff --git a/funasr/bin/train.py b/funasr/bin/train.py
index 0ff4ba1..ab49c82 100644
--- a/funasr/bin/train.py
+++ b/funasr/bin/train.py
@@ -55,6 +55,8 @@
torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
torch.backends.cudnn.benchmark = kwargs.get("cudnn_benchmark", torch.backends.cudnn.benchmark)
torch.backends.cudnn.deterministic = kwargs.get("cudnn_deterministic", True)
+ # open tf32
+ torch.backends.cuda.matmul.allow_tf32 = kwargs.get("enable_tf32", True)
local_rank = int(os.environ.get('LOCAL_RANK', 0))
if local_rank == 0:
@@ -88,7 +90,8 @@
# freeze_param
freeze_param = kwargs.get("freeze_param", None)
if freeze_param is not None:
- freeze_param = eval(freeze_param)
+ if "," in freeze_param:
+ freeze_param = eval(freeze_param)
if isinstance(freeze_param, Sequence):
freeze_param = (freeze_param,)
logging.info("freeze_param is not None: %s", freeze_param)
@@ -128,7 +131,8 @@
else:
model = model.to(device=kwargs.get("device", "cuda"))
- logging.info(f"{model}")
+ if local_rank == 0:
+ logging.info(f"{model}")
kwargs["device"] = next(model.parameters()).device
# optim
@@ -149,8 +153,8 @@
# dataset
logging.info("Build dataloader")
dataloader_class = tables.dataloader_classes.get(kwargs["dataset_conf"].get("dataloader", "DataloaderMapStyle"))
- # dataloader = dataloader_class(**kwargs)
- dataloader_tr, dataloader_val = dataloader_class(**kwargs)
+ dataloader = dataloader_class(**kwargs)
+ # dataloader_tr, dataloader_val = dataloader_class(**kwargs)
trainer = Trainer(local_rank=local_rank,
use_ddp=use_ddp,
use_fsdp=use_fsdp,
@@ -175,12 +179,12 @@
# if use_ddp or use_fsdp:
# context = Join([model])
# else:
+ # context = nullcontext()
context = nullcontext()
-
for epoch in range(trainer.start_epoch, trainer.max_epoch + 1):
time1 = time.perf_counter()
with context:
- # dataloader_tr, dataloader_val = dataloader.build_iter(epoch)
+ dataloader_tr, dataloader_val = dataloader.build_iter(epoch)
trainer.train_epoch(
model=model,
optim=optim,
@@ -191,13 +195,14 @@
epoch=epoch,
writer=writer
)
+ with context:
+ trainer.validate_epoch(
+ model=model,
+ dataloader_val=dataloader_val,
+ epoch=epoch,
+ writer=writer
+ )
scheduler.step()
- trainer.validate_epoch(
- model=model,
- dataloader_val=dataloader_val,
- epoch=epoch,
- writer=writer
- )
trainer.save_checkpoint(epoch, model=model, optim=optim, scheduler=scheduler, scaler=scaler)
@@ -212,7 +217,7 @@
if trainer.rank == 0:
- average_checkpoints(trainer.output_dir, trainer.avg_nbest_model, trainer.val_acc_list)
+ average_checkpoints(trainer.output_dir, trainer.avg_nbest_model)
trainer.close()
--
Gitblit v1.9.1