From e6fe602db3eb1209543e55f1aafa2932dfda3310 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 10 一月 2025 10:14:30 +0800
Subject: [PATCH] step_or_epoch bugfix
---
funasr/train_utils/trainer_ds.py | 144 ++++++++++++++++++++++++++++++++----------------
1 files changed, 96 insertions(+), 48 deletions(-)
diff --git a/funasr/train_utils/trainer_ds.py b/funasr/train_utils/trainer_ds.py
index 1a553f8..0b104da 100644
--- a/funasr/train_utils/trainer_ds.py
+++ b/funasr/train_utils/trainer_ds.py
@@ -29,9 +29,10 @@
with torch.cuda.amp.autocast(enabled=True, dtype=dtype, cache_enabled=False):
yield
else:
- if dtype == torch.float16:
- with autocast(enabled=True):
- yield
+ if dtype == torch.float16 or dtype == torch.bfloat16:
+ yield
+ # with autocast(enabled=True, dtype=dtype):
+ # yield
else:
yield
@@ -60,6 +61,7 @@
use_ddp: bool = False,
use_fsdp: bool = False,
use_fp16: bool = False,
+ use_bf16: bool = False,
use_deepspeed: bool = False,
output_dir: str = "./",
**kwargs,
@@ -78,7 +80,7 @@
output_dir (str): The directory where model checkpoints will be saved. Default is './'.
resume (str, optional): The file path to a checkpoint to resume training from.
"""
- self.rank = kwargs.get("rank", 0)
+ self.rank = rank
self.local_rank = local_rank
self.world_size = world_size
self.use_ddp = use_ddp
@@ -98,8 +100,11 @@
self.batch_total = 0
self.dtype = torch.float32
self.use_fp16 = use_fp16
+ self.use_bf16 = use_bf16
if self.use_fp16:
self.dtype = torch.float16
+ if self.use_bf16:
+ self.dtype = torch.bfloat16
self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
self.validate_interval = kwargs.get("validate_interval", 5000)
self.keep_nbest_models = kwargs.get("keep_nbest_models", 500)
@@ -117,8 +122,8 @@
self.saved_ckpts = {}
self.step_or_epoch = -1
self.best_step_or_epoch = ""
- self.val_acc_step_or_eoch = {}
- self.val_loss_step_or_eoch = {}
+ self.val_acc_step_or_epoch = {}
+ self.val_loss_step_or_epoch = {}
self.reset_gpu_cache = kwargs.get("reset_gpu_cache", False)
self.start_data_split_i = 0
@@ -147,6 +152,16 @@
self.use_deepspeed = use_deepspeed
self.deepspeed_config = kwargs.get("deepspeed_config", "")
+ excludes = kwargs.get("excludes", None)
+ if excludes is not None:
+ if isinstance(excludes, str):
+ excludes = excludes.split(",")
+ self.excludes = excludes
+ effective_save_name_excludes = kwargs.get("effective_save_name_excludes", None)
+ if effective_save_name_excludes is not None:
+ if isinstance(effective_save_name_excludes, str):
+ effective_save_name_excludes = effective_save_name_excludes.split(",")
+ self.effective_save_name_excludes = effective_save_name_excludes
def save_checkpoint(
self,
@@ -167,6 +182,8 @@
Args:
epoch (int): The epoch number at which the checkpoint is being saved.
"""
+ if self.use_ddp or self.use_fsdp:
+ dist.barrier()
step_in_epoch = None if step is None else step_in_epoch
if self.use_deepspeed:
@@ -178,8 +195,8 @@
# "optimizer": optim.state_dict(),
# "scheduler": scheduler.state_dict(),
"saved_ckpts": self.saved_ckpts,
- "val_acc_step_or_eoch": self.val_acc_step_or_eoch,
- "val_loss_step_or_eoch": self.val_loss_step_or_eoch,
+ "val_acc_step_or_epoch": self.val_acc_step_or_epoch,
+ "val_loss_step_or_epoch": self.val_loss_step_or_epoch,
"best_step_or_epoch": self.best_step_or_epoch,
"avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
"step": step,
@@ -217,8 +234,8 @@
if self.avg_keep_nbest_models_type == "acc":
if (
- self.val_acc_step_or_eoch[ckpt_name]
- >= self.val_acc_step_or_eoch[self.best_step_or_epoch]
+ self.val_acc_step_or_epoch[ckpt_name]
+ >= self.val_acc_step_or_epoch[self.best_step_or_epoch]
):
self.best_step_or_epoch = ckpt_name
best_ckpt = Path(os.path.join(self.output_dir, f"model.pt.best"))
@@ -228,16 +245,16 @@
save_dir=self.output_dir, tag=f"model.pt.best", client_state=state
)
logging.info(
- f"Update best acc: {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
+ f"Update best acc: {self.val_acc_step_or_epoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
)
else:
logging.info(
- f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
+ f"No improvement in acc: {self.val_acc_step_or_epoch[ckpt_name]:.4f} < {self.val_acc_step_or_epoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
)
elif self.avg_keep_nbest_models_type == "loss":
if (
- self.val_loss_step_or_eoch[ckpt_name]
- <= self.val_loss_step_or_eoch[self.best_step_or_epoch]
+ self.val_loss_step_or_epoch[ckpt_name]
+ <= self.val_loss_step_or_epoch[self.best_step_or_epoch]
):
self.best_step_or_epoch = ckpt_name
best_ckpt = Path(os.path.join(self.output_dir, f"model.pt.best"))
@@ -247,16 +264,16 @@
save_dir=self.output_dir, tag=f"model.pt.best", client_state=state
)
logging.info(
- f"Update best loss: {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
+ f"Update best loss: {self.val_loss_step_or_epoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
)
else:
logging.info(
- f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
+ f"No improvement in loss: {self.val_loss_step_or_epoch[ckpt_name]:.4f} > {self.val_loss_step_or_epoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
)
else:
print("Undo")
self.saved_ckpts[ckpt_name] = getattr(
- self, f"val_{self.avg_keep_nbest_models_type}_step_or_eoch"
+ self, f"val_{self.avg_keep_nbest_models_type}_step_or_epoch"
)[ckpt_name]
if self.keep_nbest_models > 0:
if len(self.saved_ckpts) > self.keep_nbest_models:
@@ -275,16 +292,17 @@
elif self.use_fsdp:
pass
elif self.rank == 0:
- logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
+ logging.info(
+ f"Save checkpoint: {epoch}, rank: {self.rank}, local_rank: {self.local_rank}\n"
+ )
# self.step_or_epoch += 1
state = {
"epoch": epoch,
- "state_dict": model.state_dict(),
"optimizer": optim.state_dict(),
"scheduler": scheduler.state_dict(),
"saved_ckpts": self.saved_ckpts,
- "val_acc_step_or_eoch": self.val_acc_step_or_eoch,
- "val_loss_step_or_eoch": self.val_loss_step_or_eoch,
+ "val_acc_step_or_epoch": self.val_acc_step_or_epoch,
+ "val_loss_step_or_epoch": self.val_loss_step_or_epoch,
"best_step_or_epoch": self.best_step_or_epoch,
"avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
"step": step,
@@ -297,7 +315,24 @@
}
step = step_in_epoch
if hasattr(model, "module"):
- state["state_dict"] = model.module.state_dict()
+ state_dict = model.module.state_dict()
+ else:
+ state_dict = model.state_dict()
+
+ if self.effective_save_name_excludes is not None:
+ logging.info(f"effective_save_name_excludes: {self.effective_save_name_excludes}")
+ dst_state_dict = {}
+ for k in state_dict.keys():
+ for k_ex in self.effective_save_name_excludes:
+ k_tmp = k.replace("module.", "")
+ if k.startswith(k_ex):
+ logging.info(f"key: {k} matching: {k_ex}, not save it")
+ break
+ else:
+ dst_state_dict[k] = state_dict[k]
+ state["state_dict"] = dst_state_dict
+ else:
+ state["state_dict"] = state_dict
if scaler:
state["scaler_state"] = scaler.state_dict()
@@ -318,38 +353,38 @@
if self.avg_keep_nbest_models_type == "acc":
if (
- self.val_acc_step_or_eoch[ckpt_name]
- >= self.val_acc_step_or_eoch[self.best_step_or_epoch]
+ self.val_acc_step_or_epoch[ckpt_name]
+ >= self.val_acc_step_or_epoch[self.best_step_or_epoch]
):
self.best_step_or_epoch = ckpt_name
best_ckpt = Path(os.path.join(self.output_dir, f"model.pt.best"))
torch.save(state, best_ckpt)
logging.info(
- f"Update best acc: {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
+ f"Update best acc: {self.val_acc_step_or_epoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
)
else:
logging.info(
- f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
+ f"No improvement in acc: {self.val_acc_step_or_epoch[ckpt_name]:.4f} < {self.val_acc_step_or_epoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
)
elif self.avg_keep_nbest_models_type == "loss":
if (
- self.val_loss_step_or_eoch[ckpt_name]
- <= self.val_loss_step_or_eoch[self.best_step_or_epoch]
+ self.val_loss_step_or_epoch[ckpt_name]
+ <= self.val_loss_step_or_epoch[self.best_step_or_epoch]
):
self.best_step_or_epoch = ckpt_name
best_ckpt = Path(os.path.join(self.output_dir, f"model.pt.best"))
torch.save(state, best_ckpt)
logging.info(
- f"Update best loss: {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
+ f"Update best loss: {self.val_loss_step_or_epoch[self.best_step_or_epoch]:.4f}, {best_ckpt}"
)
else:
logging.info(
- f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
+ f"No improvement in loss: {self.val_loss_step_or_epoch[ckpt_name]:.4f} > {self.val_loss_step_or_epoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
)
else:
print("Undo")
self.saved_ckpts[ckpt_name] = getattr(
- self, f"val_{self.avg_keep_nbest_models_type}_step_or_eoch"
+ self, f"val_{self.avg_keep_nbest_models_type}_step_or_epoch"
)[ckpt_name]
if self.keep_nbest_models > 0:
if len(self.saved_ckpts) > self.keep_nbest_models:
@@ -390,14 +425,14 @@
_, checkpoint = model.load_checkpoint(self.output_dir, "model.pt")
self.start_epoch = checkpoint["epoch"]
self.saved_ckpts = checkpoint["saved_ckpts"]
- self.val_acc_step_or_eoch = (
- checkpoint["val_acc_step_or_eoch"]
- if "val_acc_step_or_eoch" in checkpoint
+ self.val_acc_step_or_epoch = (
+ checkpoint["val_acc_step_or_epoch"]
+ if "val_acc_step_or_epoch" in checkpoint
else {}
)
- self.val_loss_step_or_eoch = (
- checkpoint["val_loss_step_or_eoch"]
- if "val_loss_step_or_eoch" in checkpoint
+ self.val_loss_step_or_epoch = (
+ checkpoint["val_loss_step_or_epoch"]
+ if "val_loss_step_or_epoch" in checkpoint
else {}
)
self.best_step_or_epoch = (
@@ -438,6 +473,16 @@
src_state = checkpoint["state_dict"]
dst_state = model.state_dict()
for k in dst_state.keys():
+ excludes_flag = False
+ if self.excludes is not None:
+ for k_ex in self.excludes:
+ k_tmp = k.replace("module.", "")
+ if k_tmp.startswith(k_ex):
+ logging.info(f"key: {k} matching: {k_ex}, excluded")
+ excludes_flag = True
+ break
+ if excludes_flag:
+ continue
if not k.startswith("module.") and "module." + k in src_state.keys():
k_ddp = "module." + k
elif k.startswith("module.") and "module." + k not in src_state.keys():
@@ -456,14 +501,14 @@
scaler.load_state_dict(checkpoint["scaler_state"])
self.saved_ckpts = checkpoint["saved_ckpts"]
- self.val_acc_step_or_eoch = (
- checkpoint["val_acc_step_or_eoch"]
- if "val_acc_step_or_eoch" in checkpoint
+ self.val_acc_step_or_epoch = (
+ checkpoint["val_acc_step_or_epoch"]
+ if "val_acc_step_or_epoch" in checkpoint
else {}
)
- self.val_loss_step_or_eoch = (
- checkpoint["val_loss_step_or_eoch"]
- if "val_loss_step_or_eoch" in checkpoint
+ self.val_loss_step_or_epoch = (
+ checkpoint["val_loss_step_or_epoch"]
+ if "val_loss_step_or_epoch" in checkpoint
else {}
)
self.best_step_or_epoch = (
@@ -621,7 +666,6 @@
self.train_acc_avg = train_acc_avg.detach().cpu().item() / self.world_size
def forward_step(self, model, batch, loss_dict={}):
- dtype = torch.bfloat16
with maybe_autocast(dtype=self.dtype, use_deepspeed=self.use_deepspeed):
retval = model(**batch)
@@ -639,7 +683,7 @@
scaled_loss = model.backward(loss)
else:
loss = loss / self.accum_grad
- if self.use_fp16:
+ if self.use_fp16 or self.use_bf16:
scaler.scale(loss).backward()
else:
loss.backward()
@@ -667,7 +711,7 @@
# Execute an optimization step (update model parameters)
if self.use_ddp or self.use_fsdp:
dist.barrier()
- if self.use_fp16:
+ if self.use_fp16 or self.use_bf16:
scaler.step(optim)
scaler.update()
else:
@@ -759,8 +803,12 @@
ckpt_name = f"model.pt.ep{epoch}"
else:
ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step_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
+ self.val_acc_step_or_epoch[ckpt_name] = self.val_acc_avg
+ self.val_loss_step_or_epoch[ckpt_name] = self.val_loss_avg
+
+ if self.use_ddp or self.use_fsdp or self.use_deepspeed:
+ dist.barrier()
+
model.train()
def log(
--
Gitblit v1.9.1