From 15868f623089cf70983a8b4f435ff86e7f160b8a Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 06 十二月 2023 23:50:54 +0800
Subject: [PATCH] funasr2
---
funasr/cli/train_cli.py | 59 ++++++++++++++++++++++++++---------------------------------
1 files changed, 26 insertions(+), 33 deletions(-)
diff --git a/funasr/cli/train_cli.py b/funasr/cli/train_cli.py
index 28e0e28..ed62773 100644
--- a/funasr/cli/train_cli.py
+++ b/funasr/cli/train_cli.py
@@ -37,7 +37,7 @@
@hydra.main()
def main(kwargs: DictConfig):
# preprocess_config(kwargs)
- import pdb; pdb.set_trace()
+ # import pdb; pdb.set_trace()
# set random seed
set_all_random_seed(kwargs.get("seed", 0))
torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
@@ -46,11 +46,11 @@
local_rank = int(os.environ.get('LOCAL_RANK', 0))
# Check if we are using DDP or FSDP
- use_ddp = 'WORLD_SIZE' in os.environ
+ use_ddp = 'WORLD_SIZE' in os.environ and os.environ["WORLD_SIZE"] > 1
use_fsdp = kwargs.get("use_fsdp", None)
if use_ddp or use_fsdp:
dist.init_process_group(backend=kwargs.get("backend", "nccl"), init_method='env://')
- device= torch.cuda.set_device(local_rank)
+ torch.cuda.set_device(local_rank)
# build_tokenizer
@@ -72,9 +72,24 @@
# model_class = load_class_from_path(kwargs.get("model").split(":"))
model_class = dynamic_import(kwargs.get("model"))
model = model_class(**kwargs, **kwargs["model_conf"], vocab_size=len(tokenizer.token_list))
- # model = model.to(device=kwargs.get("device", "cpu"))
-
-
+ frontend = model.frontend
+ # init_param
+ init_param = kwargs.get("init_param", None)
+ if init_param is not None:
+ init_param = eval(init_param)
+ if isinstance(init_param, Sequence):
+ init_param = (init_param,)
+ logging.info("init_param is not None: ", init_param)
+ for p in init_param:
+ logging.info(f"Loading pretrained params from {p}")
+ load_pretrained_model(
+ model=model,
+ init_param=p,
+ ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
+ oss_bucket=kwargs.get("oss_bucket", None),
+ )
+ else:
+ initialize(model, kwargs.get("init", "kaiming_normal"))
# import pdb;
# pdb.set_trace()
@@ -94,9 +109,12 @@
if use_ddp:
model = model.cuda(local_rank)
- model = DDP(model, device_ids=[local_rank])
+ model = DDP(model, device_ids=[local_rank],
+ find_unused_parameters=kwargs.get("train_conf", {}).get("find_unused_parameters", False))
elif use_fsdp:
model = FSDP(model).cuda(local_rank)
+ else:
+ model = model.to(device=kwargs.get("device", "cuda"))
# optim
@@ -111,27 +129,9 @@
scheduler_class = scheduler_choices.get(scheduler)
scheduler = scheduler_class(optim, **kwargs.get("scheduler_conf"))
- # init_param
- init_param = kwargs.get("init_param", None)
- if init_param is not None:
- init_param = eval(init_param)
- if isinstance(init_param, Sequence):
- init_param = (init_param,)
- logging.info("init_param is not None: ", freeze_param)
- for p in init_param:
- logging.info(f"Loading pretrained params from {p}")
- load_pretrained_model(
- model=model,
- init_param=p,
- ignore_init_mismatch=kwargs.get("ignore_init_mismatch", True),
- oss_bucket=kwargs.get("oss_bucket", None),
- )
- else:
- initialize(model, kwargs.get("init", "kaiming_normal"))
-
# dataset
- dataset_tr = AudioDataset(kwargs.get("train_data_set_list"), frontend=model.frontend, tokenizer=tokenizer, **kwargs.get("dataset_conf"))
+ dataset_tr = AudioDataset(kwargs.get("train_data_set_list"), frontend=frontend, tokenizer=tokenizer, **kwargs.get("dataset_conf"))
# dataloader
batch_sampler = BatchSampler(dataset_tr, **kwargs.get("dataset_conf"), **kwargs.get("dataset_conf").get("batch_conf"))
@@ -158,13 +158,6 @@
torch.distributed.destroy_process_group()
-
-def train(epoch, model, op):
- pass
-
-def val():
- pass
-
if __name__ == "__main__":
main()
\ No newline at end of file
--
Gitblit v1.9.1