From 48b9082a2d4b96e5aeb995ac7d39b01f298eb165 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 12 五月 2023 11:41:42 +0800
Subject: [PATCH] fix
---
funasr/train/trainer.py | 27 ++++++++++++++++++++++-----
1 files changed, 22 insertions(+), 5 deletions(-)
diff --git a/funasr/train/trainer.py b/funasr/train/trainer.py
index 9574a0d..a40f031 100644
--- a/funasr/train/trainer.py
+++ b/funasr/train/trainer.py
@@ -95,6 +95,7 @@
use_pai: bool
oss_bucket: Union[oss2.Bucket, None]
batch_interval: int
+ bias_grad_times: float
class Trainer:
"""Trainer having a optimizer.
@@ -186,9 +187,6 @@
logging.warning("No keep_nbest_models is given. Change to [1]")
trainer_options.keep_nbest_models = [1]
keep_nbest_models = trainer_options.keep_nbest_models
-
- #assert batch_interval is set and >0
- assert trainer_options.batch_interval > 0
output_dir = Path(trainer_options.output_dir)
reporter = Reporter()
@@ -549,8 +547,11 @@
no_forward_run = options.no_forward_run
ngpu = options.ngpu
use_wandb = options.use_wandb
+ bias_grad_times = options.bias_grad_times
distributed = distributed_option.distributed
+ if bias_grad_times != 1.0:
+ logging.warning("Using bias_grad_times: {} for gradient scaling".format(bias_grad_times))
if log_interval is None:
try:
log_interval = max(len(iterator) // 20, 10)
@@ -585,10 +586,16 @@
if num_batch_updates % batch_interval == 0:
if options.use_pai and options.oss_bucket is not None:
buffer = BytesIO()
- torch.save(model.state_dict(), buffer)
+ if hasattr(model, "module"):
+ torch.save(model.module.state_dict(), buffer)
+ else:
+ torch.save(model.state_dict(), buffer)
options.oss_bucket.put_object(os.path.join(output_dir, f"{num_batch_updates}step.pb"), buffer.getvalue())
else:
- torch.save(model.state_dict(), os.path.join(output_dir, f"{num_batch_updates}step.pb"))
+ if hasattr(model, "module"):
+ torch.save(model.module.state_dict(), os.path.join(output_dir, f"{num_batch_updates}step.pb"))
+ else:
+ torch.save(model.state_dict(), os.path.join(output_dir, f"{num_batch_updates}step.pb"))
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
@@ -687,6 +694,16 @@
scale_factor=0.55,
)
+ # for contextual training
+ if bias_grad_times != 1.0:
+ # contextual related parameter names
+ cr_pnames = ["bias_encoder", "bias_embed", "decoder.bias_decoder", "decoder.bias_output"]
+ for name, param in model.named_parameters():
+ for cr_pname in cr_pnames:
+ if cr_pname in name:
+ param.grad *= bias_grad_times
+ continue
+
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(),
--
Gitblit v1.9.1