mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 01:52:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

remove batch shaving

Browse Source
This commit is contained in:
yfyeung 2025-04-28 04:37:57 -07:00 committed by Your Name
parent f5d2aa1f5d
commit 46b9be31cc
1 changed files with 0 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
egs/speech_llm/ASR_LLM/whisper_llm_zh/train.py
View File

@ -71,7 +71,6 @@ from icefall.env import get_env_info
from icefall.utils import (
AttributeDict,
MetricsTracker,
filter_uneven_sized_batch,
setup_logger,
str2bool,
)
@ -278,13 +277,6 @@ def compute_loss(
Returns:
Return a tuple of two elements. The first element is the loss tensor.
"""
# For the uneven-sized batch, the total duration after padding would possibly
# cause OOM. Hence, for each batch, which is sorted descendingly by length,
# we simply drop the last few shortest samples, so that the retained total frames
# (after padding) would not exceed `allowed_max_frames`:
# `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`,
# where `max_frames = max_duration * 1000 // frame_shift_ms`.
# We set allowed_excess_duration_ratio=0.1.
def preprocess(
messages,
@ -375,10 +367,6 @@ def compute_loss(
text = text.replace("？", "")
return text
max_frames = params.max_duration * 1000 // params.frame_shift_ms
allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio))
batch = filter_uneven_sized_batch(batch, allowed_max_frames)
device = next(model.parameters()).device
feature = batch["inputs"]