From 24af4286d5d4a49a160370a3bc58e63be5e96e21 Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 11 六月 2024 15:26:35 +0800
Subject: [PATCH] modify the qformer adaptor (#1804)
---
funasr/models/sanm/attention.py | 16 ++++++++++++----
1 files changed, 12 insertions(+), 4 deletions(-)
diff --git a/funasr/models/sanm/attention.py b/funasr/models/sanm/attention.py
index da8850f..08f7dc7 100644
--- a/funasr/models/sanm/attention.py
+++ b/funasr/models/sanm/attention.py
@@ -100,7 +100,9 @@
n_batch = value.size(0)
if mask is not None:
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
- min_value = float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+ min_value = -float(
+ "inf"
+ ) # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
scores = scores.masked_fill(mask, min_value)
self.attn = torch.softmax(scores, dim=-1).masked_fill(
mask, 0.0
@@ -269,7 +271,9 @@
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
- min_value = float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+ min_value = -float(
+ "inf"
+ ) # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
scores = scores.masked_fill(mask, min_value)
self.attn = torch.softmax(scores, dim=-1).masked_fill(
mask, 0.0
@@ -673,7 +677,9 @@
n_batch = value.size(0)
if mask is not None:
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
- min_value = float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+ min_value = -float(
+ "inf"
+ ) # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
# logging.info(
# "scores: {}, mask_size: {}".format(scores.size(), mask.size()))
scores = scores.masked_fill(mask, min_value)
@@ -858,7 +864,9 @@
mask = mask.unsqueeze(1).eq(0) # (batch, 1, *, time2)
- min_value = float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
+ min_value = -float(
+ "inf"
+ ) # float(numpy.finfo(torch.tensor(0, dtype=scores.dtype).numpy().dtype).min)
scores = scores.masked_fill(mask, min_value)
self.attn = torch.softmax(scores, dim=-1).masked_fill(
mask, 0.0
--
Gitblit v1.9.1