Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
曾经的那一瞬间
Models
提交
012d7e74
M
Models
项目概览
曾经的那一瞬间
/
Models
大约 1 年 前同步成功
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
列表
看板
标记
里程碑
合并请求
0
DevOps
流水线
流水线任务
计划
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
M
Models
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
0
Issue
0
列表
看板
标记
里程碑
合并请求
0
合并请求
0
Pages
DevOps
DevOps
流水线
流水线任务
计划
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
流水线任务
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
012d7e74
编写于
11月 21, 2022
作者:
A
A. Unique TensorFlower
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Internal change
PiperOrigin-RevId: 490119364
上级
24a39da3
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
8 addition
and
3 deletion
+8
-3
official/nlp/modeling/layers/moe.py
official/nlp/modeling/layers/moe.py
+7
-2
official/nlp/modeling/layers/moe_test.py
official/nlp/modeling/layers/moe_test.py
+1
-1
未找到文件。
official/nlp/modeling/layers/moe.py
浏览文件 @
012d7e74
...
...
@@ -100,6 +100,7 @@ class Router(tf.keras.layers.Layer):
use_bias
:
bool
=
True
,
kernel_initializer
:
_InitializerType
=
_DEFAULT_KERNEL_INITIALIZER
,
bias_initializer
:
_InitializerType
=
_DEFAULT_BIAS_INITIALIZER
,
router_z_loss_weight
:
float
=
0.0
,
name
:
str
=
"router"
,
dtype
:
Any
=
tf
.
float32
,
**
kwargs
):
...
...
@@ -112,6 +113,8 @@ class Router(tf.keras.layers.Layer):
weights.
kernel_initializer: Kernel initializer for router weights.
bias_initializer: Bias initializer for router weights.
router_z_loss_weight: Weight for router_z_loss. Use non-zero values if
running into training instability (esp. with dtype 'bfloat16' or lower).
name: Layer name.
dtype: The dtype of the layer's computations and weights. tf.float32 is
recommended for stability.
...
...
@@ -122,7 +125,7 @@ class Router(tf.keras.layers.Layer):
self
.
num_experts
=
num_experts
# Used to check consistency with
# FeedForwardExperts.
self
.
jitter_noise
=
jitter_noise
self
.
router_z_loss_weight
=
router_z_loss_weight
self
.
router_weights
=
tf
.
keras
.
layers
.
Dense
(
num_experts
,
use_bias
=
use_bias
,
...
...
@@ -156,8 +159,10 @@ class Router(tf.keras.layers.Layer):
inputs
,
apply_jitter
=
training
)
# router_probs <float32>[num_groups, tokens_per_group, num_experts]
# router_logits <float>[num_groups, tokens_per_group, num_experts]
router_z_loss
=
_router_z_loss
(
router_logits
)
unscaled_router_z_loss
=
_router_z_loss
(
router_logits
)
router_z_loss
=
self
.
router_z_loss_weight
*
unscaled_router_z_loss
self
.
add_loss
(
router_z_loss
)
self
.
add_metric
(
unscaled_router_z_loss
,
name
=
"unscaled_router_z_loss"
)
self
.
add_metric
(
router_z_loss
,
name
=
"router_z_loss"
)
routing_instructions
=
self
.
_compute_routing_instructions
(
...
...
official/nlp/modeling/layers/moe_test.py
浏览文件 @
012d7e74
...
...
@@ -224,7 +224,7 @@ class MoeTest(tf.test.TestCase):
metrics
=
[
metric
.
name
for
metric
in
moe_layer
.
metrics
]
self
.
assertSetEqual
(
{
'router_z_loss'
,
'load_balancing_loss'
,
'router_z_loss'
,
'
unscaled_router_z_loss'
,
'
load_balancing_loss'
,
'fraction_tokens_left_behind'
,
'router_confidence'
,
'expert_usage'
},
set
(
metrics
))
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录