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
MegEngine 天元
MegEngine
提交
992a90bb
MegEngine
项目概览
MegEngine 天元
/
MegEngine
1 年多 前同步成功
通知
403
Star
4705
Fork
582
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
MegEngine
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
提交
992a90bb
编写于
3月 16, 2021
作者:
M
Megvii Engine Team
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
docs(mge/quantization): add docstring for Observer
GitOrigin-RevId: 043be3886dc05205426fd60c9af3bc6172c70ce9
上级
c45f1eb2
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
48 addition
and
2 deletion
+48
-2
imperative/python/megengine/quantization/observer.py
imperative/python/megengine/quantization/observer.py
+48
-2
未找到文件。
imperative/python/megengine/quantization/observer.py
浏览文件 @
992a90bb
...
...
@@ -26,9 +26,10 @@ logger = get_logger(__name__)
class
Observer
(
Module
,
QParamsModuleMixin
):
r
"""
A base class for Observer Module.
A base class for Observer Module. Used to record input tensor's statistics for
quantization.
:param dtype: a string indicating
to collect scale and zero_point of which dtype
.
:param dtype: a string indicating
which dtype to collect scale and zero_point of
.
"""
def
__init__
(
self
,
dtype
:
Union
[
str
,
QuantDtypeMeta
],
**
kwargs
):
...
...
@@ -72,6 +73,14 @@ class Observer(Module, QParamsModuleMixin):
class
MinMaxObserver
(
Observer
):
r
"""
A Observer Module records input tensor's running min and max values to calc scale.
:param mode: set quantization mode.
:param eps: a initial maximum value to avoid division by zero problem.
:param dtype: a string indicating which dtype to collect scale and zero_point of.
"""
def
__init__
(
self
,
mode
:
QuantMode
=
QuantMode
.
SYMMERTIC
,
...
...
@@ -119,6 +128,14 @@ class MinMaxObserver(Observer):
class
SyncMinMaxObserver
(
MinMaxObserver
):
r
"""
A distributed version of :class:`~.MinMaxObserver`.
:param mode: set quantization mode.
:param eps: a initial maximum value to avoid division by zero problem.
:param dtype: a string indicating which dtype to collect scale and zero_point of.
"""
def
forward
(
self
,
x_orig
):
if
self
.
enable
:
x
=
x_orig
.
detach
()
...
...
@@ -134,6 +151,15 @@ class SyncMinMaxObserver(MinMaxObserver):
class
ExponentialMovingAverageObserver
(
MinMaxObserver
):
r
"""
A :class:`~.MinMaxObserver` with momentum support for min/max updating.
:param momentum: momentum ratio for min/max updating.
:param mode: set quantization mode.
:param eps: a initial maximum value to avoid division by zero problem.
:param dtype: a string indicating which dtype to collect scale and zero_point of.
"""
def
__init__
(
self
,
momentum
:
float
=
0.9
,
...
...
@@ -170,6 +196,15 @@ class ExponentialMovingAverageObserver(MinMaxObserver):
class
SyncExponentialMovingAverageObserver
(
ExponentialMovingAverageObserver
):
r
"""
A distributed version of :class:`~.ExponentialMovingAverageObserver`.
:param momentum: momentum ratio for min/max updating.
:param mode: set quantization mode.
:param eps: a initial maximum value to avoid division by zero problem.
:param dtype: a string indicating which dtype to collect scale and zero_point of.
"""
def
forward
(
self
,
x_orig
):
if
self
.
enabled
:
x
=
x_orig
.
detach
()
...
...
@@ -192,6 +227,17 @@ class SyncExponentialMovingAverageObserver(ExponentialMovingAverageObserver):
class
HistogramObserver
(
MinMaxObserver
):
r
"""
A :class:`~.MinMaxObserver` using running histogram of tensor values
for min/max updating. Usually used for calibration quantization.
:param bins: number of bins to use for the histogram.
:param upsample_rate: which ratio to interpolate histograms in.
:param mode: set quantization mode.
:param eps: a initial maximum value to avoid division by zero problem.
:param dtype: a string indicating which dtype to collect scale and zero_point of.
"""
def
__init__
(
self
,
bins
:
int
=
2048
,
...
...
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