bshillingford/python-torchfile: Deserialize Lua torch-serialized objects from Py ...

原作者: [db:作者] 来自: 网络收藏邀请

开源软件名称（OpenSource Name）：

bshillingford/python-torchfile

开源软件地址(OpenSource Url)：

https://github.com/bshillingford/python-torchfile

开源编程语言(OpenSource Language)：

Python 100.0%

开源软件介绍(OpenSource Introduction)：

Torch serialization reader for Python

Mostly direct port of the torch7 Lua and C serialization implementation to Python, depending only on numpy (and the standard library: array and struct). Sharing of objects including torch.Tensors is preserved.

import torchfile
stuff = torchfile.load('a_bunch_of_stuff.t7')

Installation:

Install from PyPI:

pip install torchfile

or clone this repository, then:

python setup.py install

Supports Python 2.7, 3.4, 3.5, 3.6. Probably others too.

More examples:

Write from torch, read from Python:

Lua:

+th> torch.save('/tmp/test.t7', {hello=123, world=torch.rand(1,2,3)})

Python:

In [3]: o = torchfile.load('/tmp/test.t7')
In [4]: print o['world'].shape
(1, 2, 3)
In [5]: o
Out[5]: 
{'hello': 123, 'world': array([[[ 0.52291083,  0.29261517,  0.11113465],
         [ 0.01017287,  0.21466237,  0.26572137]]])}

Arbitary torch classes supported:

In [1]: import torchfile

In [2]: o = torchfile.load('testfiles_x86_64/gmodule_with_linear_identity.t7')

In [3]: o.forwardnodes[3].data.module
Out[3]: TorchObject(nn.Identity, {'output': array([], dtype=float64), 'gradInput': array([], dtype=float64)})

In [4]: for node in o.forwardnodes: print(repr(node.data.module))                                                                                                            
None
None
None
TorchObject(nn.Identity, {'output': array([], dtype=float64), 'gradInput': array([], dtype=float64)})
None
TorchObject(nn.Identity, {'output': array([], dtype=float64), 'gradInput': array([], dtype=float64)})
TorchObject(nn.Linear, {'weight': array([[-0.0248373 ],
       [ 0.17503954]]), 'gradInput': array([], dtype=float64), 'gradWeight': array([[  1.22317168e-312],
       [  1.22317168e-312]]), 'bias': array([ 0.05159848, -0.25367146]), 'gradBias': array([  1.22317168e-312,   1.22317168e-312]), 'output': array([], dtype=float64)})
TorchObject(nn.CAddTable, {'output': array([], dtype=float64), 'gradInput': []})
None

In [5]: o.forwardnodes[6].data.module.weight
Out[5]: 
array([[-0.0248373 ],
       [ 0.17503954]])

In [6]: o.forwardnodes[6].data.module.bias
Out[6]: array([ 0.05159848, -0.25367146])

More complex writing from torch:

Lua:

+th> f = torch.DiskFile('/tmp/test.t7', 'w'):binary()
+th> f:writeBool(false)
+th> f:writeObject({hello=123})
+th> f:writeInt(456)
+th> f:close()

Python:

In [1]: import torchfile
In [2]: with open('/tmp/test.t7','rb') as f:
   ...:     r = torchfile.T7Reader(f)
   ...:     print(r.read_boolean())
   ...:     print(r.read_obj())
   ...:     print(r.read_int())
   ...: 
False
{'hello': 123}
456

Supported types:

nil to Python None
numbers to Python floats, or by default a heuristic changes them to ints or longs if they are integral
booleans
strings: read as byte strings (Python 3) or normal strings (Python 2), like lua strings which don't support unicode, and that can contain null chars
tables converted to a special dict (*); if they are list-like (i.e. have numeric keys from 1 through n) they become a python list by default
Torch classes: supports Tensors and Storages, and most classes such as modules. Trivially extensible much like the Torch serialization code. Trivial torch classes like most nn.Module subclasses become TorchObjects. The torch_readers dict contains the mapping from class names to reading functions.
functions: loaded into the LuaFunction namedtuple, which simply wraps the raw serialized data, i.e. upvalues and code. These are mostly useless, but exist so you can deserialize anything.
tds.Hash, tds.Vec

(*) Since Lua allows you to index a table with a table but Python does not, we replace dicts with a subclass that is hashable, and change its equality comparison behaviour to compare by reference. See hashable_uniq_dict.

Test files demonstrating various features:

In [1]: import torchfile

In [2]: torchfile.load('testfiles_x86_64/list_table.t7')
Out[2]: ['hello', 'world', 'third item', 123]

In [3]: torchfile.load('testfiles_x86_64/doubletensor.t7')
Out[3]: 
array([[ 1. ,  2. ,  3. ],
       [ 4. ,  5. ,  6.9]])

# ...also other files demonstrating various types.

The example t7 files will work on any modern Intel or AMD 64-bit CPU, but the code will use the native byte ordering etc. Currently, the implementation assumes the system-dependent binary Torch format, but minor refactoring can give support for the ascii format as well.

鲜花

握手

雷人

路过

鸡蛋

该文章已有0人参与评论

请发表评论

全部评论

专题导读

More+

10-27 六六分期app的软件客服如何联系？(六六分期

11-06 可心卡盟:win10系统火狐flash插件崩溃怎么

11-06 亲亲特价:怎么删除回收站图标

11-06 济南大学虚拟社区:鲁大师节能降温的具体办

11-06 xlueops.exe:无线网络安装向导

11-06 女斗合众国:win7系统cf与主机连接不稳定怎

11-06 0xc000022-[cf烟雾头]cf怎么调烟雾头

11-06 qizideyouhuo:应用程序无法正常启动0xc0000

11-06 ipz-185:win7系统vcf文件怎么打开

11-06 傻哥蹦迪:win10系统s4怎么打开usb调试

11-06 八神浩树gtaste:回收站清空了怎么恢复

11-06 妖尾之黑色守护:win10系统电脑没有1440x900

11-06 校园至尊魔王小说:win7系统浏览网页时字体

11-06 女斗合众国:win10系统访问共享文件夹提示请

11-06 tokyo hot n0654:恢复win7系统默认字体一招

11-06 雨酷仙境:设置win7系统转移临时文件夹腾出

11-06 阿穆纳伊之杖:win7系统开始菜单在右边还原

11-06 tunespotting:win10系统火狐flash插件总是

11-06 甘尔葛分析师：计谋网站seo关键词暴涨有什

11-06 蔡贵霖: 计谋网站seo关键词暴涨有什么秘密

11-06 博益网首页:ao3网页版进入不了解决方法

11-06 漏斗子专栏: 网站数据分析小白易懂精华篇

11-06 见证双虹怎么做:win7系统开启telnet命令的

11-06 颾狐蝶蜋:系统资源不足无法完成请求的服务

11-06 国光中学校歌:提交网站到alexa查询详细步骤

11-06 西安有情天:静态网页和动态网页的区别

11-06 红木雅尚斋:外部链接构造对网站的好处

11-06 前官礼遇：防止域名劫持–增强域安全性的10

11-06 密传二转答案: 中文分词算法有哪些

11-06 金泉家园邮编:百度快照劫持的表现及应对方

mongrel2/Tir: A Simple Lua Web Framework For Mongrel2发布时间：2022-08-16

zhandouxiaojiji/behavior3lua: behavior tree for lua发布时间：2022-08-16

剪的笔顺,诠释剪的笔画,认识剪的部首

1 六六分期app的软件客服如何联系？(六六分期

六六分期app的软件客服如何联系？不知道吗？加qq群【895510560】即可！标题：六六分期

阅读：18982|2023-10-27

2 可心卡盟:win10系统火狐flash插件崩溃怎么

今天小编告诉大家如何处理win10系统火狐flash插件总是崩溃的问题，可能很多用户都不知

阅读：9920|2022-11-06

3 亲亲特价:怎么删除回收站图标

今天小编告诉大家如何对win10系统删除桌面回收站图标进行设置，可能很多用户都不知道

阅读：8302|2022-11-06

4 济南大学虚拟社区:鲁大师节能降温的具体办

今天小编告诉大家如何对win10系统电脑设置节能降温的设置方法，想必大家都遇到过需要

阅读：8667|2022-11-06

5 xlueops.exe:无线网络安装向导

我们在使用xp系统的过程中,经常需要对xp系统无线网络安装向导设置进行设置，可能很多

阅读：8600|2022-11-06

6 女斗合众国:win7系统cf与主机连接不稳定怎

今天小编告诉大家如何处理win7系统玩cf老是与主机连接不稳定的问题，可能很多用户都不

阅读：9611|2022-11-06

7 0xc000022-[cf烟雾头]cf怎么调烟雾头

电脑对日常生活的重要性小编就不多说了，可是一旦碰到win7系统设置cf烟雾头的问题，很

阅读：8590|2022-11-06

8 qizideyouhuo:应用程序无法正常启动0xc0000

我们在日常使用电脑的时候，有的小伙伴们可能在打开应用的时候会遇见提示应用程序无法

阅读：7974|2022-11-06

9 ipz-185:win7系统vcf文件怎么打开

今天小编告诉大家如何对win7系统打开vcf文件进行设置，可能很多用户都不知道怎么对win

阅读：8604|2022-11-06

10 傻哥蹦迪:win10系统s4怎么打开usb调试

今天小编告诉大家如何对win10系统s4开启USB调试模式进行设置，可能很多用户都不知道怎

阅读：7516|2022-11-06

客服电话

电子邮件

bshillingford/python-torchfile: Deserialize Lua torch-serialized objects from Py ...

开源软件名称（OpenSource Name）：

开源软件地址(OpenSource Url)：

开源编程语言(OpenSource Language)：

开源软件介绍(OpenSource Introduction)：

Torch serialization reader for Python

Installation:

More examples:

Write from torch, read from Python:

Arbitary torch classes supported:

More complex writing from torch:

Supported types:

Test files demonstrating various features:

请发表评论

全部评论

上一篇：

下一篇：

latextemplates/scientific-thesis-templat

krishnaik06/Machine-Learning-in-90-days

armancodv/building-energy-model-matlab:

美元符号为什么是“$”

FGRibreau/import-tweets-to-mastodon: How

剪的笔顺,诠释剪的笔画,认识剪的部首

六六分期app的软件客服如何联系？(六六分期

florent37/ViewAnimator: A fluent Android

florent37/Shrine-MaterialDesign2: implem

CVE-2020-36276

SimpleSoftwareIO/simple-sms: Send and re

关于我们

产品与服务

解决方案

139-2527-9053