2 Star 3 Fork 1

xiaoxin_小新 / Open_LSTM

加入 Gitee
与超过 1200万 开发者一起发现、参与优秀开源项目,私有仓库也完全免费 :)
免费加入
克隆/下载
贡献代码
同步代码
取消
提示: 由于 Git 不支持空文件夾,创建文件夹后会生成空的 .keep 文件
Loading...
README
EPL-1.0

<<<<<<< HEAD

LSTM

什么是LSTM?

RNN 会受到短时记忆的影响。如果一条序列足够长,那它们将很难将信息从较早的时间步传送到后面的时间步。

1638954756174

​ $$H^{(t)}=\sigma(W^{ht}\cdot X{(t)}+W^{hh}\cdot H^{(t-1)}+b_n) $$

1638954884318

LSTM 是解决短时记忆问题的解决方案,它们具有称为**“门”**的内部机制,可以调节信息流。

这些“”可以知道序列中哪些重要的数据是需要保留,而哪些是要删除的。 随后,它可以沿着长链序列传递相关信息以进行预测,几乎所有基于循环神经网络的技术成果都是通过这两个网络实现的。

img

LSTM体系结构

LSTM被称为门结构:一些数学运算的组合,这些运算使信息流动或从计算图的那里保留下来。因此,它能够“决定”其长期和短期记忆,并输出对序列数据的可靠预测:

1639018164481

LSTM单元中的预测序列。注意,它不仅会传递预测值,而且还会传递一个c,c是长期记忆的代表

遗忘门

遗忘门(forget gate)是输入信息与候选者一起操作的门,作为长期记忆。在输入、隐藏状态和偏差的第一个线性组合上,应用一个sigmoid函数:

1638781376444

1639018143530

sigmoid将遗忘门的输出“缩放”到0-1之间,然后,通过将其与候选者相乘,我们可以将其设置为0,表示长期记忆中的“遗忘”,或者将其设置为更大的数字,表示我们从长期记忆中记住的“多少”。

输入门

输入门是将包含在输入和隐藏状态中的信息组合起来,然后与候选和部分候选c''u t一起操作的地方:

1638781443966

1638781473090

1639018106022

在这些操作中,决定了多少新信息将被引入到内存中,如何改变——这就是为什么我们使用tanh函数

(从-1到1)。我们将短期记忆和长期记忆中的部分候选组合起来,并将其设置为候选。

输出门和隐藏状态(输出)

之后,我们可以收集o_t作为LSTM单元的输出门,然后将其乘以候选单元(长期存储器)的tanh,后者已经用正确的操作进行了更新。网络输出为h_t。1638781533876

1639018057537

LSTM单元方程

​ $${f_t} = \sigma(U_fx_t+V_fh_{t-1}+b_f)$$

​ $$i_t = \sigma(U_ix_i+V_ih_{t-1}+b_i)$$

​ $$o_t = \sigma(U_ox_i+V_oh_{t-1}+b_0)$$

​ $$g_t = tanh(U_gx_t+V_gh_{t-1}+b_g) 等价于 \tilde{c_t}$$

​ $$c_t = f_t \cdot c_{t-1} + i_t \cdot g_t $$

​ $$h_t = o_t \cdot tanh(c_t)$$

BI-LSTM

双向LSTM的结构与双向RNN基本相同。但双向LSTM能够同时利用过去时刻和未来时刻的信息,会比单向LSTM的预测更加准确。

1639013651115

代码实现部分的一些方法详解:

主要对里面的使用的方法进行一个单独的介绍

hand_lstm.py:
nn.Parameter()  

可以使用nn.Parameter()来转换一个固定的权重数值,使的其可以跟着网络训练一直调优下去,学习到一个最适合的权重值。

也就是说其作用是使得当前的参数可以被保存梯度

torch.Tensor(2,1)

# 输出结果
tensor([[0.],
        [0.]])

生成一个 size 为(a,b) 的全0张量

当使用nn.Parameter()后,这个全0张量会被赋予一个随机值

nn.Parameter(torch.Tensor(21))  

# 输出结果
Parameter containing:
tensor([[9.1477e-41],
        [1.0739e-05]], requires_grad=True)

# requires_grad=True 保留梯度信息
stdv = 1.0 / math.sqrt(128)

​ $$stdv = {1\over\sqrt{128}}$$

uniform_(a,b) 在a,b中间tensor从均匀分布中抽样数值进行填充。

b = torch.randn(2,3)
# tensor([[ 0.7628, -1.6425, -1.4857],
#        [ 0.4624,  0.4417, -1.0390]])
b.uniform_(0, 1)
# tensor([[0.4614, 0.1125, 0.6703],
#         [0.1231, 0.2875, 0.3648]])

torch.zeros() 比较基础的,就是生成一个 全0的tensor

torch.zeros(2,3)
# tensor([[0., 0., 0.],
#        [0., 0., 0.]])

unsqueeze() 增加一个维度

c = torch.zeros(2,3)
c = c.unsqueeze(0)
# tensor([[[0., 0., 0.],
#          [0., 0., 0.]]])
c.size()
# torch.Size([1, 2, 3])

cat()  按第N维进行拼接  将这个list 转化为tensor结构

"""
l1:
[tensor([[[ 0.4480,  0.4832, -0.1113],
          [ 1.0129, -0.6316,  0.2096]]]),
 tensor([[[ 1.9288,  0.8110,  0.5142],
          [-0.2250,  2.0976,  0.0366]]])]
"""
torch.cat(l1,dim=0)

"""
tensor([[[ 0.4480,  0.4832, -0.1113],
         [ 1.0129, -0.6316,  0.2096]],

        [[ 1.9288,  0.8110,  0.5142],
         [-0.2250,  2.0976,  0.0366]]])
"""

l1.size()

# torch.Size([2, 2, 3])

transpose(input, dim0, dim1)

  • input (Tensor) – 输入张量,必填
  • dim0 (int) – 转置的第一维,默认0,可选
  • dim1 (int) – 转置的第二维,默认1,可选
"""
tensor([[[ 1.0085,  1.3747, -1.2875,  0.8424],
         [-0.2851, -1.5225, -1.1119, -1.5243],
         [-0.4864,  0.8463, -0.5516,  0.2596]],

        [[-0.8208, -1.9864, -1.3729, -0.5153],
         [-0.7217,  0.5975,  1.1512,  1.1755],
         [ 0.6672, -1.9025, -0.6683, -0.4970]]])
"""
l1.transpose(0,1)


"""tensor([[[ 1.0085,  1.3747, -1.2875,  0.8424],
         [-0.8208, -1.9864, -1.3729, -0.5153]],

        [[-0.2851, -1.5225, -1.1119, -1.5243],
         [-0.7217,  0.5975,  1.1512,  1.1755]],

        [[-0.4864,  0.8463, -0.5516,  0.2596],
         [ 0.6672, -1.9025, -0.6683, -0.4970]]])
"""
lstm_train.py

1.model.train() 和 model.eval()

简单来说,

是设置了训练或者测试模式,定义模型是否需要学习。对部分层有影响,如Dropout和BN。具体影响如下:

​ 1.Dropout: 训练过程中,为防止模型过拟合,增加其泛化性,会随机屏蔽掉一些神经元,相当于输入每次走过不同的“模型”。测试模式时,所有神经元共同作用,类似于boosting。

​ 2.BN: 训练过程中,模型每次处理一个minibatch数据,BN根据一个minibatch来计算mean和std后做归一化处理,这也是为什么模型的性能和minibatch的大小关系很大,测试时,BN会利用训练时得到的参数来处理测试数据。如果不设置model.eval(),输入单张图像,会报错。

​ 3.model. train()和model. eval()可以看做是对这种训练和测试需要联动的模块进行一个统一的设置。当你在写model的时候,你写的是测试和训练通用的model,这个时候,就是通过model. train()和model. eval()来来设置model的测试阶段和训练阶段。这样在用需要训练和测试联动的模块的时候,就不用再专门写一个训练的model和一个测试的model了

2.Variable ()

tensor不能反向传播,variable可以反向传播。

Variable计算时,它会逐渐地生成计算图。这个图就是将所有的计算节点都连接起来,最后进行误差反向传递的时候,一次性将所有Variable里面的梯度都计算出来,而tensor就没有这个能力。

运行结果:

1639014974043

BI-LSTM:

1639015108720

Hand-LSTM:

1639015480828

Hand-BI-LSTM:

1639017512993

Eclipse Public License - v 1.0 THE ACCOMPANYING PROGRAM IS PROVIDED UNDER THE TERMS OF THIS ECLIPSE PUBLIC LICENSE ("AGREEMENT"). ANY USE, REPRODUCTION OR DISTRIBUTION OF THE PROGRAM CONSTITUTES RECIPIENT'S ACCEPTANCE OF THIS AGREEMENT. 1. DEFINITIONS "Contribution" means: a) in the case of the initial Contributor, the initial code and documentation distributed under this Agreement, and b) in the case of each subsequent Contributor: i) changes to the Program, and ii) additions to the Program; where such changes and/or additions to the Program originate from and are distributed by that particular Contributor. A Contribution 'originates' from a Contributor if it was added to the Program by such Contributor itself or anyone acting on such Contributor's behalf. Contributions do not include additions to the Program which: (i) are separate modules of software distributed in conjunction with the Program under their own license agreement, and (ii) are not derivative works of the Program. "Contributor" means any person or entity that distributes the Program. "Licensed Patents" mean patent claims licensable by a Contributor which are necessarily infringed by the use or sale of its Contribution alone or when combined with the Program. "Program" means the Contributions distributed in accordance with this Agreement. "Recipient" means anyone who receives the Program under this Agreement, including all Contributors. 2. GRANT OF RIGHTS a) Subject to the terms of this Agreement, each Contributor hereby grants Recipient a non-exclusive, worldwide, royalty-free copyright license to reproduce, prepare derivative works of, publicly display, publicly perform, distribute and sublicense the Contribution of such Contributor, if any, and such derivative works, in source code and object code form. b) Subject to the terms of this Agreement, each Contributor hereby grants Recipient a non-exclusive, worldwide, royalty-free patent license under Licensed Patents to make, use, sell, offer to sell, import and otherwise transfer the Contribution of such Contributor, if any, in source code and object code form. This patent license shall apply to the combination of the Contribution and the Program if, at the time the Contribution is added by the Contributor, such addition of the Contribution causes such combination to be covered by the Licensed Patents. The patent license shall not apply to any other combinations which include the Contribution. No hardware per se is licensed hereunder. c) Recipient understands that although each Contributor grants the licenses to its Contributions set forth herein, no assurances are provided by any Contributor that the Program does not infringe the patent or other intellectual property rights of any other entity. Each Contributor disclaims any liability to Recipient for claims brought by any other entity based on infringement of intellectual property rights or otherwise. As a condition to exercising the rights and licenses granted hereunder, each Recipient hereby assumes sole responsibility to secure any other intellectual property rights needed, if any. For example, if a third party patent license is required to allow Recipient to distribute the Program, it is Recipient's responsibility to acquire that license before distributing the Program. d) Each Contributor represents that to its knowledge it has sufficient copyright rights in its Contribution, if any, to grant the copyright license set forth in this Agreement. 3. REQUIREMENTS A Contributor may choose to distribute the Program in object code form under its own license agreement, provided that: a) it complies with the terms and conditions of this Agreement; and b) its license agreement: i) effectively disclaims on behalf of all Contributors all warranties and conditions, express and implied, including warranties or conditions of title and non-infringement, and implied warranties or conditions of merchantability and fitness for a particular purpose; ii) effectively excludes on behalf of all Contributors all liability for damages, including direct, indirect, special, incidental and consequential damages, such as lost profits; iii) states that any provisions which differ from this Agreement are offered by that Contributor alone and not by any other party; and iv) states that source code for the Program is available from such Contributor, and informs licensees how to obtain it in a reasonable manner on or through a medium customarily used for software exchange. When the Program is made available in source code form: a) it must be made available under this Agreement; and b) a copy of this Agreement must be included with each copy of the Program. Contributors may not remove or alter any copyright notices contained within the Program. Each Contributor must identify itself as the originator of its Contribution, if any, in a manner that reasonably allows subsequent Recipients to identify the originator of the Contribution. 4. COMMERCIAL DISTRIBUTION Commercial distributors of software may accept certain responsibilities with respect to end users, business partners and the like. While this license is intended to facilitate the commercial use of the Program, the Contributor who includes the Program in a commercial product offering should do so in a manner which does not create potential liability for other Contributors. Therefore, if a Contributor includes the Program in a commercial product offering, such Contributor ("Commercial Contributor") hereby agrees to defend and indemnify every other Contributor ("Indemnified Contributor") against any losses, damages and costs (collectively "Losses") arising from claims, lawsuits and other legal actions brought by a third party against the Indemnified Contributor to the extent caused by the acts or omissions of such Commercial Contributor in connection with its distribution of the Program in a commercial product offering. The obligations in this section do not apply to any claims or Losses relating to any actual or alleged intellectual property infringement. In order to qualify, an Indemnified Contributor must: a) promptly notify the Commercial Contributor in writing of such claim, and b) allow the Commercial Contributor to control, and cooperate with the Commercial Contributor in, the defense and any related settlement negotiations. The Indemnified Contributor may participate in any such claim at its own expense. For example, a Contributor might include the Program in a commercial product offering, Product X. That Contributor is then a Commercial Contributor. If that Commercial Contributor then makes performance claims, or offers warranties related to Product X, those performance claims and warranties are such Commercial Contributor's responsibility alone. Under this section, the Commercial Contributor would have to defend claims against the other Contributors related to those performance claims and warranties, and if a court requires any other Contributor to pay any damages as a result, the Commercial Contributor must pay those damages. 5. NO WARRANTY EXCEPT AS EXPRESSLY SET FORTH IN THIS AGREEMENT, THE PROGRAM IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, EITHER EXPRESS OR IMPLIED INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OR CONDITIONS OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Each Recipient is solely responsible for determining the appropriateness of using and distributing the Program and assumes all risks associated with its exercise of rights under this Agreement , including but not limited to the risks and costs of program errors, compliance with applicable laws, damage to or loss of data, programs or equipment, and unavailability or interruption of operations. 6. DISCLAIMER OF LIABILITY EXCEPT AS EXPRESSLY SET FORTH IN THIS AGREEMENT, NEITHER RECIPIENT NOR ANY CONTRIBUTORS SHALL HAVE ANY LIABILITY FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING WITHOUT LIMITATION LOST PROFITS), HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OR DISTRIBUTION OF THE PROGRAM OR THE EXERCISE OF ANY RIGHTS GRANTED HEREUNDER, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. 7. GENERAL If any provision of this Agreement is invalid or unenforceable under applicable law, it shall not affect the validity or enforceability of the remainder of the terms of this Agreement, and without further action by the parties hereto, such provision shall be reformed to the minimum extent necessary to make such provision valid and enforceable. If Recipient institutes patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Program itself (excluding combinations of the Program with other software or hardware) infringes such Recipient's patent(s), then such Recipient's rights granted under Section 2(b) shall terminate as of the date such litigation is filed. All Recipient's rights under this Agreement shall terminate if it fails to comply with any of the material terms or conditions of this Agreement and does not cure such failure in a reasonable period of time after becoming aware of such noncompliance. If all Recipient's rights under this Agreement terminate, Recipient agrees to cease use and distribution of the Program as soon as reasonably practicable. However, Recipient's obligations under this Agreement and any licenses granted by Recipient relating to the Program shall continue and survive. Everyone is permitted to copy and distribute copies of this Agreement, but in order to avoid inconsistency the Agreement is copyrighted and may only be modified in the following manner. The Agreement Steward reserves the right to publish new versions (including revisions) of this Agreement from time to time. No one other than the Agreement Steward has the right to modify this Agreement. The Eclipse Foundation is the initial Agreement Steward. The Eclipse Foundation may assign the responsibility to serve as the Agreement Steward to a suitable separate entity. Each new version of the Agreement will be given a distinguishing version number. The Program (including Contributions) may always be distributed subject to the version of the Agreement under which it was received. In addition, after a new version of the Agreement is published, Contributor may elect to distribute the Program (including its Contributions) under the new version. Except as expressly stated in Sections 2(a) and 2(b) above, Recipient receives no rights or licenses to the intellectual property of any Contributor under this Agreement, whether expressly, by implication, estoppel or otherwise. All rights in the Program not expressly granted under this Agreement are reserved. This Agreement is governed by the laws of the State of New York and the intellectual property laws of the United States of America. No party to this Agreement will bring a legal action under this Agreement more than one year after the cause of action arose. Each party waives its rights to a jury trial in any resulting litigation.

简介

LSTM学习 展开 收起
Python 等 2 种语言
EPL-1.0
取消

发行版

暂无发行版

贡献者

全部

近期动态

加载更多
不能加载更多了
Python
1
https://gitee.com/xiaoxin_gitee/Open_LSTM.git
git@gitee.com:xiaoxin_gitee/Open_LSTM.git
xiaoxin_gitee
Open_LSTM
Open_LSTM
master

搜索帮助

53164aa7 5694891 3bd8fe86 5694891