客服电话
APP下载
迪恩网络APP
随时随地掌握行业动态
官方微信
扫描二维码
关注迪恩网络微信公众号
开源软件名称（OpenSource Name）：lmthang/nmt.matlab开源软件地址(OpenSource Url)：https://github.com/lmthang/nmt.matlab开源编程语言(OpenSource Language)：MATLAB 52.5%开源软件介绍(OpenSource Introduction)：Effective Approaches to Attention-based Neural Machine TranslationsmCode to train Neural Machine Translation systems as described in our EMNLP paper Effective Approaches to Attention-based Neural Machine Translation. Features:
Citations:If you make use of this code in your research, please cite our paper
FilesThe code directory further divides into sub-directories: Examples
Here, we convert train/valid/test files in text format into integer format that can be handled efficiently in Matlab. The script syntax is:
In Matlab, go into the code/ directory and run: This trains a very basic model with all the default settings. We set 'isResume' to 0 so that it will train a new model each time you run the command instead of loading existing models. See trainLSTM.m for more. (To run directly from your terminal, checkout scripts/train.sh) The syntax is: The trainer code outputs logs such as: which means: at epoch 1, mini-batches 20, training speed is 6.38K words/s, learning rate is 1, train cost is 6.52, and grad norm is 11.62. Once in a while, the code will evaluate on the valid and test sets: which tells us additional information such as the current test perplexity, 34.40, the valid / test costs, and the average abs values of the model paramters.
Decode with beamSize 2, collect maximum 10 translations, batchSize 1. Note that the testLSTM implicitly decodes the test file specified during training. To specifiy a different test file, use 'testPrefix', see testLSTM.m for more. (To run directly from your terminal, checkout scripts/test.sh) Syntax:
Advanced Examples
We train a 2 stacking LSTM layers with 100 LSTM cells and 100-dim embeddings. Here, 0.8 means the dropout "keep" probability.
Apart from "obvious" hyperparameters, we want to scale our gradients whenever its norm averaged by batch size (128) is greater than 5. After training for 5 epochs, we start halving our learning rate each epoch. To have further control of learning rate schedule, see 'epochFraction' and 'finetuneRate' options in trainLSTM.m
Here, we also use source reversing 'isReverse' and the input feeding approach 'feedInput' as described in the paper. Other attention architectures can be specified as follows: 'isResume' is set to 0 to avoid loading existing models (done by default), so that you can try different attention architectures.
Then compare with the provided grad check outputs data/grad_checks.txt. They should look similar. Note: many different configurations will be run with the run_grad_checks.sh script. For many configuration, we set the 'initRange' to a large value 10, so you will notice the total gradient differences are large. This is to debug subtle mistakes; and if the total diff < 10, you can mostly be assured. We do note that with attnFunc=4, attnOpt=1, the diff is quite large; this is something to be checked though the model seems to work in practice. 
|
2023-10-27
2022-08-15
2022-08-17
2022-09-23
2022-08-13
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
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
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
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
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
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
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
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
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
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
请发表评论