开源库OpenNMT-py使用记录

一、准备

官网：http://opennmt.net/

python+torch版Github：https://github.com/OpenNMT/OpenNMT-py

python+torch版说明：http://zh.opennmt.net/OpenNMT-py/main.html

环境：ubuntu+python3+torch

需求：将图片公式转化为latex格式的半结构化字符串

案例：http://zh.opennmt.net/OpenNMT-py/im2text.html

二、过程

1.Download the data.

wget -O data/im2text.tgz http://lstm.seas.harvard.edu/latex/im2text_small.tgz; tar zxf data/im2text.tgz -C data/

2.Preprocess the data.

python3 preprocess.py -data_type img -src_dir data/im2text/images/ -train_src data/im2text/src-train.txt -train_tgt data/im2text/tgt-train.txt -valid_src data/im2text/src-val.txt -valid_tgt data/im2text/tgt-val.txt -save_data data/im2text/demo -tgt_seq_length 150 -tgt_words_min_frequency 2 -shard_size 500 -image_channel_size 1

3.Train the model.

nohup python3 train.py -model_type img -data data/im2text/demo -save_model demo-model -gpu_ranks 0 -batch_size 20 -max_grad_norm 20 -learning_rate 0.1 -word_vec_size 80 -encoder_type brnn -image_channel_size 1 &

4.Translate the images.

python3 translate.py -data_type img -model demo-model_step_*.pt -src_dir data/im2text/images -src data/im2text/src-test.txt -output pred.txt -max_length 150 -beam_size 5 -gpu 0 -verbose -image_channel_size 1

注意：-image_channel_size要设置，和前面两个步骤一致，否则会默认为3。

三、问题

问题1：

fjs@gpu1:~/OpenNMT-py$ python3 translate.py -data_type img -model demo-model_step_*.pt -src_dir data/im2text/images \
>                     -src data/im2text/src-test.txt -output pred.txt -max_length 150 -beam_size 5 -gpu 0 -verbose
Traceback (most recent call last):File "translate.py", line 33, in <module>main(opt)File "translate.py", line 14, in maintranslator = build_translator(opt, report_score=True)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 31, in build_translatorfields, model, model_opt = load_test_model(opt, dummy_opt.__dict__)File "/home/fjs/OpenNMT-py/onmt/decoders/ensemble.py", line 135, in load_test_model'Ensemble models must use the same preprocessed data'
AssertionError: Ensemble models must use the same preprocessed data

处理：在onmt/decoders/ensemble.py中注释点assert，这个不知有何影响，暂无法明确。

问题2：

fjs@gpu1:~/OpenNMT-py$ python3 translate.py -data_type img -model demo-model_step_*.pt -src_dir data/im2text/images                     -src data/im2text/src-test.txt -output pred.txt -max_length 150 -beam_size 5 -gpu 0 -verbose -image_channel_size 1
/usr/local/lib/python3.5/dist-packages/torchtext/data/field.py:359: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).var = torch.tensor(arr, dtype=self.dtype, device=device)
Traceback (most recent call last):File "translate.py", line 33, in <module>main(opt)File "translate.py", line 19, in mainattn_debug=opt.attn_debug)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 205, in translatebatch, data, attn_debug, fast=self.fastFile "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 309, in translate_batchreturn self._translate_batch(batch, data)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 621, in _translate_batchbatch, data_type)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 322, in _run_encoder'Ensemble decoding only supported for text data'
AssertionError: Ensemble decoding only supported for text data

是onmt/translate/translator.py的提示，增加代码如下：

问题3：

fjs@gpu1:~/OpenNMT-py$ python3 translate.py -data_type img -model demo-model_step_*.pt -src_dir data/im2text/images  -src data/im2text/src-test.txt -output pred.txt -max_length 150 -beam_size 5 -gpu 0 -verbose -image_channel_size 1
/usr/local/lib/python3.5/dist-packages/torchtext/data/field.py:359: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).var = torch.tensor(arr, dtype=self.dtype, device=device)
Traceback (most recent call last):File "translate.py", line 33, in <module>main(opt)File "translate.py", line 19, in mainattn_debug=opt.attn_debug)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 205, in translatebatch, data, attn_debug, fast=self.fastFile "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 309, in translate_batchreturn self._translate_batch(batch, data)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 646, in _translate_batchmemory_bank = tuple(tile(x, beam_size, dim=1) for x in memory_bank)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 646, in <genexpr>memory_bank = tuple(tile(x, beam_size, dim=1) for x in memory_bank)File "/home/fjs/OpenNMT-py/onmt/utils/misc.py", line 42, in tile.transpose(0, 1) \
RuntimeError: CUDA out of memory. Tried to allocate 435.00 MiB (GPU 0; 10.92 GiB total capacity; 9.64 GiB already allocated; 355.50 MiB free; 376.90 MiB cached)

调整执行参数：-beam_size 3

问题4：

fjs@gpu1:~/OpenNMT-py$ python3 translate.py -data_type img -model demo-model_step_*.pt -src_dir data/im2text/images  -src data/im2text/src-test.txt -output pred.txt -max_length 150 -beam_size 3 -gpu 0 -verbose -image_channel_size 1
/usr/local/lib/python3.5/dist-packages/torchtext/data/field.py:359: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).var = torch.tensor(arr, dtype=self.dtype, device=device)
Traceback (most recent call last):File "translate.py", line 33, in <module>main(opt)File "translate.py", line 19, in mainattn_debug=opt.attn_debug)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 205, in translatebatch, data, attn_debug, fast=self.fastFile "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 309, in translate_batchreturn self._translate_batch(batch, data)File "/home/fjs/OpenNMT-py/onmt/translate/translator.py", line 649, in _translate_batchmemory_lengths = tile(src_lengths, beam_size)File "/home/fjs/OpenNMT-py/onmt/utils/misc.py", line 32, in tileperm = list(range(len(x.size())))
AttributeError: 'int' object has no attribute 'size'

未解决？

开源库OpenNMT-py使用记录相关推荐

Google开源库Image Captioning部署记录
Github库:https://github.com/tensorflow/models/tree/master/research/im2txt 场景:给一张图片,描述图片内容.如下图: 环境:ubu ...
在别的地方看的给程序员介绍一些C++开源库，记录给大家共同学习
在别的地方看的<<给程序员介绍一些C++开源库>>,记录给大家共同学习首先说明这篇文章不是出自我手,大家共同学习. 引用地址:http://oss.org.cn/?actio ...
librtmp开源库与android平台编译记录
一.编写目的为了记录编译librtmp开源库过程遇到的问题,帮助后来人提供参考,希望本文能为大家提供参考. 二.简介(官方介绍) The Real-Time Messaging Protocol ( ...
记录几个Qt开源库界面库
QT 开源类库集合_Vinx Blog-CSDN博客_qt开源控件库https://blog.csdn.net/u011471873/article/details/108221797 QT 开源类库 ...
伯乐：一个易用、强大的PyTorch推荐系统开源库
来源:RUC AI Box 本文约3500字,建议阅读5分钟 4类53个模型,27个数据集,又一推荐系统神器! [ 导读 ]是否还在为推荐模型无法复现而怀疑人生?是否还在为不知如何入门推荐而踌躇不前? ...
软件经验|GDAL空间数据开源库开发介绍
GDAL(Geospatial Data Abstraction Library)是使用C/C++语言编写的用于读写空间数据的一套跨平台开源库.GDAL库可以读取.写入.转换.处理各种栅格数据格式,它 ...
人脸检测和识别的开源库总结
背景人脸相关的任务有以下几种: 人脸检测(将图片中包含的人脸进行初步定位,以及关键点如眼睛.鼻子.嘴巴等更精细的定位,文章封面即为典型人脸检测结果): 人脸跟踪(视频中跟踪人脸位置变化): 人脸验证 ...
Python 开源库及示例代码
Python 开源库及示例代码更多干货分布式实战(干货) spring cloud 实战(干货) mybatis 实战(干货) spring boot 实战(干货) React 入门实战(干货) ...
利用人工智能（Magpie开源库）给一段中文的文本内容进行分类打标签
当下人工智能是真心的火热呀,各种原来传统的业务也都在尝试用人工智能技术来处理,以此来节省人工成本,提高生产效率.既然有这么火的利器,那么我们就先来简单认识下什么是人工智能吧,人工智能是指利用语音识别. ...
常用C/C++开源库
1. 框架 Apache C++ Standard Library : 是一系列算法,容器,迭代器和其他基本组件的集合 ASL : Adobe源代码库提供了同行的评审和可移植的C++源代码库. Boo ...

开源库OpenNMT-py使用记录

开源库OpenNMT-py使用记录相关推荐

最新文章

热门文章