大模型笔记1: Longformer环境配置
论文:
https://arxiv.org/abs/2004.05150
目录
库安装
LongformerForQuestionAnswering
库安装
首先保证电脑上配置了git.
git环境配置:
https://blog.csdn.net/Andone_hsx/article/details/87937329
| 3.1、找到git安装路径中bin的位置，如：D:\Program Files\Git\bin 找到git安装路径中git-core的位置，如：D:\Program Files\Git\libexec\git-core; 注："D:\Program Files\Git\"是安装路径，可能与你的安装路径不一样，要按照你自己的路径替换"D:\Program Files\Git\" 3.2、右键“计算机”->“属性”->“高级系统设置”->“环境变量”->在下方的“系统变量”中找到“path”->选中“path”并选择“编辑”->将 3.1中找到的bin和git-core路径复制到其中->保存并退出 注：“path”中，每个路径之间要以英文输入状态下的分号——“;”作为间隔 |
| D:\Program Files\Git\mingw64\bin D:\Program Files\Git\mingw64\libexec\git-core |
安装环境
| conda create --name longformer python=3.7 y conda activate longformer conda install cudatoolkit=10.0 y pip install git+https://github.com/allenai/longformer.git |
报错:
| ERROR: Could not find a version that satisfies the requirement pandas>=0.20.3 (from test-tube) (from versions: none) ERROR: No matching distribution found for pandas>=0.20.3 |
| No module named 'pandas' |
Install装不上, 在anaconda navigator装的
更换清华源后似乎可以继续运行了, 参考:
https://www.cnblogs.com/raiuny/p/15950043.html
| conda config --add channels Index of /anaconda/cloud/pytorch/ | 清华大学开源软件镜像站 | Tsinghua Open Source Mirror conda config --set show_channel_urls yes conda config --set auto_activate_base false pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple |
好几次报错128, 也许是RP问题, 总之重新运行几次后安装成功了.
环境安装成功会显示successful.
接着跑tests文件夹的test_readme.py, 注意需要下载longformer-base-4096.tar.gz
这个文件我放在项目目录下的/tmp文件夹和tests/tmp都无法读取, 因此修改了self.model_dir路径为绝对路径, 并注释下载解压代码, 就可以运行了:
LongformerForQuestionAnswering
1)test_readme中默认使用的Longformer模型输出是embedding, 缺少LMHead把embedding映射成tokenid或logits等, 无法输出文字. 如果使用Longformer完成最终任务, 需要自己写映射并训练.
2)文档其它longformer模型. 大部分为分类模型. 其中LongformerForQuestionAnswering符合extractive summarization
3)coding过程中可以参考huggingface上的文档例子从transformers库里面调用其它种类的longformer
| from transformers import AutoTokenizer, LongformerForQuestionAnswering import torch tokenizer = AutoTokenizer.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa") model = LongformerForQuestionAnswering.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa") question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet" encoding = tokenizer(question, text, return_tensors="pt") input_ids = encoding["input_ids"] # default is local attention everywhere # the forward method will automatically set global attention on question tokens attention_mask = encoding["attention_mask"] outputs = model(input_ids, attention_mask=attention_mask) start_logits = outputs.start_logits end_logits = outputs.end_logits all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0].tolist()) answer_tokens = all_tokens[torch.argmax(start_logits) : torch.argmax(end_logits) + 1] answer = tokenizer.decode( tokenizer.convert_tokens_to_ids(answer_tokens) ) # remove space prepending space token |
如果加载其它qa模型(longformer_base_4096_QA_SQUAD)不配套会报错:
| Some weights of the model checkpoint at tmp/longformer_base_4096_QA_SQUAD were not used when initializing LongformerForQuestionAnswering: ['classifier.dense.weight', 'classifier.dense.bias', 'classifier.out_proj.weight', 'classifier.out_proj.bias'] |
按照示例代码加载longformer-large-4096-finetuned-triviaqa后报错
| start_logits = outputs.start_logits AttributeError: 'tuple' object has no attribute 'start_logits' |
这个报错的意思是返回值不是对象而是元组, 因此判断如果是元组, 则手动解析
| if isinstance(outputs, tuple): loss,start_logits, end_logits,hidden_states,attentions = outputs else: start_logits = outputs.start_logits end_logits = outputs.end_logits |
| 库中LongformerForQuestionAnswering类代码有两处可能返回 1. output = (start_logits, end_logits) + outputs[2:] 2. return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) |