@@ -6,94 +6,108 @@ | |||
![Hex.pm](https://img.shields.io/hexpm/l/plug.svg) | |||
[![Documentation Status](https://readthedocs.org/projects/fastnlp/badge/?version=latest)](http://fastnlp.readthedocs.io/?badge=latest) | |||
FastNLP is a modular Natural Language Processing system based on PyTorch, built for fast development of NLP models. | |||
fastNLP 是一款轻量级的 NLP 处理套件。你既可以使用它快速地完成一个命名实体识别(NER)、中文分词或文本分类任务; 也可以使用他构建许多复杂的网络模型,进行科研。它具有如下的特性: | |||
- 统一的Tabular式数据容器,让数据预处理过程简洁明了。内置多种数据集的DataSet Loader,省去预处理代码。 | |||
- 各种方便的NLP工具,例如预处理embedding加载; 中间数据cache等; | |||
- 详尽的中文文档以供查阅; | |||
- 提供诸多高级模块,例如Variational LSTM, Transformer, CRF等; | |||
- 封装CNNText,Biaffine等模型可供直接使用; | |||
- 便捷且具有扩展性的训练器; 提供多种内置callback函数,方便实验记录、异常捕获等。 | |||
## 安装指南 | |||
fastNLP 依赖如下包: | |||
+ numpy | |||
+ torch>=0.4.0 | |||
+ tqdm | |||
+ nltk | |||
其中torch的安装可能与操作系统及 CUDA 的版本相关,请参见 PyTorch 官网 。 | |||
在依赖包安装完成的情况,您可以在命令行执行如下指令完成安装 | |||
```shell | |||
pip install fastNLP | |||
``` | |||
## 内置组件 | |||
大部分用于的 NLP 任务神经网络都可以看做由编码(encoder)、聚合(aggregator)、解码(decoder)三种模块组成。 | |||
![](./docs/source/figures/text_classification.png) | |||
fastNLP 在 modules 模块中内置了三种模块的诸多组件,可以帮助用户快速搭建自己所需的网络。 三种模块的功能和常见组件如下: | |||
A deep learning NLP model is the composition of three types of modules: | |||
<table> | |||
<tr> | |||
<td><b> module type </b></td> | |||
<td><b> functionality </b></td> | |||
<td><b> example </b></td> | |||
<td><b> 类型 </b></td> | |||
<td><b> 功能 </b></td> | |||
<td><b> 例子 </b></td> | |||
</tr> | |||
<tr> | |||
<td> encoder </td> | |||
<td> encode the input into some abstract representation </td> | |||
<td> 将输入编码为具有具 有表示能力的向量 </td> | |||
<td> embedding, RNN, CNN, transformer | |||
</tr> | |||
<tr> | |||
<td> aggregator </td> | |||
<td> aggregate and reduce information </td> | |||
<td> 从多个向量中聚合信息 </td> | |||
<td> self-attention, max-pooling </td> | |||
</tr> | |||
<tr> | |||
<td> decoder </td> | |||
<td> decode the representation into the output </td> | |||
<td> 将具有某种表示意义的 向量解码为需要的输出 形式 </td> | |||
<td> MLP, CRF </td> | |||
</tr> | |||
</table> | |||
For example: | |||
![](docs/source/figures/text_classification.png) | |||
## Requirements | |||
- Python>=3.6 | |||
- numpy>=1.14.2 | |||
- torch>=0.4.0 | |||
- tensorboardX | |||
- tqdm>=4.28.1 | |||
## 完整模型 | |||
fastNLP 为不同的 NLP 任务实现了许多完整的模型,它们都经过了训练和测试。 | |||
## Resources | |||
你可以在以下两个地方查看相关信息 | |||
- [介绍](reproduction/) | |||
- [源码](fastNLP/models/) | |||
- [Tutorials](https://github.com/fastnlp/fastNLP/tree/master/tutorials) | |||
- [Documentation](https://fastnlp.readthedocs.io/en/latest/) | |||
- [Source Code](https://github.com/fastnlp/fastNLP) | |||
## 项目结构 | |||
![](./docs/source/figures/workflow.png) | |||
## Installation | |||
Run the following commands to install fastNLP package. | |||
```shell | |||
pip install fastNLP | |||
``` | |||
## Models | |||
fastNLP implements different models for variant NLP tasks. | |||
Each model has been trained and tested carefully. | |||
Check out models' performance, usage and source code here. | |||
- [Documentation](reproduction/) | |||
- [Source Code](fastNLP/models/) | |||
## Project Structure | |||
fastNLP的大致工作流程如上图所示,而项目结构如下: | |||
<table> | |||
<tr> | |||
<td><b> fastNLP </b></td> | |||
<td> an open-source NLP library </td> | |||
</tr> | |||
<tr> | |||
<td><b> fastNLP.api </b></td> | |||
<td> APIs for end-to-end prediction </td> | |||
<td> 开源的自然语言处理库 </td> | |||
</tr> | |||
<tr> | |||
<td><b> fastNLP.core </b></td> | |||
<td> data representation & train/test procedure </td> | |||
<td> 实现了核心功能,包括数据处理组件、训练器、测速器等 </td> | |||
</tr> | |||
<tr> | |||
<td><b> fastNLP.models </b></td> | |||
<td> a collection of NLP models </td> | |||
<td> 实现了一些完整的神经网络模型 </td> | |||
</tr> | |||
<tr> | |||
<td><b> fastNLP.modules </b></td> | |||
<td> a collection of PyTorch sub-models/components/wheels </td> | |||
<td> 实现了用于搭建神经网络模型的诸多组件 </td> | |||
</tr> | |||
<tr> | |||
<td><b> fastNLP.io </b></td> | |||
<td> readers & savers </td> | |||
<td> 实现了读写功能,包括数据读入,模型读写等 </td> | |||
</tr> | |||
</table> | |||
## 参考资源 | |||
- [教程](https://github.com/fastnlp/fastNLP/tree/master/tutorials) | |||
- [文档](https://fastnlp.readthedocs.io/en/latest/) | |||
- [源码](https://github.com/fastnlp/fastNLP) | |||
*In memory of @FengZiYjun. May his soul rest in peace. We will miss you very very much!* |
@@ -1,12 +1,7 @@ | |||
# fastNLP 教程 | |||
### 上手教程 Quick Start | |||
- 一分钟上手:`fastnlp_1min_tutorial.ipynb` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/fastnlp_1min_tutorial.ipynb) | |||
- 十分钟上手:`fastnlp_10min_tutorial.ipynb` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/fastnlp_10min_tutorial.ipynb) | |||
`quickstart.ipynb` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/quickstart.ipynb) | |||
### 进阶教程 Advanced Tutorial | |||
- `fastnlp_advanced_tutorial/advance_tutorial.ipynb` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/fastnlp_advanced_tutorial/advance_tutorial.ipynb) | |||
### 开发者指南 Developer Guide | |||
- `tutorial_for_developer.md` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/tutorial_for_developer.md) | |||
### 详细教程 Tutorial 1 | |||
十分钟上手:`tutorial_1.ipynb` [Click Here](https://github.com/fastnlp/fastNLP/tree/master/tutorials/tutorial_1.ipynb) |
@@ -1,370 +0,0 @@ | |||
{ | |||
"cells": [ | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 1, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stderr", | |||
"output_type": "stream", | |||
"text": [ | |||
"/Users/yh/miniconda2/envs/python3/lib/python3.6/site-packages/tqdm/autonotebook/__init__.py:14: TqdmExperimentalWarning: Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)\n", | |||
" \" (e.g. in jupyter console)\", TqdmExperimentalWarning)\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"DataSet({'raw_sent': this is a bad idea . type=str,\n", | |||
"'label': 0 type=int,\n", | |||
"'word_str_lst': ['this', 'is', 'a', 'bad', 'idea', '.'] type=list,\n", | |||
"'words': [4, 2, 5, 6, 7, 3] type=list},\n", | |||
"{'raw_sent': it is great . type=str,\n", | |||
"'label': 1 type=int,\n", | |||
"'word_str_lst': ['it', 'is', 'great', '.'] type=list,\n", | |||
"'words': [8, 2, 9, 3] type=list})" | |||
] | |||
}, | |||
"execution_count": 1, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 假设有以下的DataSet, 这里只是为了举例所以只选择了两个sample\n", | |||
"import sys\n", | |||
"import os\n", | |||
"sys.path.append('/Users/yh/Desktop/fastNLP/fastNLP')\n", | |||
"\n", | |||
"from fastNLP import DataSet\n", | |||
"from fastNLP import Instance\n", | |||
"from fastNLP import Vocabulary\n", | |||
"\n", | |||
"dataset = DataSet()\n", | |||
"dataset.append(Instance(raw_sent='This is a bad idea .', label=0))\n", | |||
"dataset.append(Instance(raw_sent='It is great .', label=1))\n", | |||
"\n", | |||
"# 按照fastNLP_10min_tutorial.ipynb的步骤,对数据进行一些处理。这里为了演示padding操作,把field的名称做了一些改变\n", | |||
"dataset.apply(lambda x:x['raw_sent'].lower(), new_field_name='raw_sent')\n", | |||
"dataset.apply(lambda x:x['raw_sent'].split(), new_field_name='word_str_lst')\n", | |||
"\n", | |||
"# 建立Vocabulary\n", | |||
"word_vocab = Vocabulary()\n", | |||
"dataset.apply(lambda x:word_vocab.update(x['word_str_lst']))\n", | |||
"dataset.apply(lambda x:[word_vocab.to_index(word) for word in x['word_str_lst']], new_field_name='words')\n", | |||
"\n", | |||
"# 检查以下是否得到我们想要的结果了\n", | |||
"dataset[:2]" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 2, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'word_str_lst': array([list(['this', 'is', 'a', 'bad', 'idea', '.']),\n", | |||
" list(['it', 'is', 'great', '.'])], dtype=object), 'words': tensor([[4, 2, 5, 6, 7, 3],\n", | |||
" [8, 2, 9, 3, 0, 0]])}\n", | |||
"batch_y has: {'label': tensor([0, 1])}\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"'\"\\n结果中\\n Batch会对元素类型(元素即最内层的数据,raw_sent为str,word_str_lst为str,words为int, label为int)为int或者float的数据进行默认\\n padding,而非int或float的则不进行padding。但若每个Instance中该field为二维数据,也不进行padding。因为二维数据的padding涉及到\\n 两个维度的padding,不容易自动判断padding的形式。\\n'" | |||
] | |||
}, | |||
"execution_count": 2, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 将field设置为input或者target\n", | |||
"dataset.set_input('word_str_lst')\n", | |||
"dataset.set_input('words')\n", | |||
"dataset.set_target('label')\n", | |||
"\n", | |||
"# 使用Batch取出batch数据\n", | |||
"from fastNLP.core.batch import Batch\n", | |||
"from fastNLP.core.sampler import RandomSampler\n", | |||
"\n", | |||
"batch_iterator = Batch(dataset=dataset, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
"\"\"\"\"\n", | |||
"结果中\n", | |||
" Batch会对元素类型(元素即最内层的数据,raw_sent为str,word_str_lst为str,words为int, label为int)为int或者float的数据进行默认\n", | |||
" padding,而非int或float的则不进行padding。但若每个Instance中该field为二维数据,也不进行padding。因为二维数据的padding涉及到\n", | |||
" 两个维度的padding,不容易自动判断padding的形式。\n", | |||
"\"\"\"" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 3, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'word_str_lst': array([list(['it', 'is', 'great', '.']),\n", | |||
" list(['this', 'is', 'a', 'bad', 'idea', '.'])], dtype=object), 'words': tensor([[ 8, 2, 9, 3, -100, -100],\n", | |||
" [ 4, 2, 5, 6, 7, 3]])}\n", | |||
"batch_y has: {'label': tensor([1, 0])}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 所有的pad_val都默认为0,如果需要修改某一个field的默认pad值,可以通过DataSet.set_pad_val(field_name, pad_val)进行修改\n", | |||
"# 若需要将word的padding修改为-100\n", | |||
"dataset.set_pad_val('words', pad_val=-100)\n", | |||
"batch_iterator = Batch(dataset=dataset, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
"# pad的值修改为-100了" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 4, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"DataSet({'raw_sent': this is a bad idea . type=str,\n", | |||
"'label': 0 type=int,\n", | |||
"'word_str_lst': ['this', 'is', 'a', 'bad', 'idea', '.'] type=list,\n", | |||
"'words': [4, 2, 5, 6, 7, 3] type=list,\n", | |||
"'char_str_lst': [['t', 'h', 'i', 's'], ['i', 's'], ['a'], ['b', 'a', 'd'], ['i', 'd', 'e', 'a'], ['.']] type=list,\n", | |||
"'chars': [[4, 9, 2, 5], [2, 5], [3], [10, 3, 6], [2, 6, 7, 3], [8]] type=list},\n", | |||
"{'raw_sent': it is great . type=str,\n", | |||
"'label': 1 type=int,\n", | |||
"'word_str_lst': ['it', 'is', 'great', '.'] type=list,\n", | |||
"'words': [8, 2, 9, 3] type=list,\n", | |||
"'char_str_lst': [['i', 't'], ['i', 's'], ['g', 'r', 'e', 'a', 't'], ['.']] type=list,\n", | |||
"'chars': [[2, 4], [2, 5], [11, 12, 7, 3, 4], [8]] type=list})" | |||
] | |||
}, | |||
"execution_count": 4, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 若需要使用二维padding或指定padding方式,可以通过设置该field的padder实现,下面以英文的character padding为例。在某些场景下,可能想要\n", | |||
"# 使用英文word的character作为特征,character的padding为二维padding,fastNLP默认只会进行一维padding。\n", | |||
"\n", | |||
"dataset.apply(lambda x: [[c for c in word] for word in x['word_str_lst']], new_field_name='char_str_lst')\n", | |||
"char_vocab = Vocabulary()\n", | |||
"dataset.apply(lambda x:[char_vocab.update(chars) for chars in x['char_str_lst']])\n", | |||
"dataset.apply(lambda x:[[char_vocab.to_index(c) for c in chars] for chars in x['char_str_lst']],new_field_name='chars')\n", | |||
"dataset[:2]" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 5, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'word_str_lst': array([list(['this', 'is', 'a', 'bad', 'idea', '.']),\n", | |||
" list(['it', 'is', 'great', '.'])], dtype=object), 'words': tensor([[ 4, 2, 5, 6, 7, 3],\n", | |||
" [ 8, 2, 9, 3, -100, -100]]), 'chars': array([list([[4, 9, 2, 5], [2, 5], [3], [10, 3, 6], [2, 6, 7, 3], [8]]),\n", | |||
" list([[2, 4], [2, 5], [11, 12, 7, 3, 4], [8]])], dtype=object)}\n", | |||
"batch_y has: {'label': tensor([0, 1])}\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"'\\n 其它field与之前的是相同的。chars因为存在两个维度需要padding,不能自动决定padding方式,所以直接输出了原始形式。\\n'" | |||
] | |||
}, | |||
"execution_count": 5, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 如果不针对二维的character指定padding方法\n", | |||
"dataset.set_input('chars')\n", | |||
"batch_iterator = Batch(dataset=dataset, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
" \n", | |||
"\"\"\"\n", | |||
" 其它field与之前的是相同的。chars因为存在两个维度需要padding,不能自动决定padding方式,所以直接输出了原始形式。\n", | |||
"\"\"\"" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 6, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'word_str_lst': array([list(['this', 'is', 'a', 'bad', 'idea', '.']),\n", | |||
" list(['it', 'is', 'great', '.'])], dtype=object), 'words': tensor([[ 4, 2, 5, 6, 7, 3],\n", | |||
" [ 8, 2, 9, 3, -100, -100]]), 'chars': tensor([[[ 4, 9, 2, 5],\n", | |||
" [ 2, 5, 0, 0],\n", | |||
" [ 3, 0, 0, 0],\n", | |||
" [10, 3, 6, 0],\n", | |||
" [ 2, 6, 7, 3],\n", | |||
" [ 8, 0, 0, 0]],\n", | |||
"\n", | |||
" [[ 2, 4, 0, 0],\n", | |||
" [ 2, 5, 0, 0],\n", | |||
" [11, 12, 7, 3],\n", | |||
" [ 8, 0, 0, 0],\n", | |||
" [ 0, 0, 0, 0],\n", | |||
" [ 0, 0, 0, 0]]])}\n", | |||
"batch_y has: {'label': tensor([0, 1])}\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"'\\n chars被正确padding了\\n'" | |||
] | |||
}, | |||
"execution_count": 6, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 若要使用二维padding,需要手动设置padding方式\n", | |||
"from fastNLP.core.fieldarray import EngChar2DPadder\n", | |||
"dataset.set_padder('chars', EngChar2DPadder())\n", | |||
"batch_iterator = Batch(dataset=dataset, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
" \n", | |||
"\"\"\"\n", | |||
" chars被正确padding了\n", | |||
"\"\"\"" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 7, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'raw_sent': ['this is a bad idea .', 'it is great . '], 'word_str_lst': array([list(['this', 'is', 'a', 'bad', 'idea', '.']),\n", | |||
" list(['it', 'is', 'great', '.'])], dtype=object), 'words': tensor([[ 4, 2, 5, 6, 7, 3],\n", | |||
" [ 8, 2, 9, 3, -100, -100]]), 'chars': tensor([[[ 4, 9, 2, 5],\n", | |||
" [ 2, 5, 0, 0],\n", | |||
" [ 3, 0, 0, 0],\n", | |||
" [10, 3, 6, 0],\n", | |||
" [ 2, 6, 7, 3],\n", | |||
" [ 8, 0, 0, 0]],\n", | |||
"\n", | |||
" [[ 2, 4, 0, 0],\n", | |||
" [ 2, 5, 0, 0],\n", | |||
" [11, 12, 7, 3],\n", | |||
" [ 8, 0, 0, 0],\n", | |||
" [ 0, 0, 0, 0],\n", | |||
" [ 0, 0, 0, 0]]])}\n", | |||
"batch_y has: {'label': tensor([0, 1])}\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"'\\n raw_sent正确输出,对应内容也进行了pad。\\n'" | |||
] | |||
}, | |||
"execution_count": 7, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# 如果AutoPad与EngChar2DPadder不能满足需要,可以自己实现Padder对象。这里举一个例子,比如需要把raw_sentence pad到一样长\n", | |||
"from fastNLP.core.fieldarray import PadderBase\n", | |||
"\n", | |||
"class PadStr(PadderBase):\n", | |||
" def __init__(self, pad_val=' '):\n", | |||
" super().__init__(pad_val=pad_val) #让父类管理pad_val的值,这样可以通过DataSet.set_pad_val()修改到该值\n", | |||
" \n", | |||
" def __call__(self, contents, field_name, field_ele_dtype):\n", | |||
" \"\"\"\n", | |||
" 如果以上面的例子举例,在raw_sent这个field进行pad时,传入的\n", | |||
" contents:\n", | |||
" [\n", | |||
" 'This is a bad idea .',\n", | |||
" 'It is great .'\n", | |||
" ]\n", | |||
" field_name: 'raw_sent',当前field的名称,主要用于帮助debug。\n", | |||
" field_ele_dtype: np.str. 这个参数基本都用不上,是该field中内部元素的类型\n", | |||
" \"\"\"\n", | |||
" max_len = max([len(str_) for str_ in contents])\n", | |||
" pad_strs = []\n", | |||
" for content in contents:\n", | |||
" pad_strs.append(content + (max_len-len(content))*self.pad_val)\n", | |||
" return pad_strs\n", | |||
"\n", | |||
"dataset.set_input('raw_sent')\n", | |||
"dataset.set_padder('raw_sent', PadStr())\n", | |||
"batch_iterator = Batch(dataset=dataset, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
"\n", | |||
"\"\"\"\n", | |||
" raw_sent正确输出,对应内容也进行了pad。\n", | |||
"\"\"\"" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": null, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [] | |||
} | |||
], | |||
"metadata": { | |||
"kernelspec": { | |||
"display_name": "Python 3", | |||
"language": "python", | |||
"name": "python3" | |||
}, | |||
"language_info": { | |||
"codemirror_mode": { | |||
"name": "ipython", | |||
"version": 3 | |||
}, | |||
"file_extension": ".py", | |||
"mimetype": "text/x-python", | |||
"name": "python", | |||
"nbconvert_exporter": "python", | |||
"pygments_lexer": "ipython3", | |||
"version": "3.6.7" | |||
} | |||
}, | |||
"nbformat": 4, | |||
"nbformat_minor": 2 | |||
} |
@@ -1,751 +0,0 @@ | |||
{ | |||
"cells": [ | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"fastNLP10 分钟上手教程\n", | |||
"-------\n", | |||
"\n", | |||
"fastNLP提供方便的数据预处理,训练和测试模型的功能" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"如果您还没有通过pip安装fastNLP,可以执行下面的操作加载当前模块" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 4, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"import sys\n", | |||
"sys.path.append(\"../\")" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"DataSet & Instance\n", | |||
"------\n", | |||
"\n", | |||
"fastNLP用DataSet和Instance保存和处理数据。每个DataSet表示一个数据集,每个Instance表示一个数据样本。一个DataSet存有多个Instance,每个Instance可以自定义存哪些内容。\n", | |||
"\n", | |||
"有一些read_*方法,可以轻松从文件读取数据,存成DataSet。" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 1, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"77\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"from fastNLP import DataSet\n", | |||
"from fastNLP import Instance\n", | |||
"\n", | |||
"# 从csv读取数据到DataSet\n", | |||
"dataset = DataSet.read_csv('sample_data/tutorial_sample_dataset.csv', headers=('raw_sentence', 'label'), sep='\\t')\n", | |||
"print(len(dataset))" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 2, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': A series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . type=str,\n", | |||
"'label': 1 type=str}\n", | |||
"{'raw_sentence': The plot is romantic comedy boilerplate from start to finish . type=str,\n", | |||
"'label': 2 type=str}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 使用数字索引[k],获取第k个样本\n", | |||
"print(dataset[0])\n", | |||
"\n", | |||
"# 索引也可以是负数\n", | |||
"print(dataset[-3])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"## Instance\n", | |||
"Instance表示一个样本,由一个或多个field(域,属性,特征)组成,每个field有名字和值。\n", | |||
"\n", | |||
"在初始化Instance时即可定义它包含的域,使用 \"field_name=field_value\"的写法。" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 3, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"{'raw_sentence': fake data type=str,\n", | |||
"'label': 0 type=str}" | |||
] | |||
}, | |||
"execution_count": 3, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"# DataSet.append(Instance)加入新数据\n", | |||
"dataset.append(Instance(raw_sentence='fake data', label='0'))\n", | |||
"dataset[-1]" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"## DataSet.apply方法\n", | |||
"数据预处理利器" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 4, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': a series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . type=str,\n", | |||
"'label': 1 type=str}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 将所有数字转为小写\n", | |||
"dataset.apply(lambda x: x['raw_sentence'].lower(), new_field_name='raw_sentence')\n", | |||
"print(dataset[0])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 5, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': a series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . type=str,\n", | |||
"'label': 1 type=int}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# label转int\n", | |||
"dataset.apply(lambda x: int(x['label']), new_field_name='label')\n", | |||
"print(dataset[0])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 6, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': a series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . type=str,\n", | |||
"'label': 1 type=int,\n", | |||
"'words': ['a', 'series', 'of', 'escapades', 'demonstrating', 'the', 'adage', 'that', 'what', 'is', 'good', 'for', 'the', 'goose', 'is', 'also', 'good', 'for', 'the', 'gander', ',', 'some', 'of', 'which', 'occasionally', 'amuses', 'but', 'none', 'of', 'which', 'amounts', 'to', 'much', 'of', 'a', 'story', '.'] type=list}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 使用空格分割句子\n", | |||
"def split_sent(ins):\n", | |||
" return ins['raw_sentence'].split()\n", | |||
"dataset.apply(split_sent, new_field_name='words')\n", | |||
"print(dataset[0])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 7, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': a series of escapades demonstrating the adage that what is good for the goose is also good for the gander , some of which occasionally amuses but none of which amounts to much of a story . type=str,\n", | |||
"'label': 1 type=int,\n", | |||
"'words': ['a', 'series', 'of', 'escapades', 'demonstrating', 'the', 'adage', 'that', 'what', 'is', 'good', 'for', 'the', 'goose', 'is', 'also', 'good', 'for', 'the', 'gander', ',', 'some', 'of', 'which', 'occasionally', 'amuses', 'but', 'none', 'of', 'which', 'amounts', 'to', 'much', 'of', 'a', 'story', '.'] type=list,\n", | |||
"'seq_len': 37 type=int}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 增加长度信息\n", | |||
"dataset.apply(lambda x: len(x['words']), new_field_name='seq_len')\n", | |||
"print(dataset[0])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"## DataSet.drop\n", | |||
"筛选数据" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 8, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"77\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 删除低于某个长度的词语\n", | |||
"dataset.drop(lambda x: x['seq_len'] <= 3)\n", | |||
"print(len(dataset))" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"## 配置DataSet\n", | |||
"1. 哪些域是特征,哪些域是标签\n", | |||
"2. 切分训练集/验证集" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 9, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"# 设置DataSet中,哪些field要转为tensor\n", | |||
"\n", | |||
"# set target,loss或evaluate中的golden,计算loss,模型评估时使用\n", | |||
"dataset.set_target(\"label\")\n", | |||
"# set input,模型forward时使用\n", | |||
"dataset.set_input(\"words\")" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 10, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"54\n", | |||
"23\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 分出测试集、训练集\n", | |||
"\n", | |||
"test_data, train_data = dataset.split(0.3)\n", | |||
"print(len(test_data))\n", | |||
"print(len(train_data))" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"Vocabulary\n", | |||
"------\n", | |||
"\n", | |||
"fastNLP中的Vocabulary轻松构建词表,将词转成数字" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 11, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"{'raw_sentence': the performances are an absolute joy . type=str,\n", | |||
"'label': 4 type=int,\n", | |||
"'words': [3, 1, 1, 26, 1, 1, 2] type=list,\n", | |||
"'seq_len': 7 type=int}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"from fastNLP import Vocabulary\n", | |||
"\n", | |||
"# 构建词表, Vocabulary.add(word)\n", | |||
"vocab = Vocabulary(min_freq=2)\n", | |||
"train_data.apply(lambda x: [vocab.add(word) for word in x['words']])\n", | |||
"vocab.build_vocab()\n", | |||
"\n", | |||
"# index句子, Vocabulary.to_index(word)\n", | |||
"train_data.apply(lambda x: [vocab.to_index(word) for word in x['words']], new_field_name='words')\n", | |||
"test_data.apply(lambda x: [vocab.to_index(word) for word in x['words']], new_field_name='words')\n", | |||
"\n", | |||
"\n", | |||
"print(test_data[0])" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 12, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"batch_x has: {'words': tensor([[ 15, 72, 15, 73, 74, 7, 3, 75, 6, 3, 16, 16,\n", | |||
" 76, 2],\n", | |||
" [ 15, 72, 15, 73, 74, 7, 3, 75, 6, 3, 16, 16,\n", | |||
" 76, 2]])}\n", | |||
"batch_y has: {'label': tensor([ 1, 1])}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 如果你们需要做强化学习或者GAN之类的项目,你们也可以使用这些数据预处理的工具\n", | |||
"from fastNLP.core.batch import Batch\n", | |||
"from fastNLP.core.sampler import RandomSampler\n", | |||
"\n", | |||
"batch_iterator = Batch(dataset=train_data, batch_size=2, sampler=RandomSampler())\n", | |||
"for batch_x, batch_y in batch_iterator:\n", | |||
" print(\"batch_x has: \", batch_x)\n", | |||
" print(\"batch_y has: \", batch_y)\n", | |||
" break" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"# Model\n", | |||
"定义一个PyTorch模型" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 15, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"CNNText(\n", | |||
" (embed): Embedding(\n", | |||
" 77, 50\n", | |||
" (dropout): Dropout(p=0.0)\n", | |||
" )\n", | |||
" (conv_pool): ConvMaxpool(\n", | |||
" (convs): ModuleList(\n", | |||
" (0): Conv1d(50, 3, kernel_size=(3,), stride=(1,), padding=(2,))\n", | |||
" (1): Conv1d(50, 4, kernel_size=(4,), stride=(1,), padding=(2,))\n", | |||
" (2): Conv1d(50, 5, kernel_size=(5,), stride=(1,), padding=(2,))\n", | |||
" )\n", | |||
" )\n", | |||
" (dropout): Dropout(p=0.1)\n", | |||
" (fc): Linear(\n", | |||
" (linear): Linear(in_features=12, out_features=5, bias=True)\n", | |||
" )\n", | |||
")" | |||
] | |||
}, | |||
"execution_count": 15, | |||
"metadata": {}, | |||
"output_type": "execute_result" | |||
} | |||
], | |||
"source": [ | |||
"from fastNLP.models import CNNText\n", | |||
"model = CNNText((len(vocab), 50), num_classes=5, padding=2, dropout=0.1)\n", | |||
"model" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"这是上述模型的forward方法。如果你不知道什么是forward方法,请参考我们的PyTorch教程。\n", | |||
"\n", | |||
"注意两点:\n", | |||
"1. forward参数名字叫**word_seq**,请记住。\n", | |||
"2. forward的返回值是一个**dict**,其中有个key的名字叫**output**。\n", | |||
"\n", | |||
"```Python\n", | |||
" def forward(self, word_seq):\n", | |||
" \"\"\"\n", | |||
"\n", | |||
" :param word_seq: torch.LongTensor, [batch_size, seq_len]\n", | |||
" :return output: dict of torch.LongTensor, [batch_size, num_classes]\n", | |||
" \"\"\"\n", | |||
" x = self.embed(word_seq) # [N,L] -> [N,L,C]\n", | |||
" x = self.conv_pool(x) # [N,L,C] -> [N,C]\n", | |||
" x = self.dropout(x)\n", | |||
" x = self.fc(x) # [N,C] -> [N, N_class]\n", | |||
" return {'output': x}\n", | |||
"```" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"这是上述模型的predict方法,是用来直接输出该任务的预测结果,与forward目的不同。\n", | |||
"\n", | |||
"注意两点:\n", | |||
"1. predict参数名也叫**word_seq**。\n", | |||
"2. predict的返回值是也一个**dict**,其中有个key的名字叫**predict**。\n", | |||
"\n", | |||
"```\n", | |||
" def predict(self, word_seq):\n", | |||
" \"\"\"\n", | |||
"\n", | |||
" :param word_seq: torch.LongTensor, [batch_size, seq_len]\n", | |||
" :return predict: dict of torch.LongTensor, [batch_size, seq_len]\n", | |||
" \"\"\"\n", | |||
" output = self(word_seq)\n", | |||
" _, predict = output['output'].max(dim=1)\n", | |||
" return {'predict': predict}\n", | |||
"```" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"Trainer & Tester\n", | |||
"------\n", | |||
"\n", | |||
"使用fastNLP的Trainer训练模型" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 16, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"from fastNLP import Trainer\n", | |||
"from copy import deepcopy\n", | |||
"from fastNLP.core.losses import CrossEntropyLoss\n", | |||
"from fastNLP.core.metrics import AccuracyMetric\n", | |||
"\n", | |||
"\n", | |||
"# 更改DataSet中对应field的名称,与模型的forward的参数名一致\n", | |||
"# 因为forward的参数叫word_seq, 所以要把原本叫words的field改名为word_seq\n", | |||
"# 这里的演示是让你了解这种**命名规则**\n", | |||
"train_data.rename_field('words', 'word_seq')\n", | |||
"test_data.rename_field('words', 'word_seq')\n", | |||
"\n", | |||
"# 顺便把label换名为label_seq\n", | |||
"train_data.rename_field('label', 'label_seq')\n", | |||
"test_data.rename_field('label', 'label_seq')" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"### loss\n", | |||
"训练模型需要提供一个损失函数\n", | |||
"\n", | |||
"下面提供了一个在分类问题中常用的交叉熵损失。注意它的**初始化参数**。\n", | |||
"\n", | |||
"pred参数对应的是模型的forward返回的dict的一个key的名字,这里是\"output\"。\n", | |||
"\n", | |||
"target参数对应的是dataset作为标签的field的名字,这里是\"label_seq\"。" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 17, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"loss = CrossEntropyLoss(pred=\"output\", target=\"label_seq\")" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"### Metric\n", | |||
"定义评价指标\n", | |||
"\n", | |||
"这里使用准确率。参数的“命名规则”跟上面类似。\n", | |||
"\n", | |||
"pred参数对应的是模型的predict方法返回的dict的一个key的名字,这里是\"predict\"。\n", | |||
"\n", | |||
"target参数对应的是dataset作为标签的field的名字,这里是\"label_seq\"。" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 18, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"metric = AccuracyMetric(pred=\"predict\", target=\"label_seq\")" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 19, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"input fields after batch(if batch size is 2):\n", | |||
"\tword_seq: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 11]) \n", | |||
"target fields after batch(if batch size is 2):\n", | |||
"\tlabel_seq: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2]) \n", | |||
"\n" | |||
] | |||
}, | |||
{ | |||
"ename": "NameError", | |||
"evalue": "\nProblems occurred when calling CNNText.forward(self, words, seq_len=None)\n\tmissing param: ['words']\n\tunused field: ['word_seq']\n\tSuggestion: You need to provide ['words'] in DataSet and set it as input. ", | |||
"output_type": "error", | |||
"traceback": [ | |||
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m", | |||
"\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)", | |||
"\u001b[0;32m<ipython-input-19-ff7d68caf88a>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[1;32m 7\u001b[0m \u001b[0msave_path\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mNone\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 8\u001b[0m \u001b[0mbatch_size\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m32\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m----> 9\u001b[0;31m n_epochs=5)\n\u001b[0m\u001b[1;32m 10\u001b[0m \u001b[0moverfit_trainer\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtrain\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n", | |||
"\u001b[0;32m/Users/fdujyn/anaconda3/lib/python3.6/site-packages/fastNLP/core/trainer.py\u001b[0m in \u001b[0;36m__init__\u001b[0;34m(self, train_data, model, optimizer, loss, batch_size, sampler, update_every, n_epochs, print_every, dev_data, metrics, metric_key, validate_every, save_path, prefetch, use_tqdm, device, callbacks, check_code_level)\u001b[0m\n\u001b[1;32m 447\u001b[0m _check_code(dataset=train_data, model=model, losser=losser, metrics=metrics, dev_data=dev_data,\n\u001b[1;32m 448\u001b[0m \u001b[0mmetric_key\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mmetric_key\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mcheck_level\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0mcheck_code_level\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 449\u001b[0;31m batch_size=min(batch_size, DEFAULT_CHECK_BATCH_SIZE))\n\u001b[0m\u001b[1;32m 450\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 451\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mtrain_data\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mtrain_data\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n", | |||
"\u001b[0;32m/Users/fdujyn/anaconda3/lib/python3.6/site-packages/fastNLP/core/trainer.py\u001b[0m in \u001b[0;36m_check_code\u001b[0;34m(dataset, model, losser, metrics, batch_size, dev_data, metric_key, check_level)\u001b[0m\n\u001b[1;32m 808\u001b[0m \u001b[0mprint\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minfo_str\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 809\u001b[0m _check_forward_error(forward_func=model.forward, dataset=dataset,\n\u001b[0;32m--> 810\u001b[0;31m batch_x=batch_x, check_level=check_level)\n\u001b[0m\u001b[1;32m 811\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 812\u001b[0m \u001b[0mrefined_batch_x\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0m_build_args\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmodel\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mforward\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mbatch_x\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n", | |||
"\u001b[0;32m/Users/fdujyn/anaconda3/lib/python3.6/site-packages/fastNLP/core/utils.py\u001b[0m in \u001b[0;36m_check_forward_error\u001b[0;34m(forward_func, batch_x, dataset, check_level)\u001b[0m\n\u001b[1;32m 594\u001b[0m \u001b[0msugg_str\u001b[0m \u001b[0;34m+=\u001b[0m \u001b[0msuggestions\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 595\u001b[0m \u001b[0merr_str\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;34m'\\n'\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0;34m'\\n'\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mjoin\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0merrs\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0;34m'\\n\\tSuggestion: '\u001b[0m \u001b[0;34m+\u001b[0m \u001b[0msugg_str\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 596\u001b[0;31m \u001b[0;32mraise\u001b[0m \u001b[0mNameError\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0merr_str\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m 597\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0m_unused\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m 598\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mcheck_level\u001b[0m \u001b[0;34m==\u001b[0m \u001b[0mWARNING_CHECK_LEVEL\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n", | |||
"\u001b[0;31mNameError\u001b[0m: \nProblems occurred when calling CNNText.forward(self, words, seq_len=None)\n\tmissing param: ['words']\n\tunused field: ['word_seq']\n\tSuggestion: You need to provide ['words'] in DataSet and set it as input. " | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 实例化Trainer,传入模型和数据,进行训练\n", | |||
"# 先在test_data拟合(确保模型的实现是正确的)\n", | |||
"copy_model = deepcopy(model)\n", | |||
"overfit_trainer = Trainer(model=copy_model, train_data=test_data, dev_data=test_data,\n", | |||
" loss=loss,\n", | |||
" metrics=metric,\n", | |||
" save_path=None,\n", | |||
" batch_size=32,\n", | |||
" n_epochs=5)\n", | |||
"overfit_trainer.train()" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 22, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"input fields after batch(if batch size is 2):\n", | |||
"\tword_seq: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2, 20]) \n", | |||
"target fields after batch(if batch size is 2):\n", | |||
"\tlabel_seq: (1)type:torch.Tensor (2)dtype:torch.int64, (3)shape:torch.Size([2]) \n", | |||
"\n", | |||
"training epochs started 2019-01-12 17-09-05\n" | |||
] | |||
}, | |||
{ | |||
"data": { | |||
"text/plain": [ | |||
"HBox(children=(IntProgress(value=0, layout=Layout(flex='2'), max=5), HTML(value='')), layout=Layout(display='i…" | |||
] | |||
}, | |||
"metadata": {}, | |||
"output_type": "display_data" | |||
}, | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"Evaluation at Epoch 1/5. Step:1/5. AccuracyMetric: acc=0.37037\n", | |||
"Evaluation at Epoch 2/5. Step:2/5. AccuracyMetric: acc=0.37037\n", | |||
"Evaluation at Epoch 3/5. Step:3/5. AccuracyMetric: acc=0.462963\n", | |||
"Evaluation at Epoch 4/5. Step:4/5. AccuracyMetric: acc=0.425926\n", | |||
"Evaluation at Epoch 5/5. Step:5/5. AccuracyMetric: acc=0.481481\n", | |||
"\n", | |||
"In Epoch:5/Step:5, got best dev performance:AccuracyMetric: acc=0.481481\n", | |||
"Reloaded the best model.\n", | |||
"Train finished!\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 用train_data训练,在test_data验证\n", | |||
"trainer = Trainer(model=model, train_data=train_data, dev_data=test_data,\n", | |||
" loss=CrossEntropyLoss(pred=\"output\", target=\"label_seq\"),\n", | |||
" metrics=AccuracyMetric(pred=\"predict\", target=\"label_seq\"),\n", | |||
" save_path=None,\n", | |||
" batch_size=32,\n", | |||
" n_epochs=5)\n", | |||
"trainer.train()\n", | |||
"print('Train finished!')" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": 23, | |||
"metadata": {}, | |||
"outputs": [ | |||
{ | |||
"name": "stdout", | |||
"output_type": "stream", | |||
"text": [ | |||
"[tester] \n", | |||
"AccuracyMetric: acc=0.481481\n", | |||
"{'AccuracyMetric': {'acc': 0.481481}}\n" | |||
] | |||
} | |||
], | |||
"source": [ | |||
"# 调用Tester在test_data上评价效果\n", | |||
"from fastNLP import Tester\n", | |||
"\n", | |||
"tester = Tester(data=test_data, model=model, metrics=AccuracyMetric(pred=\"predict\", target=\"label_seq\"),\n", | |||
" batch_size=4)\n", | |||
"acc = tester.test()\n", | |||
"print(acc)" | |||
] | |||
}, | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"# In summary\n", | |||
"\n", | |||
"## fastNLP Trainer的伪代码逻辑\n", | |||
"### 1. 准备DataSet,假设DataSet中共有如下的fields\n", | |||
" ['raw_sentence', 'word_seq1', 'word_seq2', 'raw_label','label']\n", | |||
" 通过\n", | |||
" DataSet.set_input('word_seq1', word_seq2', flag=True)将'word_seq1', 'word_seq2'设置为input\n", | |||
" 通过\n", | |||
" DataSet.set_target('label', flag=True)将'label'设置为target\n", | |||
"### 2. 初始化模型\n", | |||
" class Model(nn.Module):\n", | |||
" def __init__(self):\n", | |||
" xxx\n", | |||
" def forward(self, word_seq1, word_seq2):\n", | |||
" # (1) 这里使用的形参名必须和DataSet中的input field的名称对应。因为我们是通过形参名, 进行赋值的\n", | |||
" # (2) input field的数量可以多于这里的形参数量。但是不能少于。\n", | |||
" xxxx\n", | |||
" # 输出必须是一个dict\n", | |||
"### 3. Trainer的训练过程\n", | |||
" (1) 从DataSet中按照batch_size取出一个batch,调用Model.forward\n", | |||
" (2) 将 Model.forward的结果 与 标记为target的field 传入Losser当中。\n", | |||
" 由于每个人写的Model.forward的output的dict可能key并不一样,比如有人是{'pred':xxx}, {'output': xxx}; \n", | |||
" 另外每个人将target可能也会设置为不同的名称, 比如有人是label, 有人设置为target;\n", | |||
" 为了解决以上的问题,我们的loss提供映射机制\n", | |||
" 比如CrossEntropyLosser的需要的输入是(prediction, target)。但是forward的output是{'output': xxx}; 'label'是target\n", | |||
" 那么初始化losser的时候写为CrossEntropyLosser(prediction='output', target='label')即可\n", | |||
" (3) 对于Metric是同理的\n", | |||
" Metric计算也是从 forward的结果中取值 与 设置target的field中取值。 也是可以通过映射找到对应的值 \n", | |||
" \n", | |||
" \n", | |||
"\n", | |||
"## 一些问题.\n", | |||
"### 1. DataSet中为什么需要设置input和target\n", | |||
" 只有被设置为input或者target的数据才会在train的过程中被取出来\n", | |||
" (1.1) 我们只会在设置为input的field中寻找传递给Model.forward的参数。\n", | |||
" (1.2) 我们在传递值给losser或者metric的时候会使用来自: \n", | |||
" (a)Model.forward的output\n", | |||
" (b)被设置为target的field\n", | |||
" \n", | |||
"\n", | |||
"### 2. 我们是通过forwad中的形参名将DataSet中的field赋值给对应的参数\n", | |||
" (1.1) 构建模型过程中,\n", | |||
" 例如:\n", | |||
" DataSet中x,seq_lens是input,那么forward就应该是\n", | |||
" def forward(self, x, seq_lens):\n", | |||
" pass\n", | |||
" 我们是通过形参名称进行匹配的field的\n", | |||
" \n", | |||
"\n", | |||
"\n", | |||
"### 1. 加载数据到DataSet\n", | |||
"### 2. 使用apply操作对DataSet进行预处理\n", | |||
" (2.1) 处理过程中将某些field设置为input,某些field设置为target\n", | |||
"### 3. 构建模型\n", | |||
" (3.1) 构建模型过程中,需要注意forward函数的形参名需要和DataSet中设置为input的field名称是一致的。\n", | |||
" 例如:\n", | |||
" DataSet中x,seq_lens是input,那么forward就应该是\n", | |||
" def forward(self, x, seq_lens):\n", | |||
" pass\n", | |||
" 我们是通过形参名称进行匹配的field的\n", | |||
" (3.2) 模型的forward的output需要是dict类型的。\n", | |||
" 建议将输出设置为{\"pred\": xx}.\n", | |||
" \n" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": null, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [] | |||
} | |||
], | |||
"metadata": { | |||
"kernelspec": { | |||
"display_name": "Python 3", | |||
"language": "python", | |||
"name": "python3" | |||
}, | |||
"language_info": { | |||
"codemirror_mode": { | |||
"name": "ipython", | |||
"version": 3 | |||
}, | |||
"file_extension": ".py", | |||
"mimetype": "text/x-python", | |||
"name": "python", | |||
"nbconvert_exporter": "python", | |||
"pygments_lexer": "ipython3", | |||
"version": "3.6.7" | |||
} | |||
}, | |||
"nbformat": 4, | |||
"nbformat_minor": 2 | |||
} |
@@ -1,97 +0,0 @@ | |||
{ | |||
"cells": [ | |||
{ | |||
"cell_type": "markdown", | |||
"metadata": {}, | |||
"source": [ | |||
"## fastNLP测试说明\n", | |||
"### 测试环境\n", | |||
"fastNLP使用pytest对代码进行单元测试,测试代码在test文件夹下,测试所需数据在test/data_for_tests文件夹下\n", | |||
"测试的步骤主要分为准备数据,执行测试,比对结果,清除环境四步\n", | |||
"测试代码以test_xxx.py命名,以DataSet的测试代码为例,测试代码文件名为test_dataset.py" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": null, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"import os\n", | |||
"import unittest # 单元测试需要用到unittest\n", | |||
"\n", | |||
"from fastNLP.core.dataset import DataSet\n", | |||
"from fastNLP.core.fieldarray import FieldArray\n", | |||
"from fastNLP.core.instance import Instance\n", | |||
"# 在这个单元测试文件中,需要测试DataSet、FieldArray、以及Instance" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": null, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [ | |||
"class TestDataSet(unittest.TestCase): # 类名字以Test打头,继承unittest.TestCase\n", | |||
"\n", | |||
" def test_init_v1(self): # 测试样例1, 函数名称以test_打头\n", | |||
" # 该测试样例测试的是DataSet的初始化\n", | |||
" ins = Instance(x=[1, 2, 3, 4], y=[5, 6]) # 准备数据\n", | |||
" ds = DataSet([ins] * 40) # 执行测试(调用DataSet的初始化函数)\n", | |||
" self.assertTrue(\"x\" in ds.field_arrays and \"y\" in ds.field_arrays) # 比对结果:'x'跟'y'都是ds的field\n", | |||
" self.assertEqual(ds.field_arrays[\"x\"].content, [[1, 2, 3, 4], ] * 40) # 比对结果: field 'x'的内容正确\n", | |||
" self.assertEqual(ds.field_arrays[\"y\"].content, [[5, 6], ] * 40) # 比对结果: field 'y'的内容正确\n", | |||
" \n", | |||
" def test_init_v2(self): # 测试样例2,该样例测试DataSet的另一种初始化方式\n", | |||
" ds = DataSet({\"x\": [[1, 2, 3, 4]] * 40, \"y\": [[5, 6]] * 40})\n", | |||
" self.assertTrue(\"x\" in ds.field_arrays and \"y\" in ds.field_arrays)\n", | |||
" self.assertEqual(ds.field_arrays[\"x\"].content, [[1, 2, 3, 4], ] * 40)\n", | |||
" self.assertEqual(ds.field_arrays[\"y\"].content, [[5, 6], ] * 40)\n", | |||
" \n", | |||
" def test_init_assert(self): # 测试样例3,该样例测试不规范初始化DataSet时是否会报正确错误\n", | |||
" with self.assertRaises(AssertionError):\n", | |||
" _ = DataSet({\"x\": [[1, 2, 3, 4]] * 40, \"y\": [[5, 6]] * 100})\n", | |||
" with self.assertRaises(AssertionError):\n", | |||
" _ = DataSet([[1, 2, 3, 4]] * 10)\n", | |||
" with self.assertRaises(ValueError):\n", | |||
" _ = DataSet(0.00001)\n", | |||
" \n", | |||
" def test_contains(self): # 测试样例4,该样例测试DataSet的contains函数,是功能测试\n", | |||
" ds = DataSet({\"x\": [[1, 2, 3, 4]] * 40, \"y\": [[5, 6]] * 40})\n", | |||
" self.assertTrue(\"x\" in ds)\n", | |||
" self.assertTrue(\"y\" in ds)\n", | |||
" self.assertFalse(\"z\" in ds)\n", | |||
" \n", | |||
" # 更多测试样例见test/core/test_dataset.py" | |||
] | |||
}, | |||
{ | |||
"cell_type": "code", | |||
"execution_count": null, | |||
"metadata": {}, | |||
"outputs": [], | |||
"source": [] | |||
} | |||
], | |||
"metadata": { | |||
"kernelspec": { | |||
"display_name": "Python 3", | |||
"language": "python", | |||
"name": "python3" | |||
}, | |||
"language_info": { | |||
"codemirror_mode": { | |||
"name": "ipython", | |||
"version": 3 | |||
}, | |||
"file_extension": ".py", | |||
"mimetype": "text/x-python", | |||
"name": "python", | |||
"nbconvert_exporter": "python", | |||
"pygments_lexer": "ipython3", | |||
"version": "3.6.7" | |||
} | |||
}, | |||
"nbformat": 4, | |||
"nbformat_minor": 2 | |||
} |
@@ -1,283 +0,0 @@ | |||
# fastNLP开发者指南 | |||
#### 本教程涉及以下类: | |||
- DataSet | |||
- Sampler | |||
- Batch | |||
- Model | |||
- Loss | |||
- Metric | |||
- Trainer | |||
- Tester | |||
#### DataSet: 用于承载数据。 | |||
1. DataSet里面每个元素只能是以下的三类`np.float64`, `np.int64`, `np.str`。如果传入的数据是`int`则被转换为`np.int64`, `float`被转为`np.float64`。 | |||
2. DataSet可以将field设置为input或者target。其中被设置为input的field会被传递给Model.forward, 这个过程中我们是通过键匹配完成传递的。举例来说,假设DataSet中有'x1', 'x2', 'x3'被设置为了input,而 | |||
- 函数是Model.forward(self, x1, x3), 那么DataSet中'x1', 'x3'会被传递给forward函数。多余的'x2'会被忽略 | |||
- 函数是Model.forward(self, x1, x4), 这里多需要了一个'x4', 但是DataSet的input field中没有这个field,会报错。 | |||
- 函数是Model.forward(self, x1, **kwargs), 会把'x1', 'x2', 'x3'都传入。但如果是Model.forward(self, x4, **kwargs)就会发生报错,因为没有'x4'。 | |||
3. 对于设置为target的field的名称,我们建议取名为'target'(如果只有一个需要predict的值),但是不强制。后面会讲为什么target可以不强制。 | |||
DataSet应该是不需要单独再开发的,如果有不能满足的场景,请在开发群提出或者github提交issue。 | |||
#### Sampler: 给定一个DataSet,返回一个序号的list,Batch按照这个list输出数据。 | |||
Sampler需要继承fastNLP.core.sampler.BaseSampler | |||
```python | |||
class BaseSampler(object): | |||
"""The base class of all samplers. | |||
Sub-classes must implement the __call__ method. | |||
__call__ takes a DataSet object and returns a list of int - the sampling indices. | |||
""" | |||
def __call__(self, *args, **kwargs): | |||
raise NotImplementedError | |||
# 子类需要复写__call__方法。这个函数只能有一个必选参数, 且必须是DataSet类别, 否则Trainer没法调 | |||
class SonSampler(BaseSampler): | |||
def __init__(self, xxx): | |||
# 可以实现init也不可以不实现。 | |||
pass | |||
def __call__(self, data_set): | |||
pass | |||
``` | |||
#### Batch: 将DataSet中设置为input和target的field取出来构成batch_x, batch_y | |||
并且根据情况(主要根据数据类型能不能转为Tensor)将数据转换为pytorch的Tensor。batch中sample的取出顺序是由Sampler决定的。 | |||
Sampler是传入一个DataSet,返回一个与DataSet等长的序号list,Batch一次会取出batch_size个sample(最后一个batch可能数量不足batch_size个)。 | |||
举例: | |||
1. SequentialSampler是顺序采样 | |||
假设传入的DataSet长度是100, SequentialSampler返回的序号list就是[0, 1, ...,98, 99]. batch_size如果被设置为4,那么第一个batch所获取的instance就是[0, 1, 2, 3]这四个instance. 第二个batch所获取instace就是[4, 5, 6, 7], ...直到采完所有的sample。 | |||
2. RandomSampler是随机采样 | |||
假设传入的DataSet长度是100, RandomSampler返回的序号list可能是[0, 99, 20, 5, 3, 1, ...]. 依次按照batch_size的大小取出sample。 | |||
Batch应该不需要继承与开发,如果你有特殊需求请在开发群里提出。 | |||
#### Model:用户自定的Model | |||
必须是nn.Module的子类 | |||
1. 必须实现forward方法,并且forward方法不能出现*arg这种参数. 例如 | |||
```python | |||
def forward(self, word_seq, *args): #这是不允许的. | |||
# ... | |||
pass | |||
``` | |||
返回值必须是dict的 | |||
```python | |||
def forward(self, word_seq, seq_lens): | |||
xxx = "xxx" | |||
return {'pred': xxx} #return的值必须是dict的。里面的预测的key推荐使用pred,但是不做强制限制。输出元素数目不限。 | |||
``` | |||
2. 如果实现了predict方法,在做evaluation的时候将调用predict方法而不是forward。如果没有predict方法,则在evaluation时调用forward方法。predict方法也不能使用*args这种参数形式,同时结果也必须返回一个dict,同样推荐key为'pred'。 | |||
#### Loss: 根据model.forward()返回的prediction(是一个dict)和batch_y计算相应的loss | |||
1. 先介绍"键映射"。 如在DataSet, Model一节所看见的那样,fastNLP并不限制Model.forward()的返回值,也不限制DataSet中target field的key。计算的loss的时候,怎么才能知道从哪里取值呢? | |||
这里以CrossEntropyLoss为例,一般情况下, 计算CrossEntropy需要prediction和target两个值。而在CrossEntropyLoss初始化时可以传入两个参数(pred=None, target=None), 这两个参数接受的类型是str,假设(pred='output', target='label'),那么CrossEntropyLoss会使用'output'这个key在forward的output与batch_y中寻找值;'label'也是在forward的output与batch_y中寻找值。注意这里pred或target的来源并不一定非要来自于model.forward与batch_y,也可以只来自于forward的结果。 | |||
2. 如何创建一个自己的loss | |||
- 使用fastNLP.LossInForward, 在model.forward()的结果中包含一个为loss的key。 | |||
- trainer中使用loss(假设loss=CrossEntropyLoss())的时候其实是 | |||
los = loss(prediction, batch_y),即直接调用的是`loss.__call__()`方法,但是CrossEntropyLoss里面并没有自己实现`__call__`方法,这是因为`__call__`在LossBase中实现了。所有的loss必须继承fastNLP.core.loss.LossBase, 下面先说一下LossBase的几个方法,见下一节。 | |||
3. 尽量不要复写`__call__()`, `_init_param_map()`方法。 | |||
```python | |||
class LossBase(): | |||
def __init__(self): | |||
self.param_map = {} # 一般情况下也不需要自己创建。调用_init_param_map()更好 | |||
self._checked = False # 这个参数可以忽略 | |||
def _init_param_map(self, key_map=None, **kwargs): | |||
# 这个函数是用于注册Loss的“键映射”,有两种传值方法, | |||
# 第一种是通过key_map传入dict,取值是用value到forward和batch_y取 | |||
# key_map = {'pred': 'output', 'target': 'label'} | |||
# 第二种是自己写 | |||
# _init_param_map(pred='output', target='label') | |||
# 为什么会提供这么一个方法?通过调用这个方法会自动注册param_map,并会做一些检查,防止出现传入的key其实并不是get_loss | |||
# 的一个参数。注意传入这个方法的参数必须都是需要做键映射的内容,其它loss参数不要传入。如果传入(pred=None, target=None) | |||
# 则__call__()会到pred_dict与target_dict去寻找key为'pred'和'target'的值。 | |||
# 但这个参数不是必须要调用的。 | |||
def __call__(self, pred_dict, target_dict, check=False): # check=False忽略这个参数,之后应该会被删除的 | |||
# 这个函数主要会做一些check的工作,比如pred_dict与target_dict中是否包含了计算loss所必须的key等。检查通过,则调用get_loss | |||
# 方法。 | |||
fast_param = self._fast_param_map(predict_dict, target_dict): | |||
if fast_param: | |||
return self.get_loss(**fast_param) | |||
# 如果没有fast_param则通过匹配参数然后调用get_loss完成 | |||
xxxx | |||
return loss # 返回为Tensor的loss | |||
def _fast_param_map(self, pred_dict, target_dict): | |||
# 这是一种快速计算loss的机制,因为在很多情况下其实都不需要通过"键映射",比如计算loss时,pred_dict只有一个元素, | |||
# target_dict也只有一个元素,那么无歧义地就可以把预测值与实际值用于计算loss, 基类判断了这种情况(可能还有其它无歧义的情况)。 | |||
# 即_fast_param_map成功的话,就不需要使用键映射,这样即使在没有传递或者传递错误"键映射"的情况也可以直接计算loss。 | |||
# 返回值是一个dict, 如果匹配成功,应该返回类似{'pred':value, 'target': value}的结果;如果dict为空则说明匹配失败, | |||
# __call__方法会继续执行。 | |||
def get_loss(self, *args, **kwargs): | |||
# 这个是一定需要实现的,计算loss的地方。 | |||
# (1) get_loss中一定不能包含*arg这种参数形式。 | |||
# (2) 如果包含**kwargs这种参数,这会将pred_dict与target_dict中所有参数传入。但是建议不要用这个参数 | |||
raise NotImplementedError | |||
# 下面使用L1Loss举例 | |||
class L1Loss(LossBase): # 继承LossBase | |||
# 初始化需要映射的值,这里需要映射的值'pred', 'target'必须与get_loss需要参数名是对应的 | |||
def __init__(self, pred=None, target=None): | |||
super(L1Loss, self).__init__() | |||
# 这里传入_init_param_map以使得pred和target被正确注册,但这一步不是必须的, 建议调用。传入_init_param_map的是用于 | |||
# “键映射"的键值对。假设初始化__init__(pred=None, target=None, threshold=0.1)中threshold是用于控制loss计算的,则 | |||
# 不要将threshold传入_init_param_map. | |||
self._init_param_map(pred=pred, target=target) | |||
def get_loss(self, pred, target): | |||
# 这里'pred', 'target'必须和初始化的映射是一致的。 | |||
return F.l1_loss(input=pred, target=target) #直接返回一个loss即可 | |||
``` | |||
### Metric: 根据Model.forward()或者Model.predict()的结果计算metric | |||
metric的设计和loss的设计类似。都是传入pred_dict与target_dict进行计算。但是metric的pred_dict来源可能是Model.forward的返回值, 也可能是Model.predict(如果Model具有predict方法则会调用predict方法)的返回值,下面统一用pred_dict代替。 | |||
1. 这里的"键映射"与loss的"键映射"是类似的。举例来说,若Metric(pred='output', target='label'),则使用'output'到pred_dict和target_dict中寻找pred, 用'label'寻找target。 | |||
2. 如何创建一个自己的Metric方法 | |||
Metric与loss的计算不同在于,Metric的计算有两个步骤。 | |||
- **每个batch的输出**都会调用Metric的``__call__(pred_dict, target_dict)``方法,而``__call__``方法会调用evaluate()(需要实现)方法。 | |||
- 在所有batch传入之后,调用Metric的get_metric()方法得到最终的metric值。 | |||
- 所以Metric在调用evaluate方法时,根据拿到的数据: pred_dict与batch_y, 改变自己的状态(比如累加正确的次数,总的sample数等)。在调用get_metric()的时候给出一个最终计算结果。 | |||
所有的Metric必须继承自fastNLP.core.metrics.MetricBase. 例子见下一个cell | |||
3. 尽量不要复写``__call__()``,``_init_param_map()``方法。 | |||
```python | |||
class MetricBase: | |||
def __init__(self): | |||
self.param_map = {} # 一般情况下也不需要自己创建。调用_init_param_map()更好 | |||
self._checked = False # 这个参数可以忽略 | |||
def _init_param_map(self, key_map=None, **kwargs): | |||
# 这个函数是用于注册Metric的“键映射”,有两种传值方法, | |||
# 第一种是通过key_map传入dict,取值是用value到forward和batch_y取 | |||
# key_map = {'pred': 'output', 'target': 'label'} | |||
# 第二种是自己写(建议使用改种方式) | |||
# _init_param_map(pred='output', target='label') | |||
# 为什么会提供这么一个方法?通过调用这个方法会自动注册param_map,并会做一些检查,防止出现传入的key其实并不是evaluate() | |||
# 的一个参数。注意传入这个方法的参数必须都是需要做键映射的内容,其它evaluate参数不要传入。如果传入(pred=None, target=None) | |||
# 则__call__()会到pred_dict与target_dict去寻找key为'pred'和'target'的值。 | |||
# 但这个参数不是必须要调用的。 | |||
pass | |||
def __call__(self, pred_dict, target_dict, check=False): # check=False忽略这个参数,之后应该会被删除的 | |||
# 这个函数主要会做一些check的工作,比如pred_dict与target_dict中是否包含了计算evaluate所必须的key等。检查通过,则调用 | |||
# evaluate方法。 | |||
fast_param = self._fast_param_map(predict_dict, target_dict): | |||
if fast_param: | |||
return self.evaluate(**fast_param) | |||
# 如果没有fast_param则通过匹配参数然后调用get_loss完成 | |||
# xxxx | |||
def _fast_param_map(self, pred_dict, target_dict): | |||
# 这是一种快速计算loss的机制,因为在很多情况下其实都不需要通过"键映射",比如evaluate时,pred_dict只有一个元素, | |||
# target_dict也只有一个元素,那么无歧义地就可以把预测值与实际值用于计算metric, 基类判断了这种情况(可能还有其它无歧义的 | |||
# 情况)。即_fast_param_map成功的话,就不需要使用键映射,这样即使在没有传递或者传递错误"键映射"的情况也可以直接计算metric。 | |||
# 返回值是一个dict, 如果匹配成功,应该返回类似{'pred':value, 'target': value}的结果;如果dict为空则说明匹配失败, | |||
# __call__方法会继续尝试匹配。 | |||
pass | |||
def evaluate(self, *args, **kwargs): | |||
# 这个是一定需要实现的,累加metric状态 | |||
# (1) evaluate()中一定不能包含*arg这种参数形式。 | |||
# (2) 如果包含**kwargs这种参数,这会将pred_dict与target_dict中所有参数传入。但是建议不要用这个参数 | |||
raise NotImplementedError | |||
def get_metric(self, reset=True): | |||
# 这是一定需要实现的,获取最终的metric。返回值必须是一个dict。会在所有batch传入之后调用 | |||
raise NotImplementedError | |||
# 下面使用AccuracyMetric举例 | |||
class AccuracyMetric(MetricBase): # MetricBase | |||
# 初始化需要映射的值,这里需要映射的值'pred', 'target'必须与evaluate()需要参数名是对应的 | |||
def __init__(self, pred=None, target=None): | |||
super(AccuracyMetric, self).__init__() | |||
# 这里传入_init_param_map以使得pred和target被正确注册,但这一步不是必须的, 建议调用。传入_init_param_map的是用于 | |||
# “键映射"的键值对。假设初始化__init__(pred=None, target=None, threshold=0.1)中threshold是用于控制loss计算的,则 | |||
# 不要将threshold传入_init_param_map. | |||
self._init_param_map(pred=pred, target=target) | |||
self.total = 0 # 用于累加一共有多少sample | |||
self.corr = 0 # 用于累加一共有多少正确的sample | |||
def evaluate(self, pred, target): | |||
# 对pred和target做一些基本的判断或者预处理等 | |||
if pred.size()==target.size() and len(pred.size())=1: #如果pred已经做了argmax | |||
pass | |||
elif len(pred.size())==2 and len(target.size())==1: # pred还没有进行argmax | |||
pred = pred.argmax(dim=1) | |||
else: | |||
raise ValueError("The shape of pred and target should be ((B, n_classes), (B, )) or (" | |||
"(B,),(B,)).") | |||
assert pred.size(0)==target.size(0), "Mismatch batch size." | |||
# 进行相应的累加 | |||
self.total += pred.size(0) | |||
self.corr += torch.sum(torch.eq(pred, target).float()).item() | |||
def get_metric(self, reset=True): | |||
# reset用于指示是否清空累加信息。默认为True | |||
# 这个函数需要返回dict,可以包含多个metric。 | |||
metric = {} | |||
metric['acc'] = self.corr/self.total | |||
if reset: | |||
self.total = 0 | |||
self.corr = 0 | |||
return metric | |||
``` | |||
#### Tester: 用于做evaluation,应该不需要更改 | |||
重要的初始化参数有data, model, metric;比较重要的function是test()。 | |||
test中的运行过程 | |||
``` | |||
predict_func = 如果有model.predict则为model.predict, 否则是model.forward | |||
for batch_x, batch_y in batch: | |||
# (1) 同步数据与model | |||
# (2) 根据predict_func的参数从batch_x中取出数据传入到predict_func中,得到结果pred_dict | |||
# (3) 调用metric(pred_dict, batch_y | |||
# (4) 当所有batch都运行完毕,会调用metric的get_metric方法,并且以返回的值作为evaluation的结果 | |||
metric.get_metric() | |||
``` | |||
#### Trainer: 对训练过程的封装。 | |||
里面比较重要的function是train() | |||
train()中的运行过程 | |||
``` | |||
(1) 创建batch | |||
batch = Batch(dataset, batch_size, sampler=sampler) | |||
for batch_x, batch_y in batch: | |||
# ... | |||
batch_x,batch_y都是dict。batch_x是DataSet中被设置为input的field;batch_y是DataSet中被设置为target的field。 | |||
两个dict中的key就是DataSet中的key,value会根据情况做好padding的tensor。 | |||
(2)会将batch_x, batch_y中tensor移动到model所在的device | |||
(3)根据model.forward的参数列表, 从batch_x中取出需要传递给forward的数据。 | |||
(4)获取model.forward的输出结果pred_dict,并与batch_y一起传递给loss函数, 求得loss | |||
(5)对loss进行反向梯度并更新参数 | |||
(6) 如果有验证集,则需要做验证 | |||
tester = Tester(model, dev_data,metric) | |||
eval_results = tester.test() | |||
(7) 如果eval_results是当前的最佳结果,则保存模型。 | |||
``` | |||
#### 其他 | |||
Trainer中还提供了"预跑"的功能。该功能通过check_code_level管理,如果check_code_level为-1,则不进行"预跑"。 | |||
check_code_level=0,1,2代表不同的提醒级别。 | |||
目前不同提醒级别对应的是对DataSet中设置为input或target但又没有使用的field的提醒级别。 | |||
0是忽略(默认);1是会warning发生了未使用field的情况;2是出现了unused会直接报错并退出运行 | |||
"预跑"的主要目的有两个: | |||
- 防止train完了之后进行evaluation的时候出现错误。之前的train就白费了 | |||
- 由于存在"键映射",直接运行导致的报错可能不太容易debug,通过"预跑"过程的报错会有一些debug提示 | |||
"预跑"会进行以下的操作: | |||
- 使用很小的batch_size, 检查batch_x中是否包含Model.forward所需要的参数。只会运行两个循环。 | |||
- 将Model.foward的输出pred_dict与batch_y输入到loss中, 并尝试backward. 不会更新参数,而且grad会被清零 | |||
如果传入了dev_data,还将进行metric的测试 | |||
- 创建Tester,并传入少量数据,检测是否可以正常运行 | |||
"预跑"操作是在Trainer初始化的时候执行的。 | |||
正常情况下,应该不需要改动"预跑"的代码。但如果你遇到bug或者有什么好的建议,欢迎在开发群或者github提交issue。 | |||