{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# E2. 使用 Bert + prompt 完成 SST2 分类\n", "\n", "  1   基础介绍：`prompt-based model`简介、与`fastNLP`的结合\n", "\n", "  2   准备工作：`P-Tuning v2`原理概述、`P-Tuning v2`模型搭建\n", "\n", "  3   模型训练：加载`tokenizer`、预处理`dataset`、模型训练与分析" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 1. 基础介绍：prompt-based model 简介、与 fastNLP 的结合\n", "\n", "  本示例使用`GLUE`评估基准中的`SST2`数据集，通过`prompt-based tuning`方式\n", "\n", "    微调`bert-base-uncased`模型，实现文本情感的二分类，在此之前本示例\n", "\n", "    将首先简单介绍提示学习模型的研究，以及与`fastNLP v0.8`结合的优势\n", "\n", "**`prompt`**，**提示词、提词器**，最早出自**`PET`**，\n", "\n", "  \n", "\n", "**`prompt-based tuning`**，**基于提示的微调**，描述\n", "\n", "  **`prompt-based model`**，**基于提示的模型**\n", "\n", "**`prompt-based model`**，**基于提示的模型**，举例\n", "\n", "  案例一：**`P-Tuning v1`**\n", "\n", "  案例二：**`PromptTuning`**\n", "\n", "  案例三：**`PrefixTuning`**\n", "\n", "  案例四：**`SoftPrompt`**\n", "\n", "使用`fastNLP v0.8`实现`prompt-based model`的优势\n", "\n", "  \n", "\n", "  本示例仍使用了`tutorial-E1`的`SST2`数据集，将`bert-base-uncased`作为基础模型\n", "\n", "    在后续实现中，意图通过将连续的`prompt`与`model`拼接，解决`SST2`二分类任务" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "data": { "text/html": [ "
\n",
       "
\n" ], "text/plain": [ "\n" ] }, "metadata": {}, "output_type": "display_data" }, { "name": "stdout", "output_type": "stream", "text": [ "4.18.0\n" ] } ], "source": [ "import torch\n", "import torch.nn as nn\n", "from torch.optim import AdamW\n", "from torch.utils.data import DataLoader, Dataset\n", "\n", "import transformers\n", "from transformers import AutoTokenizer\n", "from transformers import AutoModelForSequenceClassification\n", "\n", "import sys\n", "sys.path.append('..')\n", "\n", "import fastNLP\n", "from fastNLP import Trainer\n", "from fastNLP.core.metrics import Accuracy\n", "\n", "print(transformers.__version__)\n", "\n", "task = 'sst2'\n", "model_checkpoint = 'bert-base-uncased'" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 2. 准备工作：P-Tuning v2 原理概述、P-Tuning v2 模型搭建\n", "\n", "  本示例使用`P-Tuning v2`作为`prompt-based tuning`与`fastNLP v0.8`结合的案例\n", "\n", "    以下首先简述`P-Tuning v2`的论文原理，并由此引出`fastNLP v0.8`的代码实践\n", "\n", "`P-Tuning v2`出自论文 [Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks](https://arxiv.org/pdf/2110.07602.pdf)\n", "\n", "  其主要贡献在于，在`PrefixTuning`等深度提示学习基础上，提升了其在分类标注等`NLU`任务的表现\n", "\n", "    并使之在中等规模模型，主要是参数量在`100M-1B`区间的模型上，获得与全参数微调相同的效果\n", "\n", "  其结构如图所示，\n", "\n", "" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [], "source": [ "class SeqClsModel(nn.Module):\n", " def __init__(self, model_checkpoint, num_labels, pre_seq_len):\n", " nn.Module.__init__(self)\n", " self.num_labels = num_labels\n", " self.back_bone = AutoModelForSequenceClassification.from_pretrained(model_checkpoint, \n", " num_labels=num_labels)\n", " self.embeddings = self.back_bone.get_input_embeddings()\n", "\n", " for param in self.back_bone.parameters():\n", " param.requires_grad = False\n", " \n", " self.pre_seq_len = pre_seq_len\n", " self.prefix_tokens = torch.arange(self.pre_seq_len).long()\n", " self.prefix_encoder = nn.Embedding(self.pre_seq_len, self.embeddings.embedding_dim)\n", " \n", " def get_prompt(self, batch_size):\n", " prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.back_bone.device)\n", " prompts = self.prefix_encoder(prefix_tokens)\n", " return prompts\n", "\n", " def forward(self, input_ids, attention_mask, labels=None):\n", " \n", " batch_size = input_ids.shape[0]\n", " raw_embedding = self.embeddings(input_ids)\n", " \n", " prompts = self.get_prompt(batch_size=batch_size)\n", " inputs_embeds = torch.cat((prompts, raw_embedding), dim=1)\n", " prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.back_bone.device)\n", " attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1)\n", "\n", " outputs = self.back_bone(inputs_embeds=inputs_embeds, \n", " attention_mask=attention_mask, labels=labels)\n", " return outputs\n", "\n", " def train_step(self, input_ids, attention_mask, labels):\n", " loss = self(input_ids, attention_mask, labels).loss\n", " return {'loss': loss}\n", "\n", " def evaluate_step(self, input_ids, attention_mask, labels):\n", " pred = self(input_ids, attention_mask, labels).logits\n", " pred = torch.max(pred, dim=-1)[1]\n", " return {'pred': pred, 'target': labels}" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "接着，通过确定分类数量初始化模型实例，同时调用`torch.optim.AdamW`模块初始化优化器\n", "\n", "  根据`P-Tuning v2`论文：*Generally, simple classification tasks prefer shorter prompts (less than 20)*\n", "\n", "  此处`pre_seq_len`参数设定为`20`，学习率相应做出调整，其他内容和`tutorial-E1`中的内容一致" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "Some weights of the model checkpoint at bert-base-uncased were not used when initializing BertForSequenceClassification: ['cls.predictions.transform.dense.bias', 'cls.predictions.decoder.weight', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias', 'cls.seq_relationship.weight', 'cls.predictions.transform.dense.weight', 'cls.predictions.bias', 'cls.seq_relationship.bias']\n", "- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n", "- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n", "Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\n", "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" ] } ], "source": [ "model = SeqClsModel(model_checkpoint=model_checkpoint, num_labels=2, pre_seq_len=20)\n", "\n", "optimizers = AdamW(params=model.parameters(), lr=1e-2)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### 3. 模型训练：加载 tokenizer、预处理 dataset、模型训练与分析\n", "\n", "  本示例沿用`tutorial-E1`中的数据集，即使用`GLUE`评估基准中的`SST2`数据集\n", "\n", "    以`bert-base-uncased`模型作为基准，基于`P-Tuning v2`方式微调\n", "\n", "    数据集加载相关代码流程见下，内容和`tutorial-E1`中的内容基本一致\n", "\n", "首先，使用`datasets.load_dataset`加载数据集，使用`transformers.AutoTokenizer`\n", "\n", "  构建`tokenizer`实例，通过`dataset.map`使用`tokenizer`将文本替换为词素序号序列" ] }, { "cell_type": "code", "execution_count": 5, "metadata": { "scrolled": false }, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "Reusing dataset glue (/remote-home/xrliu/.cache/huggingface/datasets/glue/sst2/1.0.0/dacbe3125aa31d7f70367a07a8a9e72a5a0bfeb5fc42e75c9db75b96da6053ad)\n" ] }, { "data": { "application/vnd.jupyter.widget-view+json": { "model_id": "b72eeebd34354a88a99b2e07ec9a86df", "version_major": 2, "version_minor": 0 }, "text/plain": [ " 0%| | 0/3 [00:00