【adapter-transformers】：Installation & QuickStart（一、安装与快速启动）

【要求】：

adapter-transformers是Huggingface的transformers库的直接替代品。它目前支持Python 3.8+和PyTorch 1.12.1+。因此必须先安装PyTorch。

一、安装（使用pip）

pip install adapter-transformers

二、快速启动（使用训预练的适配器进行推理）

下面的例子展示了如何使用带有适配器的预训练的Transformer模型。我们的目标是预测给定句子的情感。

我们在这个例子中使用BERT，所以我们首先使用BertAdapterModel类从HuggingFace的模型中心加载一个预训练的BertTokenizer来编码输入句子和一个预训练的BERT -base uncase检查点:

import os

import torch
from transformers import BertTokenizer
from transformers.adapters import BertAdapterModel, AutoAdapterModel

# Load pre-trained BERT tokenizer from HuggingFace
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')

# An input sentence
sentence = "It's also, clearly, great fun."

# Tokenize the input sentence and create a PyTorch input tensor
input_data = tokenizer(sentence, return_tensors="pt")

# Load pre-trained BERT model from HuggingFace Hub
# The `BertAdapterModel` class is specifically designed for working with adapters
# It can be used with different prediction heads
model = BertAdapterModel.from_pretrained('bert-base-uncased')

加载模型后，我们现在添加一个预训练的任务适配器，它对AdapterHub中的任务很有用。

在这种情况下，对于情感分类，我们使用在SST-2数据集上训练的适配器。与适配器一起加载的任务预测头（ task prediction head）为我们的句子提供了一个类标签:

# Load pre-trained task adapter from Adapter Hub
# This method call will also load a pre-trained classification head for the adapter task
adapter_name = model.load_adapter("sentiment/sst-2@ukp", config='pfeiffer')

# Activate the adapter we just loaded, so that it is used in every forward pass
model.set_active_adapters(adapter_name)

# Predict output tensor
outputs = model(**input_data)

# Retrieve the predicted class label
predicted = torch.argmax(outputs[0]).item()
assert predicted == 1

为了保存我们预先训练的模型和适配器，我们可以像下面这样轻松地存储和重新加载它们:

# For the sake of this demonstration an example path for loading and storing is given below
example_path = os.path.join(os.getcwd(), "adapter-quickstart")

# Save model
model.save_pretrained(example_path)
# Save adapter
model.save_adapter(example_path, adapter_name)

# Load model, similar to HuggingFace's AutoModel class, 
# you can also use AutoAdapterModel instead of BertAdapterModel
model = AutoAdapterModel.from_pretrained(example_path)
model.load_adapter(example_path)

最后，如果我们已经完成了适配器的工作，我们可以通过停用和删除适配器来将基本Transformer恢复到其原始形式:

# Deactivate all adapters
model.set_active_adapters(None)
# Delete the added adapter
model.delete_adapter(adapter_name)

---

原文链接：Installation — adapter-transformers documentation (adapterhub.ml)