hugging face tutorial - Chinese translation - Loading pre-trained instances with AutoClass

使用 AutoClass Load a pretrained instance

Since there are so many different ones Transformer 体系结构,为您的 checkpoint 创建一个 Transformer Architecture is a challenge.作为 Transformers core 哲学的一部分,AutoClass 可以从给定的checkoutThe correct architecture is automatically inferred and loaded,Thus making the library easy、Simple and flexible to use.来自 pretrained method It allows you to quickly load a pre-trained model for any architecture,This way you don't have to invest time and resources to train a model from scratch.这种类型的checkoutCode-agnostic means if your code works for onecheckout,then it will work for the other onecheckout——as long as it is trained for similar tasks——Even if the architecture is different.

请记住,Architecture refers to the skeleton of a model,而checkoutis the weight for a given architecture.例如,BERT 是一种架构,而 BERT-base-uncased 是一种checkout.Model is a generic term,It can represent architecture,也可以表示checkout.

在本教程中,学习:

• Load a pretrained tokenizer
• Load a trained feature extractor
• Load a pre-trained processor
• 加载一个预先训练好的模型

AutoTokenizer

几乎每个 NLP Tasks all start with a tokenizer.The tokenizer converts your input into a format that the model can handle.
用 AutoTokenizer.from_pretrained()加载一个 tokenizer:

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

Then label your input like below:

sequence = "In a hole in the ground there lived a hobbit."print(tokenizer(sequence))
{
    'input_ids': [101, 1999, 1037, 4920, 1999, 1996, 2598, 2045, 2973, 1037, 7570, 10322, 4183, 1012, 102], 
 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 
 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}

Automatic feature extractor

for audio and visual tasks,Feature extractors process audio signals or images into the correct input format.
使用 AutoFeatureExtractor.from_pretrainedLoad a feature extractor:

from transformers import AutoFeatureExtractor
feature_extractor = AutoFeatureExtractor.from_pretrained(
... "ehcalabres/wav2vec2-lg-xlsr-en-speech-emotion-recognition"... )

automatic processor

Multi-pass tasks require a processor that combines both preprocessing tools.例如,layoutlmv2The model needs a feature extractor to process the image,A tokenizer is also required to process the text; The processor combines the two.
用 AutoProcessor.from_pretrained加载处理器:

from transformers import AutoProcessor
processor = AutoProcessor.from_pretrained("microsoft/layoutlmv2-base-uncased")

AutoModel

Pytorch

最后,AutoModelFor Class allows you to load a pretrained model for a given task(See the full list of available tasks here).例如,使用 AutoModelForSequenceClassification.from_pretrainedLoad a sequence classification model

from transformers import AutoModelForSequenceClassification

Easily reuse the samecheckoutto load schemas for different tasks:

from transformers import AutoModelForTokenClassification
model = AutoModelForTokenClassification.from_pretrained("distilbert-base-uncased")

通常,我们建议使用 AutoTokenizer 类和 AutoModelFor class to load pretrained model instances.This will ensure you load the correct architecture every time.在下一个教程中,Learn how to use the newly loaded tokenizer、Feature extractors and processors preprocess the dataset for fine-tuning.

Tensorflow

最后,tfautomatodelfor Class allows you to load a pretrained model for a given task(See here for a full list of available tasks).例如,使用 tfautomatodelforsequenceclassification.
from_pretrainedLoad a sequence classification model

from transformers import TFAutoModelForSequenceClassification
model = TFAutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")

Easily reuse the samecheckoutto load schemas for different tasks:

from transformers import TFAutoModelForTokenClassification
model = TFAutoModelForTokenClassification.from_pretrained("distilbert-base-uncased")

通常,我们建议使用 AutoTokenizer 类和 TFAutoModelFor class to load pretrained model instances.This will ensure you load the correct architecture every time.在下一个教程中,Learn how to use the newly loaded tokenizer、Feature extractors and processors preprocess the dataset for fine-tuning.

This article is the translation of the hug face tutorial,仅学习记录

原文链接