The way to Construct and Practice a Transformer Mannequin from Scratch with Hugging Face Transformers

Hugging Face Transformers library supplies instruments for simply loading and utilizing pre-trained Language Fashions (LMs) based mostly on the transformer structure. However, do you know this library additionally lets you implement and practice your transformer mannequin from scratch? This tutorial illustrates how by a step-by-step sentiment classification instance.

Essential word: Coaching a transformer mannequin from scratch is computationally costly, with a coaching loop usually requiring hours to say the least. To run the code on this tutorial, it’s extremely beneficial to have entry to high-performance computing sources, be it on-premises or by way of a cloud supplier.

Step-by-Step Course of

Preliminary Setup and Dataset Loading

Relying on the kind of Python improvement setting you might be engaged on, chances are you’ll want to put in Hugging Face’s transformers and datasets libraries, in addition to the speed up library to coach your transformer mannequin in a distributed computing setting.

!pip set up transformers datasets
!pip set up speed up -U

As soon as the mandatory libraries are put in, let’s load the feelings dataset for sentiment classification of Twitter messages from Hugging Face hub:

from datasets import load_dataset
dataset = load_dataset('jeffnyman/feelings')

Utilizing the info for coaching a transformer-based LM requires tokenizing the textual content. The next code initializes a BERT tokenizer (BERT is a household of transformer fashions appropriate for textual content classification duties), defines a operate to tokenize textual content knowledge with padding and truncation, and applies it to the dataset in batches.

from transformers import AutoTokenizer

def tokenize_function(examples):
  return tokenizer(examples['text'], padding="max_length", truncation=True)

tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
tokenized_datasets = dataset.map(tokenize_function, batched=True)

Earlier than transferring on to initialize the transformer mannequin, let’s confirm the distinctive labels within the dataset. Having a verified set of present class labels helps stop GPU-related errors throughout coaching by verifying label consistency and correctness. We are going to use this label set afterward.

unique_labels = set(tokenized_datasets['train']['label'])
print(f"Unique labels in the training set: {unique_labels}")

def check_labels(dataset):
  for label in dataset['train']['label']:
    if label not in unique_labels:
      print(f"Found invalid label: {label}")

check_labels(tokenized_datasets)

Subsequent, we create and outline a mannequin configuration, after which instantiate the transformer mannequin with this configuration. That is the place we specify hyperparameters in regards to the transformer structure like embedding measurement, variety of consideration heads, and the beforehand calculated set of distinctive labels, key in constructing the ultimate output layer for sentiment classification.

from transformers import BertConfig
from transformers import BertForSequenceClassification

config = BertConfig(
vocab_size=tokenizer.vocab_size,
hidden_size=512,
num_hidden_layers=6,
num_attention_heads=8,
intermediate_size=2048,
max_position_embeddings=512,
num_labels=len(unique_labels)
)

mannequin = BertForSequenceClassification(config)

We’re virtually prepared to coach our transformer mannequin. It simply stays to instantiate two obligatory cases: TrainingArguments, with specs in regards to the coaching loop such because the variety of epochs, and Coach, which glues collectively the mannequin occasion, the coaching arguments, and the info utilized for coaching and validation.

from transformers import TrainingArguments, Coach

training_args = TrainingArguments(
  output_dir="./results",
  evaluation_strategy="epoch",
  learning_rate=2e-5,
  per_device_train_batch_size=16,
  per_device_eval_batch_size=16,
  num_train_epochs=3,
  weight_decay=0.01,
)

coach = Coach(
  mannequin=mannequin,
  args=training_args,
  train_dataset=tokenized_datasets["train"],
  eval_dataset=tokenized_datasets["test"],
)

Time to coach the mannequin, sit again, and loosen up. Bear in mind this instruction will take a big period of time to finish:

As soon as skilled, your transformer mannequin must be prepared for passing in enter examples for sentiment prediction.

Troubleshooting

If issues seem or persist when executing the coaching loop or throughout its setup, chances are you’ll want to examine the configuration of the GPU/CPU sources getting used. As an illustration, if utilizing a CUDA GPU, including these directions in the beginning of your code will help stop errors within the coaching loop:

import os
os.environ["CUDA_LAUNCH_BLOCKING"] = "1"

These strains disable the GPU and make CUDA operations synchronous, offering extra rapid and correct error messages for debugging.

Alternatively, in case you are making an attempt this code in a Google Colab occasion, likelihood is this error message exhibits up throughout execution, even if in case you have beforehand put in the speed up library:

ImportError: Utilizing the `Coach` with `PyTorch` requires `speed up>=0.21.0`: Please run `pip set up transformers[torch]` or `pip set up speed up -U`

To handle this situation, strive restarting your session within the ‘Runtime’ menu: the speed up library usually requires resetting the run setting after being put in.

Abstract and Wrap-Up

This tutorial showcased the important thing steps to construct your transformer-based LM from scratch utilizing Hugging Face libraries. The principle steps and parts concerned will be summarized as:

• Loading the dataset and tokenizing the textual content knowledge.
• Initializing your mannequin by utilizing a mannequin configuration occasion for the kind of mannequin (language process) it’s supposed for, e.g. BertConfig.
• Organising a Coach and TrainingArguments cases and working the coaching loop.

As a subsequent studying step, we encourage you to discover how one can make predictions and inferences together with your newly skilled mannequin.
 
 

Iván Palomares Carrascosa is a pacesetter, author, speaker, and adviser in AI, machine studying, deep studying & LLMs. He trains and guides others in harnessing AI in the actual world.