train.py

import os

os.environ["HF_HOME"] = "/home/jovyan/work/learn-ml/huggingface"

import torch
from torch.utils.data import DataLoader

from transformers import AdamW, AutoTokenizer, AutoModelForSequenceClassification, DataCollatorWithPadding, get_scheduler
from datasets import load_dataset, load_metric
from accelerate import Accelerator

from tqdm.auto import tqdm



checkpoint = 'bert-base-uncased'
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)


raw_datasets = load_dataset("glue", "mrpc")



def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)

tokenized_dataset = raw_datasets.map(tokenize_function, batched=True)


data_collator = DataCollatorWithPadding(tokenizer=tokenizer)


samples = tokenized_dataset["train"][:8]
samples = {k: v for k,v in samples.items() if k not in ["idx", "sentence1", "sentence2"]}
print([len(x) for x in samples["input_ids"]])


tokenized_dataset = tokenized_dataset.remove_columns(["sentence1","sentence2","idx"])
tokenized_dataset = tokenized_dataset.rename_column("label", "labels")
tokenized_dataset.set_format("torch")
tokenized_dataset.column_names["train"]


train_dataloader = DataLoader(
    tokenized_dataset["train"], shuffle=True, batch_size=8, collate_fn=data_collator,
)

eval_dataloader = DataLoader(
    tokenized_dataset["validation"], batch_size=8, collate_fn=data_collator,
)



accelerator = Accelerator()

model = AutoModelForSequenceClassification.from_pretrained(checkpoint, num_labels=2)

optimizer = AdamW(model.parameters(), lr=3e-5)


train_dataloader, eval_dataloader, model, optimizer = accelerator.prepare(train_dataloader, eval_dataloader, model, optimizer)

num_epochs = 3
num_training_steps = num_epochs * len(train_dataloader)
lr_scheduler = get_scheduler(
    "linear",
    optimizer=optimizer,
    num_warmup_steps=0,
    num_training_steps=num_training_steps,
)

print(num_training_steps)


progress_bar = tqdm(range(num_training_steps))


model.train()

for epoch in range(num_epochs):
    for batch in train_dataloader:
        outputs = model(**batch)
        loss = outputs.loss
        accelerator.backward(loss)
        
        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()
        progress_bar.update(1)