train_model.py

# -*- coding: utf-8 -*-
"""train_model.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1BMInZz4vjJ1PfgTbbqIknpJYcbM5cwV0
"""

import torch
import numpy as np
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments

print("Downloading dataset...")
dataset = load_dataset("papluca/language-identification", split="train")

target_langs = {'en', 'fr', 'es', 'de'}
filtered_dataset = dataset.filter(lambda example: example['labels'] in target_langs)

label2id = {"en": 0, "fr": 1, "es": 2, "de": 3}
id2label = {0: "en", 1: "fr", 2: "es", 3: "de"}

model_ckpt = "distilbert-base-multilingual-cased"
tokenizer = AutoTokenizer.from_pretrained(model_ckpt)

def preprocess(examples):
    tokenized = tokenizer(examples["text"], truncation=True, padding="max_length", max_length=64)
    tokenized["labels"] = [label2id[lang] for lang in examples["labels"]]
    return tokenized

print("Preprocessing data...")

train_subset = filtered_dataset.shuffle(seed=42).select(range(1500))
tokenized_data = train_subset.map(preprocess, batched=True)

model = AutoModelForSequenceClassification.from_pretrained(
    model_ckpt,
    num_labels=4,
    id2label=id2label,
    label2id=label2id
)

args = TrainingArguments(
    output_dir="my_real_model",
    learning_rate=2e-5,
    per_device_train_batch_size=16,
    num_train_epochs=2,
    weight_decay=0.01,
    save_strategy="no",
    use_cpu=True

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_data,
    tokenizer=tokenizer,
)

print("Starting training...")
trainer.train()

print("Saving model to './production_model'...")
trainer.save_model("production_model")
print("Done!")