train.py

from datasets import load_dataset
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
import torch
from PIL import Image
import requests

# Load dataset
dataset = load_dataset("nielsr/funsd")

# Load pre-trained model and processor
processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-printed")
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-printed")

# Preprocess the dataset
def preprocess_images(examples):
    images = [Image.open(img).convert("RGB") for img in examples['image_path']]
    pixel_values = processor(images=images, return_tensors="pt").pixel_values
    return {"pixel_values": pixel_values}

encoded_dataset = dataset.map(preprocess_images, batched=True)

# Preprocess the labels
max_length = 64

def preprocess_labels(examples):
    labels = processor.tokenizer(examples['words'], is_split_into_words=True, padding="max_length", max_length=max_length, truncation=True)
    return labels

encoded_dataset = encoded_dataset.map(preprocess_labels, batched=True)

# Prepare for training
model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
model.config.pad_token_id = processor.tokenizer.pad_token_id

# Define training arguments
from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments

training_args = Seq2SeqTrainingArguments(
    output_dir="./trocr-finetuned-funsd",
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    learning_rate=5e-5,
    num_train_epochs=3,
    weight_decay=0.01,
    logging_dir="./trocr-finetuned-funsd/logs",
    logging_steps=10,
    evaluation_strategy="epoch",
    save_strategy="epoch",
)
trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=encoded_dataset["train"],
    eval_dataset=encoded_dataset["test"],
    tokenizer=processor.tokenizer,
)

# Train the model
trainer.train()