app.py

import os
from datasets import load_dataset
from transformers import TrOCRProcessor, VisionEncoderDecoderModel, Seq2SeqTrainer, Seq2SeqTrainingArguments, default_data_collator

# Check if model already exists
if os.path.exists("trained_model"):
    print("✅ Model already exists. Skipping training.")
    exit()

print("🚀 Starting training...")

# Load only 100 samples for faster CPU training
ds = load_dataset("Azu/Handwritten-Mathematical-Expression-Convert-LaTeX", split="train[:100]")

# DEBUG: Inspect a few labels
print("\n🔍 Sample labels from dataset:")
for i in range(5):
    print(f"{i}: {ds[i]['label']} (type: {type(ds[i]['label'])})")

processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-handwritten")

# Safely extract label string from possible dict or str
def safe_get_label(example):
    label = example.get("label")
    if isinstance(label, dict) and "latex" in label:
        return label["latex"]
    elif isinstance(label, str):
        return label
    else:
        return None

def preprocess(example):
    label_str = safe_get_label(example)
    if not isinstance(label_str, str) or label_str.strip() == "":
        return {}  # Skip if label is invalid

    # Convert image to RGB
    img = example["image"].convert("RGB")
    inputs = processor(images=img, return_tensors="pt")

    # Tokenize label
    labels = processor.tokenizer(
        label_str,
        truncation=True,
        padding="max_length",
        max_length=128
    ).input_ids

    return {
        "pixel_values": inputs.pixel_values[0],
        "labels": labels
    }

# Preprocess and filter
ds = ds.map(preprocess, remove_columns=["image", "label"])
ds = ds.filter(lambda ex: "labels" in ex and ex["labels"] is not None)

# Check number of remaining examples
print(f"✅ Total usable training samples: {len(ds)}")
if len(ds) == 0:
    raise RuntimeError("❌ No usable training samples after preprocessing.")

# Model setup
model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-handwritten")
model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
model.config.pad_token_id = processor.tokenizer.pad_token_id

training_args = Seq2SeqTrainingArguments(
    output_dir="trained_model",
    per_device_train_batch_size=2,
    num_train_epochs=1,
    learning_rate=5e-5,
    logging_steps=10,
    save_steps=500,
    fp16=False,
    push_to_hub=False,
)

trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    train_dataset=ds,
    tokenizer=processor.tokenizer,
    data_collator=default_data_collator,
)

trainer.train()
print("✅ Training completed")

# Save model
model.save_pretrained("trained_model")
processor.save_pretrained("trained_model")
print("✅ Model saved to trained_model/")