train_sapbert_extended_fixed.py

# /// script
# dependencies = [
#   "transformers>=4.38.0",
#   "datasets>=2.16.0",
#   "torch>=2.1.0",
#   "scikit-learn>=1.3.0",
#   "accelerate>=0.26.0",
# ]
# ///

"""
SAPBERT Training on Extended FDA LOINC2SDTM Dataset
Multi-label classification for 8 SDTM fields
FIXED VERSION with better error handling and logging
"""

import os
import sys
import json
import traceback
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModel,
    TrainingArguments,
    Trainer,
)
import torch
import torch.nn as nn

def log(msg):
    """Print with flush to ensure immediate output"""
    print(msg, flush=True)

try:
    log("=" * 80)
    log("SAPBERT TRAINING - Extended FDA Dataset (8 SDTM Fields)")
    log("FIXED VERSION - Enhanced error handling and logging")
    log("=" * 80)

    # Configuration
    BASE_MODEL = "cambridgeltl/SapBERT-from-PubMedBERT-fulltext"
    DATASET_NAME = "panikos/loinc2sdtm-fda-extended"
    OUTPUT_DIR = "loinc2sdtm-sapbert-extended-model"
    HF_USERNAME = "panikos"

    # Fields to train on (using only the 8 core SDTM fields)
    TRAIN_FIELDS = [
        'lbtestcd',
        'lbtest',
        'lbspec',
        'lbstresu',
        'lbmethod',
        'lbptfl',
        'lbrestyp',
        'lbresscl',
    ]

    log("\n[1/7] Loading extended FDA structured dataset...")
    log(f"  Dataset: {DATASET_NAME}")

    try:
        dataset = load_dataset(DATASET_NAME, split="train")
        log(f"  ✓ Loaded {len(dataset)} examples from FDA source")
        log(f"  ✓ Training on {len(TRAIN_FIELDS)} SDTM fields")
        log(f"  ✓ Dataset features: {list(dataset.features.keys())}")
    except Exception as e:
        log(f"  ✗ FAILED to load dataset!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    # Build vocabularies
    log("\n[2/7] Building field vocabularies...")
    vocabularies = {field: set() for field in TRAIN_FIELDS}

    try:
        for i, example in enumerate(dataset):
            if i % 500 == 0:
                log(f"  Processing example {i}/{len(dataset)}...")
            for field in TRAIN_FIELDS:
                value = example.get(field, '')
                if value and value.strip():
                    vocabularies[field].add(value.upper().strip())

        vocabularies = {k: sorted(list(v)) for k, v in vocabularies.items()}
        log("  ✓ Vocabulary sizes:")
        for field, vocab in vocabularies.items():
            log(f"    {field.upper()}: {len(vocab)} unique values")
    except Exception as e:
        log(f"  ✗ FAILED to build vocabularies!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    # Create label mappings
    try:
        label2id = {
            field: {label: idx for idx, label in enumerate(vocab)}
            for field, vocab in vocabularies.items()
        }
        id2label = {
            field: {idx: label for label, idx in mapping.items()}
            for field, mapping in label2id.items()
        }
        log("  ✓ Label mappings created")
    except Exception as e:
        log(f"  ✗ FAILED to create label mappings!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    log("\n[3/7] Loading SAPBERT model...")
    log(f"  Base model: {BASE_MODEL}")

    try:
        tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
        log("  ✓ Tokenizer loaded")
        base_model = AutoModel.from_pretrained(BASE_MODEL)
        log("  ✓ Base model loaded successfully!")
    except Exception as e:
        log(f"  ✗ FAILED to load SAPBERT model!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    # Multi-label classifier with LOINC metadata as input
    class LOINC2SDTMClassifier(nn.Module):
        def __init__(self, base_model, num_classes_dict):
            super().__init__()
            self.encoder = base_model
            self.config = base_model.config
            self.hidden_size = base_model.config.hidden_size

            self.classifiers = nn.ModuleDict({
                field: nn.Sequential(
                    nn.Linear(self.hidden_size, self.hidden_size // 2),
                    nn.ReLU(),
                    nn.Dropout(0.1),
                    nn.Linear(self.hidden_size // 2, num_classes)
                )
                for field, num_classes in num_classes_dict.items()
            })

        def forward(self, input_ids, attention_mask, labels=None):
            outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
            cls_embedding = outputs.last_hidden_state[:, 0, :]

            logits = {
                field: classifier(cls_embedding)
                for field, classifier in self.classifiers.items()
            }

            loss = None
            if labels is not None:
                loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
                losses = []
                for field in logits.keys():
                    if field in labels:
                        field_loss = loss_fct(logits[field], labels[field])
                        if not torch.isnan(field_loss):
                            losses.append(field_loss)
                if losses:
                    loss = sum(losses) / len(losses)

            return {'loss': loss, 'logits': logits}

    try:
        num_classes_dict = {field: len(vocab) for field, vocab in vocabularies.items()}
        model = LOINC2SDTMClassifier(base_model, num_classes_dict)
        total_params = sum(p.numel() for p in model.parameters())
        trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
        log(f"\n[4/7] Classifier created:")
        log(f"  Total parameters: {total_params:,}")
        log(f"  Trainable parameters: {trainable_params:,}")
        log(f"  ✓ Model architecture initialized")
    except Exception as e:
        log(f"  ✗ FAILED to create classifier!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    # Prepare dataset
    class LOINC2SDTMDataset(torch.utils.data.Dataset):
        def __init__(self, dataset, tokenizer, label2id, train_fields):
            self.examples = []
            log(f"  Creating dataset wrapper for {len(dataset)} examples...")

            for i, example in enumerate(dataset):
                if i % 500 == 0:
                    log(f"    Processed {i}/{len(dataset)} examples...")

                # Create rich input combining LOINC code and metadata
                loinc_code = example['loinc_code']
                component = example.get('component', '')
                property_val = example.get('property', '')
                system = example.get('system', '')

                # Rich input: LOINC code + key metadata
                input_text = f"{loinc_code} {component} {property_val} {system}"

                # Tokenize input
                encoding = tokenizer(
                    input_text,
                    padding='max_length',
                    truncation=True,
                    max_length=64,
                    return_tensors='pt'
                )

                # Get labels for trained fields
                labels = {}
                for field in train_fields:
                    value = example.get(field, '')
                    if value and value.strip():
                        value_upper = value.upper().strip()
                        if value_upper in label2id[field]:
                            labels[field] = label2id[field][value_upper]
                        else:
                            labels[field] = -100
                    else:
                        labels[field] = -100

                self.examples.append({
                    'input_ids': encoding['input_ids'].squeeze(0),
                    'attention_mask': encoding['attention_mask'].squeeze(0),
                    'labels': labels
                })

        def __len__(self):
            return len(self.examples)

        def __getitem__(self, idx):
            return self.examples[idx]

    log("\n[5/7] Preparing training data...")
    try:
        train_dataset = LOINC2SDTMDataset(dataset, tokenizer, label2id, TRAIN_FIELDS)
        log(f"  ✓ Prepared {len(train_dataset)} training examples")
    except Exception as e:
        log(f"  ✗ FAILED to prepare training data!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    # Custom collator
    def collate_fn(batch):
        input_ids = torch.stack([item['input_ids'] for item in batch])
        attention_mask = torch.stack([item['attention_mask'] for item in batch])
        labels = {
            field: torch.tensor([item['labels'][field] for item in batch])
            for field in TRAIN_FIELDS
        }
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask,
            'labels': labels
        }

    # Training args
    training_args = TrainingArguments(
        output_dir=OUTPUT_DIR,
        num_train_epochs=10,
        per_device_train_batch_size=32,
        gradient_accumulation_steps=1,
        learning_rate=2e-5,
        lr_scheduler_type="cosine",
        warmup_ratio=0.1,
        logging_steps=10,  # More frequent logging
        logging_first_step=True,
        save_strategy="epoch",
        save_total_limit=2,
        fp16=False,
        bf16=True,
        report_to="none",
        push_to_hub=True,
        hub_model_id=f"{HF_USERNAME}/{OUTPUT_DIR}",
        hub_strategy="end",
    )

    log("\n[6/7] Training configuration:")
    log(f"  Epochs: {training_args.num_train_epochs}")
    log(f"  Batch size: {training_args.per_device_train_batch_size}")
    log(f"  Learning rate: {training_args.learning_rate}")
    log(f"  Steps per epoch: ~{len(train_dataset) // training_args.per_device_train_batch_size}")
    log(f"  Total steps: ~{(len(train_dataset) // training_args.per_device_train_batch_size) * training_args.num_train_epochs}")
    log(f"  Input: LOINC code + metadata (component, property, system)")
    log(f"  Output: {len(TRAIN_FIELDS)} SDTM fields")
    log(f"  Mixed precision: {'BF16' if training_args.bf16 else 'FP16' if training_args.fp16 else 'FP32'}")

    # Custom trainer
    class MultiLabelTrainer(Trainer):
        def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
            labels = inputs.pop("labels")
            outputs = model(**inputs, labels=labels)
            loss = outputs["loss"]

            # Log loss periodically
            if self.state.global_step % 10 == 0:
                log(f"  Step {self.state.global_step}: loss = {loss.item():.4f}")

            return (loss, outputs) if return_outputs else loss

        def get_train_dataloader(self):
            from torch.utils.data import DataLoader
            return DataLoader(
                self.train_dataset,
                batch_size=self.args.per_device_train_batch_size,
                collate_fn=collate_fn,
                shuffle=True
            )

    try:
        trainer = MultiLabelTrainer(
            model=model,
            args=training_args,
            train_dataset=train_dataset,
        )
        log("  ✓ Trainer initialized")
    except Exception as e:
        log(f"  ✗ FAILED to initialize trainer!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    log("\n[7/7] Starting training...")
    log("=" * 80)
    log("This will take approximately 15-20 minutes on A10G GPU")
    log("=" * 80)

    try:
        trainer.train()
        log("\n" + "=" * 80)
        log("✓ Training completed successfully!")
        log("=" * 80)
    except Exception as e:
        log(f"\n✗ TRAINING FAILED!")
        log(f"Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    log("\nSaving model and vocabularies...")
    try:
        trainer.save_model(OUTPUT_DIR)
        log("  ✓ Model saved")
        tokenizer.save_pretrained(OUTPUT_DIR)
        log("  ✓ Tokenizer saved")

        # Save vocabularies and metadata
        vocab_file = os.path.join(OUTPUT_DIR, "vocabularies.json")
        with open(vocab_file, 'w') as f:
            json.dump({
                'vocabularies': vocabularies,
                'label2id': label2id,
                'id2label': id2label,
                'train_fields': TRAIN_FIELDS
            }, f, indent=2)
        log("  ✓ Vocabularies saved")
    except Exception as e:
        log(f"  ✗ FAILED to save model!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    log("\nPushing to Hub...")
    try:
        trainer.push_to_hub()
        log("  ✓ Model pushed to Hub")
    except Exception as e:
        log(f"  ✗ FAILED to push to Hub!")
        log(f"  Error: {str(e)}")
        traceback.print_exc()
        sys.exit(1)

    log("\n" + "=" * 80)
    log("✓ SUCCESS! Model training and upload complete!")
    log("=" * 80)
    log(f"Model available at: https://huggingface.co/{HF_USERNAME}/{OUTPUT_DIR}")
    log(f"Trained on {len(TRAIN_FIELDS)} SDTM fields with rich LOINC metadata")
    log(f"Total examples: {len(train_dataset)}")
    log("=" * 80)

except Exception as e:
    log("\n" + "=" * 80)
    log("✗ FATAL ERROR - Training script crashed!")
    log("=" * 80)
    log(f"Error: {str(e)}")
    log("\nFull traceback:")
    traceback.print_exc()
    sys.exit(1)