import json import argparse import os import random import torch import torch.nn as nn from torch.utils.data import Dataset, DataLoader, random_split # Using LayoutLMv3TokenizerFast, LayoutLMv3Model from transformers import LayoutLMv3TokenizerFast, LayoutLMv3Model from transformers.utils import cached_file from safetensors.torch import load_file from TorchCRF import CRF from torch.optim import AdamW from tqdm import tqdm from sklearn.metrics import precision_recall_fscore_support # --- Configuration for Augmentation --- MAX_BBOX_DIMENSION = 1000 # Corrected to 1000 to match LayoutLMv3 requirement MAX_SHIFT = 30 AUGMENTATION_FACTOR = 1 # ------------------------------------- # --- Hugging Face Model ID --- HF_MODEL_ID = "heerjtdev/edugenius" # ----------------------------- # ------------------------- # Step 1: Preprocessing (Label Studio → BIO + bboxes) # ------------------------- def preprocess_labelstudio(input_path, output_path): with open(input_path, "r", encoding="utf-8") as f: data = json.load(f) processed = [] total_items = len(data) # Added for potential verbose logging print(f"🔄 Starting preprocessing of {total_items} documents My name is Aastik!! BOOBS...") for item in data: words = item["data"]["original_words"] bboxes = item["data"]["original_bboxes"] labels = ["O"] * len(words) # --- NEW: Bounding Box Normalization/Clamping --- # Defensively ensures all coordinates are within the [0, 1000] range # required by LayoutLMv3's spatial position embeddings. clamped_bboxes = [] for bbox in bboxes: # Clamp coordinates to [0, 1000] x_min, y_min, x_max, y_max = bbox new_x_min = max(0, min(x_min, 1000)) new_y_min = max(0, min(y_min, 1000)) new_x_max = max(0, min(x_max, 1000)) new_y_max = max(0, min(y_max, 1000)) # Safety check: ensure min <= max (this should rarely trigger # if the original bboxes were valid, but is good practice) if new_x_min > new_x_max: new_x_min = new_x_max if new_y_min > new_y_max: new_y_min = new_y_max clamped_bboxes.append([new_x_min, new_y_min, new_x_max, new_y_max]) # Use the clamped bboxes for the rest of the pipeline final_bboxes = clamped_bboxes # ------------------------------------------------ if "annotations" in item: for ann in item["annotations"]: for res in ann["result"]: # Check if the result item is a span annotation if "value" in res and "labels" in res["value"]: text = res["value"]["text"] tag = res["value"]["labels"][0] # Some tokenizers may split words, so we must find a consecutive word match. text_tokens = text.split() for i in range(len(words) - len(text_tokens) + 1): if words[i:i + len(text_tokens)] == text_tokens: labels[i] = f"B-{tag}" for j in range(1, len(text_tokens)): labels[i + j] = f"I-{tag}" break # Move to next annotation if a match is found processed.append({"tokens": words, "labels": labels, "bboxes": final_bboxes}) with open(output_path, "w", encoding="utf-8") as f: json.dump(processed, f, indent=2, ensure_ascii=False) print(f"✅ Preprocessed data saved to {output_path}") return output_path # ------------------------- # Step 1.5: Bounding Box Augmentation # ------------------------- def translate_bbox(bbox, shift_x, shift_y): """ Translates a single bounding box [x_min, y_min, x_max, y_max] by (shift_x, shift_y) and clamps the coordinates to the valid range [0, MAX_BBOX_DIMENSION]. """ x_min, y_min, x_max, y_max = bbox new_x_min = x_min + shift_x new_y_min = y_min + shift_y new_x_max = x_max + shift_x new_y_max = y_max + shift_y # Clamp the new coordinates (MAX_BBOX_DIMENSION is 1000) new_x_min = max(0, min(new_x_min, MAX_BBOX_DIMENSION)) new_y_min = max(0, min(new_y_min, MAX_BBOX_DIMENSION)) new_x_max = max(0, min(new_x_max, MAX_BBOX_DIMENSION)) new_y_max = max(0, min(new_y_max, MAX_BBOX_DIMENSION)) # Safety check if new_x_min > new_x_max: new_x_min = new_x_max if new_y_min > new_y_max: new_y_min = new_y_max return [new_x_min, new_y_min, new_x_max, new_y_max] def augment_sample(sample): """ Generates a new sample by translating all bounding boxes. """ shift_x = random.randint(-MAX_SHIFT, MAX_SHIFT) shift_y = random.randint(-MAX_SHIFT, MAX_SHIFT) new_sample = sample.copy() # Ensure tokens and labels are copied (they remain unchanged) new_sample["tokens"] = sample["tokens"] new_sample["labels"] = sample["labels"] # Translate all bounding boxes new_bboxes = [translate_bbox(bbox, shift_x, shift_y) for bbox in sample["bboxes"]] new_sample["bboxes"] = new_bboxes return new_sample def augment_and_save_dataset(input_json_path, output_json_path): """ Loads preprocessed data, performs augmentation, and saves the result. """ print(f"🔄 Loading preprocessed data from {input_json_path} for augmentation...") with open(input_json_path, 'r', encoding="utf-8") as f: training_data = json.load(f) augmented_data = [] original_count = len(training_data) print(f"🔄 Starting augmentation (Factor: {AUGMENTATION_FACTOR}, {original_count} documents)...") for i, original_sample in enumerate(training_data): # 1. Add the original sample augmented_data.append(original_sample) # 2. Generate augmented samples for _ in range(AUGMENTATION_FACTOR): if "tokens" in original_sample and "labels" in original_sample and "bboxes" in original_sample: augmented_data.append(augment_sample(original_sample)) else: print(f"Warning: Skipping augmentation for sample {i} due to missing keys.") augmented_count = len(augmented_data) print(f"Dataset Augmentation: Original samples: {original_count}, Total samples: {augmented_count}") # Save the augmented dataset with open(output_json_path, 'w', encoding="utf-8") as f: json.dump(augmented_data, f, indent=2, ensure_ascii=False) print(f"✅ Augmented data saved to {output_json_path}") return output_json_path # ------------------------- # Step 2: Dataset Class # ------------------------- class LayoutDataset(Dataset): def __init__(self, json_path, tokenizer, label2id, max_len=512): with open(json_path, "r", encoding="utf-8") as f: self.data = json.load(f) self.tokenizer = tokenizer self.label2id = label2id self.max_len = max_len def __len__(self): return len(self.data) def __getitem__(self, idx): item = self.data[idx] words, bboxes, labels = item["tokens"], item["bboxes"], item["labels"] # Tokenize encodings = self.tokenizer( words, boxes=bboxes, padding="max_length", truncation=True, max_length=self.max_len, return_offsets_mapping=True, return_tensors="pt" ) # Align labels to word pieces word_ids = encodings.word_ids(batch_index=0) label_ids = [] for word_id in word_ids: if word_id is None: label_ids.append(self.label2id["O"]) # [CLS], [SEP], padding else: label_ids.append(self.label2id.get(labels[word_id], self.label2id["O"])) encodings.pop("offset_mapping") encodings["labels"] = torch.tensor(label_ids) return {key: val.squeeze(0) for key, val in encodings.items()} # ------------------------- # Step 3: Model Architecture (PATCHED TO LOAD WEIGHTS CORRECTLY) # ------------------------- class LayoutLMv3CRF(nn.Module): def __init__(self, model_name, num_labels, device): super().__init__() # 1. Initialize the LayoutLMv3 model using the base class # We start by initializing from the base configuration to ensure all weights are present self.layoutlm = LayoutLMv3Model.from_pretrained("microsoft/layoutlmv3-base") # 2. Try to load the fine-tuned weights from the Hugging Face Hub/Cache try: # This resolves the path to the downloaded model.safetensors in the cache # Assumes you have renamed your file on the Hugging Face Hub to 'model.safetensors' weights_path = cached_file(model_name, "model.safetensors") fine_tuned_weights = load_file(weights_path) # 3. Strip the Mismatching Prefix (Assuming 'layoutlm.' prefix from a previous wrapper) new_state_dict = {} prefix_to_strip = "layoutlm." for key, value in fine_tuned_weights.items(): if key.startswith(prefix_to_strip): new_key = key[len(prefix_to_strip):] new_state_dict[new_key] = value else: new_state_dict[key] = value # 4. Load the fixed state dictionary into the LayoutLMv3Model # strict=False allows us to ignore classifier/CRF weights not in LayoutLMv3Model print("🔄 Successfully loaded and stripped keys. Loading base LayoutLMv3 weights...") # Load only the weights for the transformer body missing_keys, unexpected_keys = self.layoutlm.load_state_dict(new_state_dict, strict=False) print(f"Weights loading done: {len(missing_keys)} missing, {len(unexpected_keys)} unexpected keys.") except Exception as e: print(f"❌ Fine-tuned weights could not be loaded directly and mapped. Starting with random weights.") print(f"Error: {e}") # Fallback: Load the LayoutLMv3 component directly from the Hub ID (will result in random weights for layers) self.layoutlm = LayoutLMv3Model.from_pretrained(model_name) # 5. Initialize the new heads (CRF layer and Classifier) self.dropout = nn.Dropout(0.1) self.classifier = nn.Linear(self.layoutlm.config.hidden_size, num_labels) self.crf = CRF(num_labels) def forward(self, input_ids, bbox, attention_mask, labels=None): outputs = self.layoutlm(input_ids=input_ids, bbox=bbox, attention_mask=attention_mask) sequence_output = self.dropout(outputs.last_hidden_state) emissions = self.classifier(sequence_output) if labels is not None: # Training mode: calculate loss log_likelihood = self.crf(emissions, labels, mask=attention_mask.bool()) return -log_likelihood.mean() else: # Inference mode: decode best path best_paths = self.crf.viterbi_decode(emissions, mask=attention_mask.bool()) return best_paths # ------------------------- # Step 4: Training + Evaluation # ------------------------- def train_one_epoch(model, dataloader, optimizer, device): model.train() total_loss = 0 for batch in tqdm(dataloader, desc="Training"): batch = {k: v.to(device) for k, v in batch.items()} labels = batch.pop("labels") optimizer.zero_grad() loss = model(**batch, labels=labels) loss.backward() optimizer.step() total_loss += loss.item() return total_loss / len(dataloader) def evaluate(model, dataloader, device, id2label): model.eval() all_preds, all_labels = [], [] with torch.no_grad(): for batch in tqdm(dataloader, desc="Evaluating"): batch = {k: v.to(device) for k, v in batch.items()} labels = batch.pop("labels").cpu().numpy() # The model returns a list of lists of predicted labels in inference mode preds = model(**batch) for p, l, mask in zip(preds, labels, batch["attention_mask"].cpu().numpy()): valid = mask == 1 l = l[valid].tolist() all_labels.extend(l) # Ensure pred length matches label length for the unmasked tokens all_preds.extend(p[:len(l)]) # Exclude the "O" label and other special tokens if necessary, but using 'micro' average precision, recall, f1, _ = precision_recall_fscore_support(all_labels, all_preds, average="micro", zero_division=0) return precision, recall, f1 # ------------------------- # Step 5: Main Pipeline (Training) - MODIFIED MODEL/TOKENIZER LOADING # ------------------------- def main(args): # LABELS UPDATED: Added SECTION_HEADING and PASSAGE labels = [ "O", "B-QUESTION", "I-QUESTION", "B-OPTION", "I-OPTION", "B-ANSWER", "I-ANSWER", "B-SECTION_HEADING", "I-SECTION_HEADING", "B-PASSAGE", "I-PASSAGE" ] label2id = {l: i for i, l in enumerate(labels)} id2label = {i: l for l, i in label2id.items()} # --- SETUP: Use a temporary directory for intermediate files --- TEMP_DIR = "temp_intermediate_files" os.makedirs(TEMP_DIR, exist_ok=True) print(f"\n--- SETUP PHASE: Created temp directory: {TEMP_DIR} ---") # 1. Preprocess print("\n--- START PHASE: PREPROCESSING ---") initial_bio_json = os.path.join(TEMP_DIR, "training_data_bio_bboxes.json") preprocess_labelstudio(args.input, initial_bio_json) # 2. Augment print("\n--- START PHASE: AUGMENTATION ---") augmented_bio_json = os.path.join(TEMP_DIR, "augmented_training_data_bio_bboxes.json") final_data_path = augment_and_save_dataset(initial_bio_json, augmented_bio_json) # 3. Load and split augmented dataset print("\n--- START PHASE: MODEL/DATASET SETUP ---") # Load tokenizer from the specified Hugging Face ID tokenizer = LayoutLMv3TokenizerFast.from_pretrained(HF_MODEL_ID) dataset = LayoutDataset(final_data_path, tokenizer, label2id, max_len=args.max_len) val_size = int(0.2 * len(dataset)) train_size = len(dataset) - val_size # Use a fixed seed for reproducibility in split torch.manual_seed(42) train_dataset, val_dataset = random_split(dataset, [train_size, val_size]) train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True) val_loader = DataLoader(val_dataset, batch_size=args.batch_size) # 4. Initialize and load model device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"Using device: {device}") # Pass the Hugging Face ID and device to the custom model wrapper model = LayoutLMv3CRF(HF_MODEL_ID, num_labels=len(labels), device=device).to(device) ckpt_path = "checkpoints/layoutlmv3_crf_passage.pth" os.makedirs("checkpoints", exist_ok=True) if os.path.exists(ckpt_path): print(f"⚠️ Starting fresh training. Old checkpoint {ckpt_path} may be incompatible with new label count.") optimizer = AdamW(model.parameters(), lr=args.lr) # 5. Training loop for epoch in range(args.epochs): print(f"\n--- START PHASE: EPOCH {epoch + 1}/{args.epochs} TRAINING ---") avg_loss = train_one_epoch(model, train_loader, optimizer, device) print(f"\n--- START PHASE: EPOCH {epoch + 1}/{args.epochs} EVALUATION ---") precision, recall, f1 = evaluate(model, val_loader, device, id2label) print( f"Epoch {epoch + 1}/{args.epochs} | Loss: {avg_loss:.4f} | P: {precision:.3f} R: {recall:.3f} F1: {f1:.3f}") torch.save(model.state_dict(), ckpt_path) print(f"💾 Model saved at {ckpt_path}") # ------------------------- # Step 7: Main Execution # ------------------------- if __name__ == "__main__": parser = argparse.ArgumentParser(description="LayoutLMv3 Fine-tuning and Inference Script.") parser.add_argument("--mode", type=str, required=True, choices=["train", "infer"], help="Select mode: 'train' or 'infer'") parser.add_argument("--input", type=str, help="Path to input file (Label Studio JSON for train, PDF for infer).") parser.add_argument("--batch_size", type=int, default=4) parser.add_argument("--epochs", type=int, default=5) parser.add_argument("--lr", type=float, default=5e-5) parser.add_argument("--max_len", type=int, default=512) args = parser.parse_args() if args.mode == "train": if not args.input: parser.error("--input is required for 'train' mode.") main(args)