Upload 4.2_add_span_rerank_branch_4.5's state dict

.gitattributes

@@ -93,3 +93,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 4_margin_loss_C_4.1/logs/4_margin_loss_C_4.1_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
 4.1_entities_no_ce_4.2/logs/4.1_entities_no_ce_4.2_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
 4.2_add_span_rerank_branch_4.3/logs/4.2_add_span_rerank_branch_4.3_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
+4.2_add_span_rerank_branch_4.5/logs/4.2_add_span_rerank_branch_4.5_log_plot.jpg filter=lfs diff=lfs merge=lfs -text

4.2_add_span_rerank_branch_4.5/4.2_add_span_rerank_branch_4.5.py

@@ -0,0 +1,2751 @@
+# %% [code]
+get_ipython().system('pip install evaluate seqeval underthesea positional-encodings[pytorch] pytorch-crf')
+
+# %% [code]
+import warnings
+warnings.filterwarnings('ignore')
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.utils.data import Dataset, TensorDataset, DataLoader
+import torch.nn.functional as F
+import albumentations as albu
+from transformers import AutoTokenizer, AutoModel
+import torch.distributed as dist
+from torch.nn.parallel import DistributedDataParallel as DDP
+from positional_encodings.torch_encodings import PositionalEncoding1D
+from torchcrf import CRF
+
+from sklearn.metrics import f1_score
+from sklearn.preprocessing import MinMaxScaler, StandardScaler
+from scipy.spatial.transform import Rotation as R
+from sklearn.model_selection import KFold, StratifiedGroupKFold, GroupKFold, StratifiedKFold
+from sklearn.metrics import precision_recall_fscore_support
+from timm.utils import ModelEmaV3
+import timm
+
+import os
+import gc
+import json
+from pathlib import Path
+import pickle
+from tqdm.auto import tqdm
+import copy
+import numpy as np
+import pandas as pd
+import polars as pl
+from PIL import Image
+import time
+from tqdm import tqdm
+from matplotlib import pyplot as plt
+import seaborn as sns
+from multiprocessing import Manager as MemoryManager
+from functools import lru_cache
+import shutil
+import glob
+import cv2
+import random
+import re
+import joblib
+import math
+from huggingface_hub import HfApi, snapshot_download
+import evaluate
+from underthesea import word_tokenize as vi_tokenize_tool
+import spacy
+en_tokenize_tool = spacy.load("en_core_web_sm")
+from collections import defaultdict, Counter
+
+# %% [code]
+# Global config
+SEEDS = [26092004]
+topk = 1
+nfolds = 5
+only_fold_idx = 0
+test_only = 0
+debug_only = 0
+
+# Config thư mục
+dataset = 'kltn/only_entities' # conll003, ontonotes, phoner, vietbio, vietmed, vimed, kltn/only_entities, kltn/raw
+root_dir = f'/kaggle/input/notebooks/sambui22022517/kltn-data/{dataset}' ## Thư mục chứa file train, val, test
+train_dir = f'{root_dir}'
+# val_dir = f'{root_dir}/val'
+test_dir = f'{root_dir}'
+
+# Config checkpoints
+
+# Config training
+epochs = 18 if not debug_only else 2
+batch_size = 32
+device = "cuda" if torch.cuda.is_available() else "cpu"
+# # Thêm biến toàn cục nào đó vào đây
+repo_name = 'SS3M/kltn-experiments'
+state_dict_save_name = "4.2_add_span_rerank_branch_4.5"
+checkpoints_dir = state_dict_save_name
+pretrained_dir = "/kaggle/working"
+os.makedirs(f'{checkpoints_dir}', exist_ok=True)
+
+backbone_model_name = "bert-base-uncased" if dataset in ["conll003", "ontonotes"] else "vinai/phobert-base"
+word_tokenize = lambda text: [token.text for token in en_tokenize_tool(text)] if dataset == dataset in ["conll003", "ontonotes"] else vi_tokenize_tool(text)
+max_len_dict = {
+    'kltn/raw': 256, 
+    'kltn/only_entities': 68, 
+    'conll003': 46, 
+    'ontonotes': 61, 
+    'phoner': 68, 
+    'vietbio': 125, 
+    'vietmed': 36, 
+    'vimed': 100, 
+}
+zero_entities_rate_dict = {
+    'kltn/raw': 1000, 
+    'kltn/only_entities': 0.2, 
+    'conll003': 1000, # mean keep all zero-entities samples
+    'ontonotes': 1000, 
+    'phoner': 1000, 
+    'vietbio': 1000, 
+    'vietmed': 1000, 
+    'vimed': 1000, 
+}
+
+max_len = max_len_dict[dataset]
+max_n_parts = 1
+max_span_len = 10
+zero_entities_rate = zero_entities_rate_dict[dataset]
+
+# Trainer
+trainer_params = {
+    "training_time": "00:11:30:00",
+    "eval_mode": "max",
+    "topk": topk,
+    "save_name": state_dict_save_name,
+    "save_best": True,
+    "save_last": True,
+    "device": device, 
+    "logging": True, 
+    "logging_file": True, 
+    "checkpoints_dir": checkpoints_dir, 
+    "early_stopping": 30, 
+    "eval_from_ratio": 0.4, 
+    "eval_every": 1, 
+    "schedule_in_step": False, 
+    "use_ema": True, 
+    "ema_from_ratio": 0.3, 
+    "ema_decay": 0.9995, 
+    "max_grad_norm": 200.0, 
+    "return_best": True, 
+    "return_last": True, 
+}
+
+# Memory
+train_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+}
+val_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+}
+
+# Data Loader
+def seed_worker(worker_id):
+    worker_seed = torch.initial_seed() % 2**32
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+    
+train_loader_params = {
+    'batch_size': batch_size, 
+    'shuffle': True,
+    'pin_memory':True, 
+    'num_workers': 2, 
+    'drop_last': False, 
+    'worker_init_fn': seed_worker, 
+    'persistent_workers': False, 
+}
+val_loader_params = {
+    'batch_size': batch_size, 
+    'shuffle': False,
+    'pin_memory':True, 
+    'num_workers': 1, 
+    'drop_last': False, 
+    'worker_init_fn': seed_worker, 
+    'persistent_workers': False, 
+}
+
+# Model
+model_params = {
+    'backbone_model_name': backbone_model_name, 
+    'max_r': 6, 
+}
+
+# Loss Func
+loss_func_params = {
+    'lambda_span': 1.0,
+    'lambda_margin': 1.0, 
+    'margin': 0.5, 
+}
+eval_func_params = {}
+
+# Optim
+optim_params = {
+    'name': 'AdamW', 
+    'lr': 1e-4,
+    'weight_decay': 1e-4,
+}
+scheduler_params = {
+    'name': 'CosineAnnealingLR',
+    'T_max': 20,  # Số epoch để hoàn thành một chu kỳ giảm LR
+    'eta_min': 1e-6  # Learning rate nhỏ nhất trong chu kỳ
+}
+
+# %% [code]
+def set_seed(seed=42):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)  # if using multi-GPU
+    torch.use_deterministic_algorithms(False)
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+    os.environ['PYTHONHASHSEED'] = str(seed)
+
+# %% [code]
+class CustomLoss(nn.Module):
+    def __init__(
+        self,
+        lambda_margin=1.0,
+        lambda_span=1.0,
+        margin=1.0
+    ):
+        super().__init__()
+
+        self.lambda_margin = lambda_margin
+        self.lambda_span = lambda_span
+        self.margin = margin
+
+    # =========================================================
+    # token margin
+    # =========================================================
+
+    def margin_loss(self, logits, labels):
+        """
+        logits: (N, C)
+        labels: (N,)
+        """
+
+        valid_mask = labels != -100
+
+        logits = logits[valid_mask]
+        labels = labels[valid_mask]
+
+        if len(labels) == 0:
+            return logits.new_tensor(0.0)
+
+        # positive logit
+        pos_logit = logits.gather(
+            dim=1,
+            index=labels.unsqueeze(-1)
+        ).squeeze(-1)
+
+        # hardest negative
+        neg_logits = logits.clone()
+
+        neg_logits.scatter_(
+            1,
+            labels.unsqueeze(-1),
+            float("-inf")
+        )
+
+        hardest_neg = neg_logits.max(dim=-1).values
+
+        # margin ranking
+        loss = F.relu(
+            self.margin - pos_logit + hardest_neg
+        )
+
+        return loss.mean()
+
+    # =========================================================
+    # span listwise margin
+    # =========================================================
+
+    def span_loss(self, span_scores, labels):
+        """
+        span_scores: (B, N, M)
+        labels:      (B, N, M)
+
+        labels:
+            -100 -> ignore
+             0   -> negative
+            >0   -> positive
+        """
+
+        device = span_scores.device
+
+        B, N, M = span_scores.shape
+        if N == 0 or M == 0:
+            return span_scores.new_tensor(0.0)
+
+        # =====================================================
+        # masks
+        # =====================================================
+
+        valid_mask = labels != -100
+        pos_mask = labels > 0
+        neg_mask = labels == 0
+
+        # =====================================================
+        # mỗi N phải có ít nhất 1 positive
+        # =====================================================
+
+        has_pos = pos_mask.any(dim=-1)
+        # (B, N)
+
+        # =====================================================
+        # masked scores
+        # =====================================================
+        pos_scores = span_scores.masked_fill(
+            ~pos_mask,
+            float("-inf")
+        )
+
+        neg_scores = span_scores.masked_fill(
+            ~neg_mask,
+            float("-inf")
+        )
+
+        # =====================================================
+        # hardest positive / hardest negative
+        # =====================================================
+
+        best_pos = pos_scores.max(dim=-1).values
+        # (B, N)
+
+        best_neg = neg_scores.max(dim=-1).values
+        # (B, N)
+
+        # =====================================================
+        # nếu không có negative
+        # =====================================================
+
+        no_neg = ~neg_mask.any(dim=-1)
+
+        best_neg = torch.where(
+            no_neg,
+            torch.zeros_like(best_neg),
+            best_neg
+        )
+
+        # =====================================================
+        # margin ranking
+        # =====================================================
+
+        loss = F.relu(
+            self.margin - best_pos + best_neg
+        )
+
+        # =====================================================
+        # chỉ tính những N có positive
+        # =====================================================
+
+        loss = loss[has_pos]
+
+        if loss.numel() == 0:
+            return span_scores.new_tensor(0.0)
+
+        return loss.mean()
+
+    # =========================================================
+    # forward
+    # =========================================================
+
+    def forward(
+        self,
+        start_logits, start_labels,
+        end_logits, end_labels,
+        span_scores, labels
+    ):
+
+        # =====================================================
+        # flatten token logits
+        # =====================================================
+
+        B, L, C = start_logits.shape
+
+        start_logits_flat = start_logits.view(B * L, C)
+        start_labels_flat = start_labels.view(-1)
+
+        end_logits_flat = end_logits.view(B * L, C)
+        end_labels_flat = end_labels.view(-1)
+
+        # =====================================================
+        # token margin
+        # =====================================================
+
+        start_margin = self.margin_loss(
+            start_logits_flat,
+            start_labels_flat
+        )
+
+        end_margin = self.margin_loss(
+            end_logits_flat,
+            end_labels_flat
+        )
+
+        token_margin_loss = (
+            start_margin + end_margin
+        )
+
+        # =====================================================
+        # span loss
+        # =====================================================
+
+        span_loss = self.span_loss(
+            span_scores,
+            labels
+        )
+
+        # =====================================================
+        # total
+        # =====================================================
+
+        total_loss = (
+            self.lambda_margin * token_margin_loss
+            +
+            self.lambda_span * span_loss
+        )
+
+        return {
+            "total": total_loss,
+            "token_margin_loss": token_margin_loss,
+            "span_loss": span_loss,
+            "start_margin": start_margin,
+            "end_margin": end_margin,
+        }
+
+# %% [code]
+## Viết eval_fn vào đây
+
+# Bỏ hết eval_fn và trọng số vào đây
+class CustomEvalFn(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def compute_f1(self, tp, fp, fn):
+        precision = tp / (tp + fp + 1e-8)
+        recall = tp / (tp + fn + 1e-8)
+        f1 = 2 * precision * recall / (precision + recall + 1e-8)
+        return precision, recall, f1
+
+    def forward(self, pred, gold):
+        pred_set = set(pred)
+        gold_set = set(gold)
+
+        tp = len(pred_set & gold_set)
+        fp = len(pred_set - gold_set)
+        fn = len(gold_set - pred_set)
+
+        precision, recall, f1 = self.compute_f1(tp, fp, fn)
+
+        return {
+                f"precision": precision,
+                f"recall": recall,
+                f"f1": f1,
+            }
+
+class SpanErrorAnalyzer:
+    def __init__(self, pad_token_id=0):
+        self.pad_token_id = pad_token_id
+
+    # ===== helper =====
+    def _to_set(self, data):
+        """
+        data: list of (b, tuple(ids))
+        -> dict[b] = set(tuple(ids))
+        """
+        res = defaultdict(set)
+        for b, ids in data:
+            ids = tuple([i for i in ids if i != self.pad_token_id])
+            if len(ids) > 0:
+                res[b].add(ids)
+        return res
+
+    def _iou(self, a, b):
+        """
+        a, b: tuple(ids)
+        """
+        set_a, set_b = set(a), set(b)
+        inter = len(set_a & set_b)
+        union = len(set_a | set_b)
+        if union == 0:
+            return 0.0
+        return inter / union
+
+    def _boundary_error(self, pred, gold):
+        """
+        đo lệch boundary dựa trên overlap prefix/suffix
+        """
+        # left match
+        left = 0
+        for i in range(min(len(pred), len(gold))):
+            if pred[i] == gold[i]:
+                left += 1
+            else:
+                break
+
+        # right match
+        right = 0
+        for i in range(1, min(len(pred), len(gold)) + 1):
+            if pred[-i] == gold[-i]:
+                right += 1
+            else:
+                break
+
+        return {
+            "left_match": left,
+            "right_match": right,
+            "pred_len": len(pred),
+            "gold_len": len(gold),
+        }
+
+    # ===== main =====
+    def analyze(self, preds, golds):
+        pred_map = self._to_set(preds)
+        gold_map = self._to_set(golds)
+
+        all_batches = set(pred_map.keys()) | set(gold_map.keys())
+
+        stats = Counter()
+
+        detailed_errors = []
+
+        for b in all_batches:
+            pset = pred_map.get(b, set())
+            gset = gold_map.get(b, set())
+
+            matched_gold = set()
+
+            # ===== check predictions =====
+            for p in pset:
+                if p in gset:
+                    stats["exact_match"] += 1
+                    matched_gold.add(p)
+                else:
+                    # tìm gold gần nhất
+                    best_iou = 0
+                    best_g = None
+
+                    for g in gset:
+                        iou = self._iou(p, g)
+                        if iou > best_iou:
+                            best_iou = iou
+                            best_g = g
+
+                    if best_iou > 0:
+                        stats["partial_match"] += 1
+
+                        boundary = self._boundary_error(p, best_g)
+
+                        detailed_errors.append({
+                            "type": "boundary_error",
+                            "batch": b,
+                            "pred": p,
+                            "gold": best_g,
+                            "iou": best_iou,
+                            **boundary
+                        })
+                    else:
+                        if b not in gold_map:
+                            stats["no_event_sample"] += 1
+                            err_type = "no_event_sample"
+                        else:
+                            stats["completely_wrong"] += 1
+                            err_type = "completely_wrong"
+                    
+                        detailed_errors.append({
+                            "type": err_type,
+                            "batch": b,
+                            "pred": p
+                        })
+
+            # ===== check missing =====
+            for g in gset:
+                if g not in matched_gold:
+                    # check if any pred overlaps
+                    overlap = any(self._iou(p, g) > 0 for p in pset)
+
+                    if overlap:
+                        stats["miss_with_overlap"] += 1
+                    else:
+                        stats["miss"] += 1
+
+                    detailed_errors.append({
+                        "type": "miss",
+                        "batch": b,
+                        "gold": g
+                    })
+
+        return {
+            "summary": {
+                "exact_match": (stats["exact_match"], stats["exact_match"] / len(preds)),
+                "partial_match": (stats["partial_match"], stats["partial_match"] / len(preds)),
+                "no_event_sample": (stats["no_event_sample"], stats["no_event_sample"] / len(preds)),
+                "completely_wrong": (stats["completely_wrong"], stats["completely_wrong"] / len(preds)),
+                "miss": (stats["miss"], stats["miss"] / len(golds)),
+                "miss_with_overlap": (stats["miss_with_overlap"], stats["miss_with_overlap"] / len(golds)),
+            },
+            "details": detailed_errors
+        }
+
+# %% [code]
+class DataParallelProxy(nn.DataParallel):
+
+    def __getattr__(self, name):
+        try:
+            return super().__getattr__(name)
+
+        except AttributeError:
+
+            attr = getattr(self.module, name)
+
+            if callable(attr):
+
+                def wrapper(*args, **kwargs):
+                    return self._parallel_apply_method(
+                        name,
+                        *args,
+                        **kwargs
+                    )
+
+                return wrapper
+
+            return attr
+
+    # =========================================================
+    # parallel custom method
+    # =========================================================
+
+    def _parallel_apply_method(self, method_name, *inputs, **kwargs):
+
+        if not self.device_ids:
+            return getattr(self.module, method_name)(*inputs, **kwargs)
+
+        inputs_scattered, kwargs_scattered = self.scatter(
+            inputs,
+            kwargs,
+            self.device_ids
+        )
+
+        replicas = self.replicate(
+            self.module,
+            self.device_ids[:len(inputs_scattered)]
+        )
+
+        outputs = self.parallel_apply(
+            [getattr(replica, method_name) for replica in replicas],
+            inputs_scattered,
+            kwargs_scattered
+        )
+
+        return self._custom_gather(
+            outputs,
+            self.output_device
+        )
+
+    # =========================================================
+    # OVERRIDE FORWARD GATHER
+    # =========================================================
+
+    def gather(self, outputs, output_device):
+
+        return self._custom_gather(
+            outputs,
+            output_device
+        )
+
+    # =========================================================
+    # recursive gather
+    # =========================================================
+
+    def _custom_gather(self, outputs, output_device):
+
+        first = outputs[0]
+
+        # =====================================================
+        # tensor
+        # =====================================================
+
+        if torch.is_tensor(first):
+
+            return self._gather_tensor(
+                outputs,
+                output_device
+            )
+
+        # =====================================================
+        # tuple
+        # =====================================================
+
+        if isinstance(first, tuple):
+
+            return tuple(
+                self._custom_gather(
+                    list(items),
+                    output_device
+                )
+                for items in zip(*outputs)
+            )
+
+        # =====================================================
+        # list
+        # =====================================================
+
+        if isinstance(first, list):
+
+            # list[tensor]
+            if len(first) > 0 and torch.is_tensor(first[0]):
+
+                return self._gather_tensor_list(
+                    outputs,
+                    output_device
+                )
+
+            merged = []
+
+            for out in outputs:
+                merged.extend(out)
+
+            return merged
+
+        # =====================================================
+        # dict
+        # =====================================================
+
+        if isinstance(first, dict):
+
+            return {
+                k: self._custom_gather(
+                    [o[k] for o in outputs],
+                    output_device
+                )
+                for k in first.keys()
+            }
+
+        # =====================================================
+        # fallback
+        # =====================================================
+
+        return outputs
+
+    # =========================================================
+    # gather tensor with auto pad
+    # =========================================================
+
+    def _gather_tensor(self, tensors, output_device):
+
+        # move same device first
+        tensors = [
+            t.to(output_device)
+            for t in tensors
+        ]
+
+        # =====================================================
+        # fast path
+        # =====================================================
+
+        try:
+            return torch.cat(tensors, dim=0)
+
+        except RuntimeError:
+            pass
+
+        # =====================================================
+        # auto max shape
+        # =====================================================
+
+        max_shape = list(tensors[0].shape)
+
+        for t in tensors[1:]:
+
+            for d in range(len(max_shape)):
+
+                max_shape[d] = max(
+                    max_shape[d],
+                    t.shape[d]
+                )
+
+        # =====================================================
+        # pad tensors
+        # =====================================================
+
+        padded = []
+
+        for t in tensors:
+
+            pad = []
+
+            # reverse order for F.pad
+            for d in reversed(range(len(max_shape))):
+
+                # never pad batch dim
+                if d == 0:
+                    pad.extend([0, 0])
+                    continue
+
+                diff = max_shape[d] - t.shape[d]
+
+                pad.extend([0, diff])
+
+            t = F.pad(t, pad)
+
+            padded.append(t)
+
+        return torch.cat(padded, dim=0)
+
+    # =========================================================
+    # list[tensor]
+    # =========================================================
+
+    def _gather_tensor_list(self, outputs, output_device):
+
+        merged = []
+
+        for out in outputs:
+            merged.extend(out)
+
+        return self._gather_tensor(
+            merged,
+            output_device
+        )
+
+# %% [code]
+## Viết cấu trúc model vào đây
+def extract_spans_and_labels(start_logits, end_logits):
+    """
+    Args:
+        start_logits: Tensor (B, L, C)
+        end_logits:   Tensor (B, L, C)
+
+    Returns:
+        spans:  Tensor (B, N, 2)
+                padding = (0, 0)
+
+        labels: Tensor (B, N)
+                padding = -100
+
+        Nếu không extract được span nào:
+            spans.shape  = (B, 0, 2)
+            labels.shape = (B, 0)
+    """
+
+    start_labels = start_logits.argmax(dim=-1)  # (B, L)
+    end_labels = end_logits.argmax(dim=-1)      # (B, L)
+
+    B, L = start_labels.shape
+
+    batch_spans = []
+    batch_labels = []
+
+    max_n = 0
+
+    for bidx in range(B):
+
+        used_start = set()
+        used_end = set()
+
+        spans = []
+        labels = []
+
+        for s in range(L):
+
+            s_label = start_labels[bidx, s].item()
+
+            # bỏ qua O
+            if s_label == 0:
+                continue
+
+            if s in used_start:
+                continue
+
+            nearest_e = None
+
+            # tìm end gần nhất cùng class
+            for e in range(s, L):
+
+                if e in used_end:
+                    continue
+
+                e_label = end_labels[bidx, e].item()
+
+                if e_label == s_label:
+                    nearest_e = e
+                    break
+
+            if nearest_e is None:
+                continue
+
+            used_start.add(s)
+            used_end.add(nearest_e)
+
+            spans.append([s, nearest_e])
+            labels.append(s_label)
+
+        batch_spans.append(spans)
+        batch_labels.append(labels)
+
+        max_n = max(max_n, len(spans))
+
+    # =========================================================
+    # không extract được gì
+    # =========================================================
+
+    if max_n == 0:
+
+        spans = torch.empty(
+            (B, 0, 2),
+            dtype=torch.long,
+            device=start_logits.device
+        )
+
+        labels = torch.empty(
+            (B, 0),
+            dtype=torch.long,
+            device=start_logits.device
+        )
+
+        return spans, labels
+
+    # =========================================================
+    # padding
+    # =========================================================
+
+    padded_spans = []
+    padded_labels = []
+
+    for spans, labels in zip(batch_spans, batch_labels):
+
+        pad_n = max_n - len(spans)
+
+        spans = spans + [[0, 0]] * pad_n
+        labels = labels + [-100] * pad_n
+
+        padded_spans.append(spans)
+        padded_labels.append(labels)
+
+    spans = torch.tensor(
+        padded_spans,
+        dtype=torch.long,
+        device=start_logits.device
+    )  # (B, N, 2)
+
+    labels = torch.tensor(
+        padded_labels,
+        dtype=torch.long,
+        device=start_logits.device
+    )  # (B, N)
+
+    return spans, labels
+
+def expand_spans(spans, r, attention_mask):
+    """
+    Args:
+        spans: Tensor (B, N, 2)
+        r:     Tensor (B, N)
+        attention_mask: Tensor (B, L)
+
+    Returns:
+        expanded_spans: Tensor (B, N, M, 2)
+
+        Với:
+            M = (2*max_r + 1)^2
+
+        padding = (0, 0)
+
+        Nếu N = 0:
+            return shape (B, 0, 0, 2)
+    """
+
+    device = spans.device
+
+    B, N, _ = spans.shape
+
+    # =========================================================
+    # empty case
+    # =========================================================
+
+    if N == 0:
+        return torch.empty(
+            (B, 0, 0, 2),
+            dtype=torch.long,
+            device=device
+        )
+
+    # =========================================================
+    # max radius
+    # =========================================================
+
+    max_r = r.max().item()
+
+    # =========================================================
+    # create all (ds, de)
+    # =========================================================
+
+    shifts = torch.arange(
+        -max_r,
+        max_r + 1,
+        device=device
+    )
+
+    ds_grid, de_grid = torch.meshgrid(
+        shifts,
+        shifts,
+        indexing='ij'
+    )
+
+    ds = ds_grid.reshape(-1)  # (M,)
+    de = de_grid.reshape(-1)  # (M,)
+
+    M = ds.size(0)
+
+    # =========================================================
+    # original spans
+    # =========================================================
+
+    spans = spans.unsqueeze(2)  # (B, N, 1, 2)
+
+    start = spans[..., 0]  # (B, N, 1)
+    end   = spans[..., 1]  # (B, N, 1)
+
+    # =========================================================
+    # expand independently
+    # =========================================================
+
+    expanded_s = start + ds.view(1, 1, M)
+    expanded_e = end   + de.view(1, 1, M)
+
+    expanded = torch.stack(
+        [expanded_s, expanded_e],
+        dim=-1
+    )  # (B, N, M, 2)
+
+    # =========================================================
+    # valid shift range
+    # =========================================================
+
+    valid_shift = (
+        (ds.abs().view(1, 1, M) <= r.unsqueeze(-1))
+        &
+        (de.abs().view(1, 1, M) <= r.unsqueeze(-1))
+    )
+
+    # =========================================================
+    # non-padding spans
+    # =========================================================
+
+    non_pad_span = (
+        spans.squeeze(2).sum(dim=-1) != 0
+    )  # (B, N)
+
+    # =========================================================
+    # boundary check
+    # =========================================================
+
+    valid_length = attention_mask.sum(dim=-1)
+    valid_length = valid_length.view(B, 1, 1)
+
+    valid_boundary = (
+        (expanded_s > 0)
+        &
+        (expanded_e < valid_length)
+        &
+        (expanded_s <= expanded_e)
+    )
+
+    # =========================================================
+    # final valid mask
+    # =========================================================
+
+    valid_mask = (
+        valid_shift
+        &
+        non_pad_span.unsqueeze(-1)
+        &
+        valid_boundary
+    )
+
+    # =========================================================
+    # flatten valid spans
+    # =========================================================
+
+    expanded_flat = expanded.reshape(B, N, M, 2)
+    valid_flat = valid_mask.reshape(B, N, M)
+
+    # =========================================================
+    # số valid max toàn batch
+    # =========================================================
+
+    valid_counts = valid_flat.sum(dim=-1)  # (B, N)
+
+    M_max = valid_counts.max().item()
+
+    # =========================================================
+    # allocate output
+    # =========================================================
+
+    filtered_expanded = torch.zeros(
+        (B, N, M_max, 2),
+        dtype=expanded.dtype,
+        device=device
+    )
+
+    # =========================================================
+    # gather only valid spans
+    # =========================================================
+
+    for b in range(B):
+        for n in range(N):
+
+            cur_valid = valid_flat[b, n]  # (M,)
+
+            valid_spans = expanded_flat[b, n][cur_valid]
+
+            k = valid_spans.size(0)
+
+            if k > 0:
+                filtered_expanded[b, n, :k] = valid_spans
+
+    return filtered_expanded
+
+def get_span_reprs(hidden, spans):
+    """
+    Args:
+        hidden: (B, L, H)
+        spans:  (B, N, M, 2)
+
+    Return:
+        span_repr: (B, N, M, 4H)
+
+        Nếu N = 0:
+            return shape (B, 0, 0, 4H)
+    """
+
+    B, N, M, _ = spans.shape
+    H = hidden.size(-1)
+
+    # =========================================================
+    # empty case
+    # =========================================================
+
+    if N == 0:
+        return torch.empty(
+            (B, 0, 0, 4 * H),
+            dtype=hidden.dtype,
+            device=hidden.device
+        )
+
+    batch_idx = torch.arange(
+        B,
+        device=hidden.device
+    ).view(B, 1, 1)
+
+    start_idx = spans[..., 0]  # (B, N, M)
+    end_idx   = spans[..., 1]  # (B, N, M)
+
+    # =========================================================
+    # gather hidden
+    # =========================================================
+
+    start_h = hidden[batch_idx, start_idx]
+    end_h   = hidden[batch_idx, end_idx]
+
+    # (B, N, M, H)
+
+    # =========================================================
+    # span representation
+    # =========================================================
+
+    span_repr = torch.cat(
+        [
+            start_h,
+            end_h,
+            end_h - start_h,
+            end_h * start_h
+        ],
+        dim=-1
+    )  # (B, N, M, 4H)
+
+    return span_repr
+    
+class MLP(nn.Module):
+    def __init__(self, in_size, hid_size, out_size):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(in_size, hid_size),
+            nn.ReLU(),
+            nn.Linear(hid_size, out_size)
+        )
+
+    def forward(self, x):
+        return self.mlp(x)
+
+class IEModel(nn.Module):
+    def __init__(self, backbone_model_name, num_labels, max_r):
+        super().__init__()
+        self.max_r = max_r
+        
+        self.encoder = AutoModel.from_pretrained(backbone_model_name)
+        hidden_size = self.encoder.config.hidden_size
+
+        self.start_classifier = MLP(hidden_size, hidden_size, num_labels)
+        self.end_classifier = MLP(hidden_size, hidden_size, num_labels)
+
+        self.span_scorer = MLP(4*hidden_size, hidden_size, 1)
+
+    def encode(self, input_ids, attention_mask):
+        B, n_parts, L = input_ids.shape
+        input_ids = input_ids.view(-1, L)
+        attention_mask = attention_mask.view(-1, L)
+        
+        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
+        hidden_states = outputs.last_hidden_state # B * n_parts, L, H
+        
+        hidden_states = hidden_states.view(B, n_parts, L, -1).reshape(B, n_parts*L, -1) # B, L, H
+        attention_mask = attention_mask.view(B, n_parts, L).reshape(B, n_parts*L) # B, L
+        return hidden_states, attention_mask
+
+    def get_logits(self, hidden_states):
+        start_logits = self.start_classifier(hidden_states)      # B, N, classes
+        end_logits = self.end_classifier(hidden_states)      # B, N, classes
+        return start_logits, end_logits
+
+    def get_scores(self, expanded_span_reprs):
+        return self.span_scorer(expanded_span_reprs).squeeze(-1)
+
+    def forward(self, input_ids, attention_mask, spans=None):
+        hidden_states, attention_mask = self.encode(input_ids, attention_mask)
+        start_logits, end_logits = self.get_logits(hidden_states)
+
+        if spans is None:
+            spans, _ = extract_spans_and_labels(start_logits, end_logits) # (B, N, 2)
+        B, N, _ = spans.shape
+        r = torch.randint(1, self.max_r+1, (B, N), device=spans.device)
+        expanded_spans = expand_spans(spans, r, attention_mask) # (B, N, M, 2)
+
+        expanded_span_reprs = get_span_reprs(hidden_states, expanded_spans) # (B, N, M, 4H)
+        expanded_span_scores = self.get_scores(expanded_span_reprs) # (B, N, M)
+        
+        return start_logits, end_logits, expanded_spans, expanded_span_scores
+
+def test_model():
+    model = DataParallelProxy(IEModel(backbone_model_name, 7, 6)).to(device)
+    model.eval()
+    total_params = sum(p.numel() for p in model.parameters())
+    print(f"Total params: {total_params:,}")
+
+    vocab_size = model.module.encoder.config.vocab_size
+    max_len = model.module.encoder.config.max_position_embeddings
+
+    bz = 32
+    i = torch.randint(0, vocab_size, (bz, 5, 10)).to(device)
+    a = torch.ones(bz, 5, 10).to(device)
+    g = torch.ones(bz, 3, 2, dtype=torch.long).to(device)
+
+    with torch.no_grad():
+        r = model(i, a)
+
+    if type(r) == tuple:
+        print([r[i].shape if type(r[i]) == type(torch.Tensor()) else len(r[i]) for i in range(len(r))])
+    else:
+        print(r.shape)
+
+test_model()
+
+# %% [code]
+def configure_optimizers(network, optim_params, scheduler_params):
+    try:
+        optim_params = copy.copy(optim_params)
+        scheduler_params = copy.copy(scheduler_params)
+        
+        optim_name = optim_params.pop('name')
+        scheduler_name = scheduler_params.pop('name')
+    
+        optimizer_cls = globals().get(optim_name) or getattr(optim, optim_name, None)
+        scheduler_cls = globals().get(scheduler_name) or getattr(optim.lr_scheduler, scheduler_name, None)
+    
+        if optimizer_cls is None:
+            raise ValueError(f"Optimizer '{optim_name}' is not available!")
+    
+        optimizer = optimizer_cls(network.parameters(), **optim_params)
+        
+        scheduler = None
+        if scheduler_params and scheduler_cls:  # Chỉ tạo scheduler nếu có tham số
+            scheduler = scheduler_cls(optimizer, **scheduler_params)
+    
+        return optimizer, scheduler
+
+    except KeyError as e:
+        raise ValueError(f"Missing {e} in config!!")
+            
+def freeze(self, model):
+    model.eval()
+    for param in model.parameters():
+        param.requires_grad = False
+
+def unfreeze(self, model):
+    model.train()
+    for param in model.parameters():
+        param.requires_grad = True
+
+def reduce_batch_size(loader, ratio=0.5):
+    new_bs = max(1, int(loader.batch_size * ratio))
+
+    shuffle = isinstance(loader.sampler, RandomSampler)
+
+    new_loader = DataLoader(
+        dataset=loader.dataset,
+        batch_size=new_bs,
+        shuffle=shuffle,
+        sampler=None if shuffle else loader.sampler,
+        num_workers=loader.num_workers,
+        collate_fn=loader.collate_fn,
+        pin_memory=loader.pin_memory,
+        drop_last=loader.drop_last,
+        timeout=loader.timeout,
+        worker_init_fn=loader.worker_init_fn,
+        multiprocessing_context=loader.multiprocessing_context,
+        generator=loader.generator,
+        prefetch_factor=loader.prefetch_factor if loader.num_workers > 0 else None,
+        persistent_workers=loader.persistent_workers,
+        pin_memory_device=loader.pin_memory_device
+    )
+
+    return new_loader
+
+def list_to_tuple(x):
+    if isinstance(x, (list, tuple)):
+        return tuple(list_to_tuple(i) for i in x)
+    return x
+
+def fmt(x):
+    if isinstance(x, float):
+        return round(x, 5)
+    if isinstance(x, dict):
+        return {k: fmt(v) for k, v in x.items()}
+    if isinstance(x, list):
+        return [fmt(v) for v in x]
+    return x
+    
+class ModelEmaV3Proxy(ModelEmaV3):
+    def __getattr__(self, name):
+        try:
+            return super().__getattr__(name)
+        except AttributeError:
+            return getattr(self.module, name)
+
+def align(spans, gold_spans, gold_labels):
+    """
+    Args:
+        spans:       (B, N1, M, 2)
+        gold_spans:  (B, N2, 2)
+        gold_labels: (B, N2)
+
+    Returns:
+        labels: (B, N1, M)
+
+        matched -> gold_labels
+        padding (0,0) -> -100
+        unmatched -> 0
+    """
+
+    device = spans.device
+
+    B, N1, M, _ = spans.shape
+    N2 = gold_spans.shape[1]
+
+    # =========================================================
+    # compare spans
+    # =========================================================
+
+    pred = spans.unsqueeze(3)
+    # (B, N1, M, 1, 2)
+
+    gold = gold_spans.unsqueeze(1).unsqueeze(1)
+    # (B, 1, 1, N2, 2)
+
+    matched = (pred == gold).all(dim=-1)
+    # (B, N1, M, N2)
+
+    # =========================================================
+    # default labels = 0
+    # =========================================================
+
+    labels = torch.zeros(
+        (B, N1, M),
+        dtype=gold_labels.dtype,
+        device=device
+    )
+
+    # =========================================================
+    # assign matched labels
+    # =========================================================
+
+    has_match = matched.any(dim=-1)
+    # (B, N1, M)
+
+    matched_idx = matched.float().argmax(dim=-1)
+    # (B, N1, M)
+
+    expanded_gold_labels = gold_labels.unsqueeze(1).unsqueeze(1).expand(
+        B, N1, M, N2
+    )
+
+    gathered_labels = torch.gather(
+        expanded_gold_labels,
+        dim=-1,
+        index=matched_idx.unsqueeze(-1)
+    ).squeeze(-1)
+
+    labels = torch.where(
+        has_match,
+        gathered_labels,
+        labels
+    )
+
+    # =========================================================
+    # padding span -> -100
+    # =========================================================
+
+    is_padding = (spans.sum(dim=-1) == 0)
+
+    labels = torch.where(
+        is_padding,
+        torch.full_like(labels, -100),
+        labels
+    )
+
+    return labels
+
+def select_best_spans(spans, span_scores):
+    """
+    Args:
+        spans:       (B, N, M, 2)
+        span_scores: (B, N, M)
+
+    Returns:
+        best_spans:  (B, N, 2)
+    """
+
+    # =========================================================
+    # padding spans -> -inf
+    # =========================================================
+
+    is_padding = (spans.sum(dim=-1) == 0)
+    # (B, N, M)
+
+    masked_scores = span_scores.masked_fill(
+        is_padding,
+        float("-inf")
+    )
+
+    # =========================================================
+    # best index theo chiều M
+    # =========================================================
+
+    best_idx = masked_scores.argmax(dim=-1)
+    # (B, N)
+
+    # =========================================================
+    # gather spans
+    # =========================================================
+
+    best_spans = torch.gather(
+        spans,
+        dim=2,
+        index=best_idx.unsqueeze(-1).unsqueeze(-1).expand(
+            -1, -1, 1, 2
+        )
+    ).squeeze(2)
+
+    # =========================================================
+    # nếu toàn bộ M đều là padding
+    # -> output = (0,0)
+    # =========================================================
+
+    all_padding = is_padding.all(dim=-1)
+    # (B, N)
+
+    best_spans = torch.where(
+        all_padding.unsqueeze(-1),
+        torch.zeros_like(best_spans),
+        best_spans
+    )
+
+    return best_spans
+
+def extract_entities(input_ids, spans, labels, id2label):
+    """
+    Args:
+        input_ids: (B, L)
+        spans:     (B, N, 2)
+        labels:    (B, N)
+
+    Returns:
+        List[
+            (
+                bidx,
+                (
+                    tuple(token_ids),
+                    label_name
+                )
+            )
+        ]
+    """
+
+    B, N, _ = spans.shape
+
+    results = []
+
+    for bidx in range(B):
+
+        for n in range(N):
+
+            lb = labels[bidx, n].item()
+
+            # ignore padding / negative
+            if lb <= 0:
+                continue
+
+            s, e = spans[bidx, n].tolist()
+
+            # skip padding span
+            if s == 0 and e == 0:
+                continue
+
+            token_ids = tuple(
+                input_ids[bidx, s:e + 1].tolist()
+            )
+
+            results.append(
+                (
+                    bidx,
+                    (
+                        token_ids,
+                        id2label[lb]
+                    )
+                )
+            )
+
+    return results
+    
+class Trainer:
+    def __init__(
+        self, training_time="00:11:30:00", eval_mode="max", topk=1, save_name="network", save_best=True, save_last=False, max_grad_norm=200.0, 
+        logging=0, logging_file=False, checkpoints_dir="", early_stopping=False, eval_from_ratio=-1, eval_every=1, device='cpu', 
+        schedule_in_step=True, use_ema=True, ema_from_ratio=-1, ema_decay=0.999, return_best=True, return_last=True
+    ):
+        self.ema_net = None
+        
+        self.training_time = self._time_str_to_seconds(training_time)
+        self.mode = eval_mode
+        self.topk = topk
+        self.device = device
+        self.logging = logging if logging < epochs else 1
+        self.logging_file = logging_file
+        self.checkpoints_dir = checkpoints_dir
+        self.early_stopping = early_stopping
+        self.eval_from_ratio = eval_from_ratio
+        self.eval_every = eval_every
+        self.save_name = save_name
+        self.save_best = save_best
+        self.save_last = save_last
+        self.return_best = return_best
+        self.return_last = return_last
+        self.max_grad_norm = max_grad_norm
+        self.schedule_in_step = schedule_in_step
+        self.use_ema = use_ema
+        self.ema_from_ratio = ema_from_ratio
+        self.ema_decay = ema_decay
+        
+        self.best_stage = [[float('-inf') if self.mode == 'max' else float('inf'), None, None]]
+        self.grad_scaler = torch.amp.GradScaler(self.device, init_scale=1024.0)
+
+    def fit(self, network, optimizer, scheduler, loss_fn, epochs, train_loader, val_loader=None, eval_fn=None, start_epoch=1, start_training_time=None, id2label=None):
+        if eval_fn is None:
+            if self.mode == "max":
+                eval_fn = lambda *x: -loss_fn(*x)
+            else:
+                eval_fn = lambda *x: loss_fn(*x)
+
+        if torch.cuda.device_count() > 1:
+            network = DataParallelProxy(network)
+        network = network.to(self.device)
+            
+        if not start_training_time:
+            start_training_time = time.time()
+            
+        start_ema = int(epochs * self.ema_from_ratio)
+        start_eval = int(epochs * self.eval_from_ratio)
+        
+        if val_loader is None:
+            print(f'[Trainer CallBack] 📢 Không có Val Set, không thể đánh giá và Early Stopping!')
+        else:
+            model_to_use_str = 'mô hình EMA' if self.use_ema else 'mô hình gốc'
+            start_model_update_str = f'Bắt đầu cập nhật EMA từ epoch {start_epoch + start_ema}!' if self.use_ema else ''
+            print(f'[Trainer CallBack] 📢 Đánh giá bằng {model_to_use_str} từ epoch {start_epoch + start_eval}!', start_model_update_str)
+
+        training_log = {}
+        for epoch in range(start_epoch, epochs+start_epoch):
+            if self.use_ema and self.ema_net is None and epoch - start_epoch >= start_ema:
+                self.ema_net = ModelEmaV3Proxy(network, self.ema_decay, device=self.device)
+                
+            try:
+                teaching_rate = math.cos(math.pi / 2 * epoch / epochs)
+                train_loss_epoch, train_loss_epoch_dict = self._train_epoch(network, train_loader, optimizer, scheduler, loss_fn, teaching_rate)
+                logging_dict = {'lr': [group['lr'] for group in optimizer.param_groups], 'train_loss': train_loss_epoch}
+                logging_dict.update(train_loss_epoch_dict)
+                    
+                if val_loader is not None and epoch - start_epoch >= start_eval and (epoch - start_epoch - start_eval) % self.eval_every == 0:
+                    eval_net = self.ema_net.module if (self.use_ema and self.ema_net is not None) else network
+                    
+                    val_score, val_score_dict, _ = self._eval_epoch(eval_net, val_loader, eval_fn, id2label)
+                    update = self._update_best_network(eval_net, val_score, epoch)
+                    logging_dict.update({'val_score': val_score, 'best_score': self.best_stage[0][0], 'new_best_model': update})
+                    logging_dict.update(val_score_dict)
+                if not self.schedule_in_step and scheduler:
+                    scheduler.step()
+                    
+            except RuntimeError as e:
+                if "out of memory" in str(e).lower():
+                    print(f"[Trainer CallBack] ⚠️ Epoch {epoch}/{epochs}: CUDA Out of Memory! Clearing GPU cache...")
+                    torch.cuda.empty_cache()
+                    gc.collect()
+                    if torch.cuda.is_available():
+                        torch.cuda.synchronize()
+                    print(f"[Trainer CallBack] ✅ Epoch {epoch}/{epochs}: GPU memory cleared")
+                    
+                    train_loader = reduce_batch_size(train_loader, ratio=0.5)
+                    if val_loader is not None:
+                        val_loader = reduce_batch_size(val_loader, ratio=0.5)
+                        
+                    logging_dict = {'lr': [group['lr'] for group in optimizer.param_groups], 'train_loss': float('inf')}
+                else:
+                    raise
+                    
+            training_log[epoch] = logging_dict                    
+            if self.is_early_stopping(epoch):
+                print(f'[Trainer CallBack] 📢 Epoch {epoch}/{epochs}: Detect Overfitting! Breaking Training Process...')
+                break
+            if self.logging:
+                if epoch % self.logging == 0:
+                    print(f'[Trainer CallBack] 📢 Epoch {epoch}/{epochs}:', fmt(logging_dict))
+            else:
+                print(f'{epoch}...', end=' ')
+                    
+            if self._at_time_limit(start_training_time):
+                print(f'[Trainer CallBack] ⚠️ Epoch {epoch}/{epochs}: Thời gian training giới hạn là {self.training_time}, hết giờ tại epoch {epoch}/{epochs}')
+                break
+                
+        if self.logging_file:
+            os.makedirs(f'{self.checkpoints_dir}/logs', exist_ok=True)
+            with open(f"{self.checkpoints_dir}/logs/{self.save_name}_logging.json", "a", encoding="utf-8") as f:
+                f.write(json.dumps(training_log))
+                
+        if self.use_ema and self.ema_net is not None:
+            self._save_state_dict(self.ema_net.module)
+        else:
+            self._save_state_dict(network)
+        print(f'[Trainer CallBack] 📢 Kết thúc training.\n')
+
+        best_model, last_model = None, None
+        eval_net = self.ema_net.module if (self.use_ema and self.ema_net is not None) else network
+        if self.return_best :
+            best_model = self.best_stage[0][2] if self.best_stage[0][2] is not None else eval_net.state_dict()
+            best_model = {k.replace("module.", ""): v.detach().cpu().clone() for k, v in best_model.items()}
+        if self.return_last:
+            last_model = eval_net.state_dict()
+            last_model = {k.replace("module.", ""): v.detach().cpu().clone() for k, v in last_model.items()}
+
+        del network
+        torch.cuda.empty_cache()
+        gc.collect()
+        return training_log, best_model, last_model
+
+    def _time_str_to_seconds(self, time_str):
+        days, hours, minutes, seconds = map(int, time_str.split(":"))
+        return days * 86400 + hours * 3600 + minutes * 60 + seconds
+
+    def _update_best_network(self, network, val_score, epoch):
+        topk = max(1, self.topk)
+        self.best_stage.append([val_score, epoch, {k: v.detach().cpu().clone() for k, v in network.state_dict().items()}])
+        self.best_stage = sorted(self.best_stage, reverse=(self.mode == 'max'), key=lambda x: x[0])[:topk]
+        if val_score in [x[0] for x in self.best_stage]:
+            return True
+        return False
+
+    def is_early_stopping(self, epoch):
+        if self.best_stage[0][1] is None:
+            return False
+        if not self.early_stopping:
+            return False
+        return epoch - self.best_stage[0][1] >= self.early_stopping
+
+    def _at_time_limit(self, start_training_time):
+        return time.time() - start_training_time >= self.training_time
+
+    def _save_state_dict(self, network):
+        if self.topk <= 0:
+            return
+
+        if self.save_best:
+            for r in range(self.topk):
+                os.makedirs(f'{self.checkpoints_dir}/r{r+1}s', exist_ok=True)
+                
+            for rank, (score, epoch, state_dict) in enumerate(self.best_stage):
+                if state_dict is None:
+                    continue
+                state_dict = {k.replace("module.", ""): v.detach().cpu().clone() for k, v in state_dict.items()}
+                torch.save(state_dict, f'{self.checkpoints_dir}/r{rank+1}s/{self.save_name}_r{rank+1}_vs{score:.5f}_{"ema" if self.ema_net is not None else ""}.pth')
+        if self.save_last:
+            os.makedirs(f'{self.checkpoints_dir}/lasts', exist_ok=True)
+            state_dict = {k.replace("module.", ""): v.detach().cpu().clone() for k, v in network.state_dict().items()}
+            torch.save(state_dict, f'{self.checkpoints_dir}/lasts/{self.save_name}_last_{"ema" if self.ema_net is not None else ""}.pth')
+        
+    def _train_epoch(self, network, train_loader, optimizer, scheduler, loss_fn, teaching_rate):
+        network.train()
+        total_loss = 0
+        total_loss_dict = {}
+        for batch_idx, batch in enumerate(train_loader):
+            optimizer.zero_grad()
+            with torch.autocast(device_type=self.device, dtype=torch.float16):
+                loss, loss_dict = self._cal_loss(network, batch, batch_idx, loss_fn, teaching_rate)
+
+                for k, v in loss_dict.items():
+                    t = total_loss_dict.get(k, 0)
+                    total_loss_dict[k] = t + v
+            self.grad_scaler.scale(loss).backward()
+            self.grad_scaler.unscale_(optimizer)
+            grad_norm = nn.utils.clip_grad_norm_(network.parameters(), self.max_grad_norm)
+            # print(grad_norm) # Bỏ cmt dòng này để biết nên chọn max_grad_norm bằng bao nhiêu...
+            self.grad_scaler.step(optimizer)
+            self.grad_scaler.update()
+            if self.schedule_in_step and scheduler:
+                scheduler.step()
+            if self.use_ema and self.ema_net is not None:
+                self.ema_net.update(network)
+            total_loss += loss
+        return (total_loss / len(train_loader)).item(), {k: v.item() / len(train_loader) for k, v in total_loss_dict.items()}
+
+    def _eval_epoch(self, network, val_loader, eval_fn, id2label):
+        network.eval()
+        total_score = 0.0
+        total_score_dict = {}
+        object_lists = None  # sẽ init sau
+    
+        with torch.no_grad():
+            for batch_idx, batch in enumerate(val_loader):
+                score, score_dict, objects = self._cal_val_score(network, batch, batch_idx, eval_fn, id2label)
+                total_score += score
+
+                for k, v in score_dict.items():
+                    t = total_score_dict.get(k, 0)
+                    total_score_dict[k] = t + v
+    
+                if objects:
+                    if object_lists is None:
+                        object_lists = [[] for _ in range(len(objects))]
+    
+                    for i, obj in enumerate(objects):
+                        object_lists[i].append(obj.detach())
+    
+        if object_lists is not None:
+            object_arrays = [
+                torch.concat(obj_list, dim=0).cpu().numpy()
+                for obj_list in object_lists
+            ]
+        else:
+            object_arrays = []
+    
+        return total_score / len(val_loader), {k: v / len(val_loader) for k, v in total_score_dict.items()}, object_arrays
+
+    def _cal_loss(self, network, batch, batch_idx, loss_fn, teaching_rate):
+        # Bạn cần override _cal_loss để tính loss
+        input_ids = batch['input_ids'].to(self.device)
+        attention_mask = batch['attention_mask'].to(self.device)
+        gold_spans = batch['gold_spans'].to(self.device)
+        gold_labels = batch['gold_labels'].to(self.device)
+        start_labels = batch['start_labels'].to(self.device)
+        end_labels = batch['end_labels'].to(self.device)
+
+        choice = random.random()
+        if choice < teaching_rate:
+            start_logits, end_logits, spans, span_scores = network(input_ids, attention_mask, gold_spans)
+        else:
+            start_logits, end_logits, spans, span_scores = network(input_ids, attention_mask)
+
+        labels = align(spans, gold_spans, gold_labels)
+        
+        loss_dict = loss_fn(
+            start_logits, start_labels, 
+            end_logits, end_labels, 
+            span_scores, labels
+        )
+        return loss_dict['total'], loss_dict
+
+    def _cal_val_score(self, network, batch, batch_idx, eval_fn, id2label):
+        # Bạn cần override _cal_val_score để tính val score, list bên cạnh là để trả về y hay pred gì đó (nếu cần)
+        input_ids = batch['input_ids'].to(self.device)
+        attention_mask = batch['attention_mask'].to(self.device)
+        gold_entities = batch['gold_entities']
+
+        B, _, _ = input_ids.shape
+
+        start_logits, end_logits, spans, span_scores = network(input_ids, attention_mask)
+        _, labels = extract_spans_and_labels(start_logits, end_logits)
+        spans = select_best_spans(spans, span_scores)
+        
+        pred_ids = extract_entities(input_ids.reshape(B, -1), spans, labels, id2label)
+        pred_ids = list_to_tuple(pred_ids)
+            
+        gold_ids = list_to_tuple(gold_entities)
+        
+        score_dict = eval_fn(pred_ids, gold_ids)
+        return score_dict['f1'], score_dict, []
+
+# %% [code]
+class PhoBERTSpanAligner:
+    def __init__(self, tokenizer, max_len):
+        self.tokenizer = tokenizer
+        self.max_len = max_len
+
+    # ===== 1. Extract discontinuous spans =====
+    def extract_spans(self, sample):
+        entity_spans = []
+        
+        for event in sample["entities"]:
+            entity_type = event["label"]
+            spans = [tuple(event["offset"])]
+            entity_spans.append({
+                "spans": spans,
+                "label": entity_type
+            })
+        
+        return entity_spans
+
+    # ===== 2. Word offsets =====
+    def build_word_offsets(self, text, words):
+        offsets = []
+        pointer = 0
+
+        for word in words:
+            start = text.find(word, pointer)
+            end = start + len(word)
+            offsets.append((start, end))
+            pointer = end
+
+        return offsets
+
+    # ===== 3. Char → word =====
+    def char_span_to_word_span(self, word_offsets, start, end):
+        start_word = None
+        end_word = None
+
+        for i, (w_start, w_end) in enumerate(word_offsets):
+            if w_start <= start < w_end:
+                start_word = i
+            if w_start < end <= w_end:
+                end_word = i
+
+        return start_word, end_word
+
+    # ===== 4. Word → subword =====
+    def word_to_subword_map(self, words):
+        mapping = []
+        subword_index = 1  # <s>
+
+        for word in words:
+            sub_tokens = self.tokenizer.tokenize(word)
+            start = subword_index
+            end = subword_index + len(sub_tokens) - 1
+            mapping.append((start, end))
+            subword_index += len(sub_tokens)
+
+        return mapping
+
+    # ===== 5. Span → subword =====
+    def span_to_subword(self, word_offsets, word_subword_map, spans):
+        sub_spans = []
+
+        for span_start, span_end in spans:
+            w_start, w_end = self.char_span_to_word_span(
+                word_offsets, span_start, span_end
+            )
+            if w_start is None or w_end is None:
+                continue
+
+            sub_start = word_subword_map[w_start][0]
+            sub_end = word_subword_map[w_end][1]
+            sub_spans.append((sub_start, sub_end))
+
+        return sub_spans
+
+    def extract_valid_spans(self, sub_spans):
+        valid_spans = []
+        for s, e in sub_spans:
+            if s < 0 or e < 0 or s >= self.max_len or e >= self.max_len or s > e:
+                continue
+            valid_spans.append((s, e))
+        return valid_spans
+
+    def encode(self, sample):
+        text = sample["text"]
+        entities = self.extract_spans(sample)
+    
+        # ===== 1. Word tokenize =====
+        words = word_tokenize(text)
+        sentence = " ".join(words)
+    
+        # ===== 2. Mapping =====
+        word_offsets = self.build_word_offsets(text, words)
+        word_subword_map = self.word_to_subword_map(words)
+    
+        # ===== 3. Tokenize FULL =====
+        encoding = self.tokenizer(
+            sentence,
+            max_length=self.max_len,
+            truncation=True,
+            padding="max_length",
+            return_tensors="pt"
+        )
+        input_ids = encoding["input_ids"][0]
+        attention_mask = encoding["attention_mask"][0]
+    
+        # ===== 5. Convert spans =====
+        entities_gold_spans = []
+    
+        for ent in entities:
+            label = ent["label"]
+            
+            sub_spans = self.span_to_subword(
+                word_offsets,
+                word_subword_map,
+                ent["spans"]
+            )
+            valid_spans = self.extract_valid_spans(sub_spans)
+            if len(valid_spans) == 0:
+                continue    
+            entities_gold_spans.append((tuple(valid_spans), label))
+                
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "entities_gold_spans": entities_gold_spans,
+        }
+
+def generate_spans(attention_mask, max_span_len):
+    seq_len = attention_mask.sum().item() - 2
+    spans = []
+    for i in range(1, seq_len+1):
+        for j in range(i, min(i+max_span_len, seq_len+1)):
+            spans.append((i, j))
+    return spans
+
+def match_gold_labels(
+    gold_spans,      # (N, 2)
+    gold_labels,     # (N,)
+    pred_spans,      # (M, 2)
+    default_label=-100
+):
+    """
+    Return:
+        pred_labels: (M,)
+    """
+
+    pred_labels = torch.full(
+        (pred_spans.size(0),),
+        default_label,
+        dtype=gold_labels.dtype,
+        device=gold_labels.device
+    )
+    if gold_spans.size(0) == 0:
+        return pred_labels
+
+    # (M, N)
+    matched = (pred_spans[:, None, :] == gold_spans[None, :, :]).all(dim=-1)
+    has_match = matched.any(dim=1)
+
+    # lấy index gold đầu tiên match
+    gold_idx = matched.float().argmax(dim=1)
+
+    pred_labels[has_match] = gold_labels[gold_idx[has_match]]
+
+    return pred_labels
+    
+class KLTNDataset(Dataset):
+    def __init__(self, all_data, using_idxes, label2id, tokenizer, max_len, max_n_parts):
+        super().__init__()
+        self.tokenizer = tokenizer
+        self.aligner = PhoBERTSpanAligner(tokenizer, max_len*max_n_parts)
+        self.all_data = all_data
+        self.using_idxes = using_idxes
+        self.label2id = label2id
+        self.max_len = max_len
+        self.max_n_parts = max_n_parts
+
+    def __len__(self):
+        return len(self.using_idxes)
+
+    def __getitem__(self, idx):
+        ridx = self.using_idxes[idx]
+        sample = self.all_data[ridx]
+        result = self.aligner.encode(sample)
+        
+        input_ids = result["input_ids"].squeeze(0)
+        attention_mask = result["attention_mask"].squeeze(0)
+        entities_gold_spans = result["entities_gold_spans"]
+        
+        all_spans = torch.tensor(generate_spans(attention_mask, 10))
+        gold_spans = torch.tensor([spans[0] for spans, _ in entities_gold_spans], dtype=torch.long) if entities_gold_spans else torch.empty(0, 2, dtype=torch.long)
+        gold_labels = torch.tensor([self.label2id[label] for _, label in entities_gold_spans], dtype=torch.long) if entities_gold_spans else torch.empty(0, dtype=torch.long)
+        all_labels = match_gold_labels(
+            gold_spans,      # (N, 2)
+            gold_labels,     # (N,)
+            all_spans,      # (M, 2)
+            default_label=0
+        )
+
+        # Get label
+        gold_entities = []
+        start_labels = torch.ones_like(input_ids) * (1-attention_mask) * (-100)
+        end_labels = torch.ones_like(input_ids) * (1-attention_mask) * (-100)
+        for spans, label in entities_gold_spans:
+            s, e = spans[0]
+            
+            start_labels[s] = self.label2id[f'{label}']
+            end_labels[e] = self.label2id[f'{label}']
+            
+            gold_entities.append((tuple(input_ids[s:e+1].tolist()), label))
+        
+        input_ids = input_ids.reshape(self.max_n_parts, self.max_len)
+        attention_mask = attention_mask.reshape(self.max_n_parts, self.max_len)
+
+        n_valid_parts = math.ceil(attention_mask.sum().item() / self.max_len)
+        input_ids = input_ids[:n_valid_parts]
+        attention_mask = attention_mask[:n_valid_parts]
+        start_labels = start_labels[:n_valid_parts*self.max_len]
+        end_labels = end_labels[:n_valid_parts*self.max_len]
+    
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            
+            "all_spans": all_spans,
+            "all_labels": all_labels,
+            
+            "gold_spans": gold_spans,
+            "gold_labels": gold_labels,
+            
+            "start_labels": start_labels, 
+            "end_labels": end_labels, 
+            
+            "gold_entities": gold_entities,
+        }
+
+def _pad_batch(tensor_list, pad_value=0):
+    """
+    tensor_list: list of tensors
+        mỗi tensor shape: (Nk, n_parts_i, max_len_i)
+
+    return:
+        padded tensor shape: (B, max_Nk, max_n_parts, max_len)
+    """
+
+    # lấy max toàn batch
+    max_Nk = max(t.size(0) for t in tensor_list)
+    max_n_parts = max(t.size(1) for t in tensor_list)
+    max_len = max(t.size(2) for t in tensor_list)
+
+    padded = []
+
+    for t in tensor_list:
+        Nk, n_parts_i, max_len_i = t.shape
+
+        # pad chiều n_parts và max_len trước
+        if n_parts_i < max_n_parts or max_len_i < max_len:
+            new_t = t.new_full(
+                (Nk, max_n_parts, max_len),
+                pad_value
+            )
+            new_t[:, :n_parts_i, :max_len_i] = t
+            t = new_t
+
+        # pad chiều Nk
+        if Nk < max_Nk:
+            pad_tensor = t.new_full(
+                (max_Nk - Nk, max_n_parts, max_len),
+                pad_value
+            )
+            t = torch.cat([t, pad_tensor], dim=0)
+
+        padded.append(t)
+
+    return torch.stack(padded)   # (B, max_Nk, max_n_parts, max_len)
+
+def collate_fn(batch):
+    gold_entities = []
+    for bidx, b in enumerate(batch):
+        for entity in b['gold_entities']:
+            gold_entities.append([bidx, entity])
+
+    input_ids = [b["input_ids"].unsqueeze(-1) for b in batch]
+    attention_mask = [b["attention_mask"].unsqueeze(-1) for b in batch]
+    
+    all_spans = [b["all_spans"].unsqueeze(-1) for b in batch]
+    all_labels = [b["all_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    
+    gold_spans = [b["gold_spans"].unsqueeze(-1) for b in batch]
+    gold_labels = [b["gold_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    
+    start_labels = [b["start_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    end_labels = [b["end_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+
+    # pad theo Nk
+    input_ids = _pad_batch(input_ids, pad_value=0).squeeze(-1)
+    attention_mask = _pad_batch(attention_mask, pad_value=0).squeeze(-1)
+    
+    all_spans = _pad_batch(all_spans, pad_value=0).squeeze(-1)
+    all_labels = _pad_batch(all_labels, pad_value=-100).squeeze(-1).squeeze(-1)
+    
+    gold_spans = _pad_batch(gold_spans, pad_value=0).squeeze(-1)
+    gold_labels = _pad_batch(gold_labels, pad_value=-100).squeeze(-1).squeeze(-1)
+    
+    start_labels = _pad_batch(start_labels, pad_value=-100).squeeze(-1).squeeze(-1)
+    end_labels = _pad_batch(end_labels, pad_value=-100).squeeze(-1).squeeze(-1)
+    
+    return {
+        "input_ids": input_ids,
+        "attention_mask": attention_mask,
+        
+        "all_spans": all_spans,
+        "all_labels": all_labels,
+        
+        "gold_spans": gold_spans,
+        "gold_labels": gold_labels,
+        
+        "start_labels": start_labels, 
+        "end_labels": end_labels, 
+        
+        "gold_entities": gold_entities,
+    }
+
+# %% [code]
+def shift_bidx(spans, batch_idx):
+    shifted = []
+    for bidx, ent in spans:
+        new_bidx = bidx + batch_idx * batch_size
+        shifted.append((new_bidx, ent))
+    return shifted
+
+def refactor_entities(entities, save_dict):
+    i, c = [], []
+    for bidx, (ids, lb) in entities:
+        if (bidx, ids) not in i:
+            i.append((bidx, ids))
+            
+        if (bidx, (ids, lb)) not in c:
+            c.append((bidx, (ids, lb)))
+
+    save_dict['Ent-I'].extend(i)
+    save_dict['Ent-C'].extend(c)
+
+def check_spans(
+    all_spans,                 # (N, 2)
+    all_labels,                # (N,)
+    ensemble_start_logits,     # (L, C)
+    ensemble_end_logits,       # (L, C)
+    expand_dict,
+    max_expand=10
+):
+    """
+    Check minimum expansion radius required to achieve recall = 1.
+
+    Args:
+        all_spans: gold spans
+        all_labels: gold labels
+        ensemble_start_logits: (L, C)
+        ensemble_end_logits: (L, C)
+        expand_dict:
+            dict like:
+            {
+                0: count,
+                1: count,
+                ...
+            }
+
+    Return:
+        matched_expand_k
+    """
+
+    # =========================================================
+    # Gold spans
+    # =========================================================
+
+    gold_set = set()
+
+    valid = all_labels > 0
+
+    valid_idxes = valid.nonzero(as_tuple=False).squeeze(-1)
+
+    for idx in valid_idxes:
+
+        s = int(all_spans[idx, 0])
+        e = int(all_spans[idx, 1])
+
+        gold_set.add((s, e))
+
+    # no gold
+    if len(gold_set) == 0:
+        expand_dict[0] += 1
+        return 0
+
+    # =========================================================
+    # Decode base spans
+    # =========================================================
+
+    start_labels = ensemble_start_logits.argmax(dim=-1)  # (L,)
+    end_labels   = ensemble_end_logits.argmax(dim=-1)    # (L,)
+
+    L = start_labels.shape[0]
+
+    base_pred = []
+
+    used_start = set()
+    used_end = set()
+
+    for s in range(L):
+
+        s_label = start_labels[s].item()
+
+        if s_label == 0:
+            continue
+
+        if s in used_start:
+            continue
+
+        nearest_e = None
+
+        for e in range(s, L):
+
+            if e in used_end:
+                continue
+
+            e_label = end_labels[e].item()
+
+            if e_label == s_label:
+                nearest_e = e
+                break
+
+        if nearest_e is None:
+            continue
+
+        used_start.add(s)
+        used_end.add(nearest_e)
+
+        base_pred.append((s, nearest_e))
+
+    # =========================================================
+    # Try expansion radius
+    # =========================================================
+
+    for k in range(max_expand + 1):
+
+        pred_set = set()
+
+        for s, e in base_pred:
+
+            for ds in range(-k, k + 1):
+                for de in range(-k, k + 1):
+
+                    ns = s + ds
+                    ne = e + de
+
+                    if ns > 0 and ne > 0 and ns <= ne:
+                        pred_set.add((ns, ne))
+
+        # recall = 1
+        if gold_set.issubset(pred_set):
+
+            expand_dict[k] += 1
+            return k
+
+    # =========================================================
+    # cannot recover
+    # =========================================================
+
+    expand_dict[-1] = expand_dict.get(-1, 0) + 1
+
+    return -1
+
+def test(network, state_dicts, test_loader, eval_fn, analyzer, device, id2label, tokenizer):
+    if torch.cuda.device_count() > 1:
+        network = DataParallelProxy(network)
+    network = network.to(device)
+    network.eval()
+
+    eval_types = ['Ent-I', 'Ent-C']
+    
+    all_pred = {eval_type: [] for eval_type in eval_types}
+    all_gold = {eval_type: [] for eval_type in eval_types}
+
+    list_input_ids = []
+    expand_dict = {i: 0 for i in range(13)}
+
+    with torch.no_grad():
+        for batch_idx, batch in enumerate(test_loader):
+            input_ids = batch['input_ids'].to(device)
+            attention_mask = batch['attention_mask'].to(device)
+            all_spans = batch['all_spans'].to(device)
+            all_labels = batch['all_labels'].to(device)
+            gold_entities = batch['gold_entities']
+
+            B, _, _ = input_ids.shape
+            list_input_ids.extend(input_ids.reshape(B, -1).tolist())
+
+            list_hidden_states = []
+            list_start_logits = []
+            list_end_logits = []
+            list_scores = []
+            for sd in state_dicts:
+                if torch.cuda.device_count() > 1:
+                    network.module.load_state_dict(sd)
+                else:
+                    network.load_state_dict(sd)
+                
+                hidden_states, attention_mask = network.encode(input_ids, attention_mask)
+                start_logits, end_logits = network.get_logits(hidden_states)
+                list_hidden_states.append(hidden_states)
+                list_start_logits.append(start_logits)
+                list_end_logits.append(end_logits)
+                
+            ensemble_start_logits = torch.stack(list_start_logits, dim=0).mean(dim=0)
+            ensemble_end_logits = torch.stack(list_end_logits, dim=0).mean(dim=0)
+            spans, pred_labels = extract_spans_and_labels(ensemble_start_logits, ensemble_end_logits)
+            B, N, _ = spans.shape
+            r = torch.ones((B, N), device=spans.device, dtype=torch.long) * network.max_r
+            expanded_spans = expand_spans(spans, r, attention_mask)
+
+            for sd, hidden_states in zip(state_dicts, list_hidden_states):
+                if torch.cuda.device_count() > 1:
+                    network.module.load_state_dict(sd)
+                else:
+                    network.load_state_dict(sd)
+                expanded_span_reprs = get_span_reprs(hidden_states, expanded_spans)
+                expanded_span_scores = network.get_scores(expanded_span_reprs)
+                list_scores.append(expanded_span_scores)
+
+            ensemble_scores = torch.stack(list_scores, dim=0).mean(dim=0)
+            pred_spans = select_best_spans(expanded_spans, ensemble_scores)
+                
+            pred_entities = extract_entities(input_ids.reshape(B, -1), pred_spans, pred_labels, id2label)
+            pred_entities = shift_bidx(pred_entities, batch_idx)
+            refactor_entities(pred_entities, all_pred)
+                
+            gold_entities = shift_bidx(gold_entities, batch_idx)
+            refactor_entities(gold_entities, all_gold)
+
+            for b in range(B):
+                check_spans(
+                    all_spans[b],                 # (N, 2)
+                    all_labels[b],                # (N,)
+                    ensemble_start_logits[b],     # (L, C)
+                    ensemble_end_logits[b],       # (L, C)
+                    expand_dict,
+                    max_expand=10
+                )
+
+    # ===== GLOBAL EVAL =====
+    final_score = {}
+    for eval_type in eval_types:
+        score = eval_fn(list_to_tuple(all_pred[eval_type]), list_to_tuple(all_gold[eval_type]))
+        final_score[eval_type] = score
+
+    analyze_result = analyzer.analyze(list_to_tuple(all_pred['Ent-I']),  list_to_tuple(all_gold['Ent-I']))
+
+    # ===== PREDICT =====
+    predictions = []
+    for input_ids in list_input_ids:
+        predictions.append([tokenizer.decode(input_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)])
+    for bidx, (ids, lb) in all_pred['Ent-C']:
+        predictions[bidx].append((tokenizer.decode(ids, skip_special_tokens=True, clean_up_tokenization_spaces=True), lb))
+
+    return final_score, analyze_result, predictions, expand_dict
+
+# %% [code]
+with open(f'{train_dir}/train.json', "r", encoding="utf-8") as f:
+    data_train = json.load(f)
+    
+with open(f'{test_dir}/test.json', "r", encoding="utf-8") as f:
+    data_test = json.load(f)
+
+print('Train:', len(data_train))
+print('Test:', len(data_test))
+
+# %% [code]
+entity_types = ['O'] + sorted(list(set([e['label'] for d in data_train + data_test for e in d['entities']])))
+# bio_entity_type = ['O'] + [f'{prefix}-{ent}' for ent in entity_types for prefix in ['B', 'I']]
+label2id = {l: i for i, l in enumerate(entity_types)}
+id2label = {i: l for l, i in label2id.items()}
+
+# %% [code]
+zero_entities_idxes = []
+for idx, d in enumerate(data_train):
+    if len(d['entities']) == 0:
+        zero_entities_idxes.append(idx)
+
+n_zero_entities_samples = len(zero_entities_idxes)
+n_has_entities_samples = len(data_train) - n_zero_entities_samples
+
+random.seed(42)
+k = min(int(n_has_entities_samples * zero_entities_rate), len(zero_entities_idxes))
+sampled_zero_entities_idxes = random.sample(zero_entities_idxes, k)
+
+new_data_train = []
+for idx, d in enumerate(data_train):
+    if len(d['entities']) == 0:
+        if idx in sampled_zero_entities_idxes:
+            new_data_train.append(d)
+    else:
+        new_data_train.append(d)
+data_train = new_data_train
+
+print('Train:', len(data_train))
+
+# %% [code]
+if debug_only:
+    data_train = data_train[:10]
+    data_test = data_test[:10]
+
+    print('Train:', len(data_train))
+    print('Test:', len(data_test))
+
+# %% [code]
+tokenizer = AutoTokenizer.from_pretrained(backbone_model_name)
+
+# %% [code]
+print('Experiment name:', state_dict_save_name)
+
+# %% [code]
+if not test_only:
+    full_idxes = np.array(range(len(data_train)))
+    training_logs, best_models, last_models = [], [], []
+    start_training_time = time.time()
+    for seed in SEEDS:
+        kf = KFold(n_splits=nfolds, shuffle=True, random_state=seed)
+        for fold_idx, (tr_idx, va_idx) in enumerate(kf.split(full_idxes)):
+            if only_fold_idx is not None and only_fold_idx >= 0 and only_fold_idx != fold_idx:
+                continue
+            set_seed(seed)
+    
+            train_idxes, val_idxes = full_idxes[tr_idx], full_idxes[va_idx]
+    
+            trainset = KLTNDataset(data_train, train_idxes, label2id, tokenizer, **train_memory_params)
+            valset = KLTNDataset(data_train, val_idxes, label2id, tokenizer, **val_memory_params)
+            
+            generator = torch.Generator()
+            generator.manual_seed(seed)
+            train_loader = DataLoader(trainset, generator=generator, collate_fn=collate_fn, **train_loader_params)
+            val_loader = DataLoader(valset, generator=generator, collate_fn=collate_fn, **val_loader_params)
+    
+            my_model = IEModel(
+                num_labels=len(label2id), 
+                **model_params
+            )
+            total_params = sum(p.numel() for p in my_model.parameters())
+            print(f"Total params: {total_params:,}")
+            
+            # optimizer, scheduler = configure_optimizers(my_model, optim_params, scheduler_params)
+            encoder_params = set(map(id, my_model.encoder.parameters()))
+            other_params = [
+                p for p in my_model.parameters()
+                if id(p) not in encoder_params
+            ]        
+            optimizer = optim.AdamW([
+                {"params": my_model.encoder.parameters(), "lr": 2e-5},
+                {"params": other_params}
+            ], lr=5e-4)
+            scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=20, eta_min=1e-6)
+            
+            loss_fn = CustomLoss(
+                **loss_func_params
+            )
+            eval_fn = CustomEvalFn(**eval_func_params)
+            trainer_params['save_name'] = f'{state_dict_save_name}_s{seed}_f{fold_idx}'
+            trainer = Trainer(**trainer_params)
+    
+            print(f'Start Training Fold {fold_idx}...')
+            training_log, best_model, last_model = trainer.fit(
+                my_model, optimizer, scheduler, loss_fn, epochs, train_loader, val_loader, eval_fn, 
+                start_epoch=1, start_training_time=start_training_time, id2label=id2label
+            )
+            
+            training_logs.append(training_log)
+            best_models.append(best_model)
+            last_models.append(last_model)
+
+# %% [code]
+def load_all_state_dicts(folder):
+    files = []
+    
+    for file in os.listdir(folder):
+        if file.endswith(".pt") or file.endswith(".pth"):
+            m = re.search(r"f(\d+)", file)   # tìm f<số>
+            if m:
+                fold = int(m.group(1))
+                files.append((fold, file))
+
+    # sort theo fold
+    files.sort(key=lambda x: x[0])
+
+    state_dicts = []
+    for fold, file in files:
+        path = os.path.join(folder, file)
+        print(f"Loading fold {fold}: {file}")
+        state_dict = torch.load(path, map_location="cpu")
+        state_dicts.append(state_dict)
+
+    return state_dicts
+
+if test_only:
+    snapshot_download(repo_id=repo_name, local_dir="", repo_type="model", allow_patterns=[f"{state_dict_save_name}/**"])
+    get_ipython().system('rm -rf .cache .gitattributes')
+    
+    best_models = load_all_state_dicts(f"{state_dict_save_name}/r1s")
+    last_models = load_all_state_dicts(f"{state_dict_save_name}/lasts")
+
+# %% [code]
+os.makedirs(f'{checkpoints_dir}/results', exist_ok=True)
+testset = KLTNDataset(data_test, range(len(data_test)), label2id, tokenizer, **val_memory_params)
+generator = torch.Generator()
+test_loader = DataLoader(testset, generator=generator, collate_fn=collate_fn, **val_loader_params)
+eval_fn = CustomEvalFn(**eval_func_params)
+analyzer = SpanErrorAnalyzer()
+my_model = IEModel(
+    num_labels=len(label2id), 
+    **model_params
+)
+total_params = sum(p.numel() for p in my_model.parameters())
+print(f"Total params: {total_params:,}")
+
+# %% [code]
+start_time = time.time()
+result_test = None
+analyze_result = None
+
+best_score, best_analyze_result, best_pred_test, expand_dict = test(my_model, best_models, test_loader, eval_fn, analyzer, device, id2label, tokenizer)
+last_score, last_analyze_result, last_pred_test, _ = test(my_model, last_models, test_loader, eval_fn, analyzer, device, id2label, tokenizer)
+
+result_test = {"Best model": best_score, "Last model": last_score}
+analyze_result = {"Best model": best_analyze_result, "Last model": last_analyze_result}
+analyze_result_sumary = {"Best model": best_analyze_result['summary'], "Last model": last_analyze_result['summary']}
+pred_test = {"Best model": best_pred_test, "Last model": last_pred_test}
+
+with open(f"{checkpoints_dir}/results/{state_dict_save_name}_test.json", "w", encoding="utf-8") as f:
+    json.dump(result_test, f, ensure_ascii=False, indent=2)
+
+with open(f"{checkpoints_dir}/results/{state_dict_save_name}_error_analyze_result.json", "w", encoding="utf-8") as f:
+    json.dump(analyze_result, f, ensure_ascii=False, indent=2)
+
+with open(f"{checkpoints_dir}/results/{state_dict_save_name}_pred_test.json", "w", encoding="utf-8") as f:
+    json.dump(pred_test, f, ensure_ascii=False, indent=2)
+    
+print('Test:', time.time() - start_time, 's --> Done!')
+print(json.dumps(analyze_result_sumary, ensure_ascii=False, indent=4))
+
+# %% [code]
+expand_dict
+
+# %% [code]
+expand_dict_sum = sum(list(expand_dict.values()))
+{key: value / expand_dict_sum for key, value in expand_dict.items()}
+
+# %% [code]
+best_pred_test[:10]
+
+# %% [code]
+last_pred_test[:10]
+
+# %% [code]
+def dict_to_df(data):
+    row_tuples = []
+    row_values = []
+
+    metrics = ["precision", "recall", "f1"]
+
+    # Lấy model đầu tiên
+    first_model = next(iter(data.values()))
+
+    # eval_keys
+    eval_keys = list(first_model.keys())
+
+    for eval_key in eval_keys:
+        row_tuples.append(eval_key)
+        row = {}
+
+        for model_name, model_data in data.items():
+            for metric in metrics:
+                row[(model_name, metric)] = model_data[eval_key][metric]
+
+        row_values.append(row)
+
+    # ===== DataFrame =====
+    df = pd.DataFrame(row_values)
+
+    # MultiIndex columns
+    df.columns = pd.MultiIndex.from_tuples(df.columns)
+
+    # Index
+    df.index = pd.Index(row_tuples, name="evaluation")
+
+    # ===== Sort =====
+    sort_keys = []
+    if ("Best model", "f1") in df.columns:
+        sort_keys.append(("Best model", "f1"))
+    if ("Last model", "f1") in df.columns:
+        sort_keys.append(("Last model", "f1"))
+
+    if sort_keys:
+        df = df.sort_values(by=sort_keys, ascending=False)
+
+    return df
+
+result_test_df = dict_to_df(result_test)
+result_test_df.to_excel(f"{checkpoints_dir}/results/{state_dict_save_name}_test_df.xlsx")
+result_test_df
+
+# %% [code]
+key = ("Best model", "f1")
+result_test_df_best = result_test_df.sort_values(by=key, ascending=False).groupby(level="evaluation").head(1)
+result_test_df_best.to_excel(f"{checkpoints_dir}/results/{state_dict_save_name}_test_df_best.xlsx")
+result_test_df_best
+
+# %% [code]
+def get_avg_best_score(logs):
+    return float(np.mean([list(log.values())[-1]['best_score'] for log in logs]))
+    
+def get_avg_log(logs, epochs):
+    avg_log = {}
+
+    for epoch in range(1, epochs + 1):
+        val_score = 0.0
+        train_loss = 0.0
+        n_eval = 0
+
+        for idx in range(len(logs)):
+            log = logs[idx].get(epoch, logs[idx].get(str(epoch)))
+            if log is None:
+                continue
+
+            val_score += log.get('val_score', 0.0)
+            train_loss += log.get('train_loss', 0.0)
+            n_eval += 1
+
+        if n_eval == 0:
+            continue
+
+        avg_log[epoch] = {
+            'train_loss': train_loss / n_eval,
+            'val_score': val_score / n_eval if val_score != 0 else float('inf')
+        }
+
+    return avg_log
+
+def parse_label_key(label: str):
+    try:
+        first = float(label.split('_', 1)[0]) # số đầu: trước dấu _
+        last = float(re.findall(r'_(\d+(?:\.\d+)?)$', label)[0])
+        return first, last
+    except:
+        return (0, 0)
+
+def plot_training_logs(logs_dict, save_path=None, figsize=(24, 10)):
+    fig, axes = plt.subplots(1, 2, figsize=figsize)
+
+    # ===== Plot Train Loss =====
+    for name, log in logs_dict.items():
+        epochs = sorted(log.keys())
+        train_loss = [log[e]['train_loss'] for e in epochs]
+        axes[0].plot(epochs, train_loss, label=name)
+
+    axes[0].set_xlabel('Epoch')
+    axes[0].set_ylabel('Train Loss')
+    axes[0].set_title('Training Loss')
+    axes[0].grid(True)
+
+    # ===== Plot Validation Score =====
+    for name, log in logs_dict.items():
+        epochs = sorted(log.keys())
+        val_score = [log[e]['val_score'] for e in epochs]
+        axes[1].plot(epochs, val_score, label=name)
+
+    axes[1].set_xlabel('Epoch')
+    axes[1].set_ylabel('Validation Score')
+    axes[1].set_title('Validation Score')
+    axes[1].grid(True)
+
+    # ===== Shared Legend =====
+    handles, labels = axes[0].get_legend_handles_labels()
+    pairs = list(zip(handles, labels))
+    pairs_sorted = sorted(
+        pairs,
+        key=lambda x: parse_label_key(x[1])
+    )
+    handles_sorted, labels_sorted = zip(*pairs_sorted)
+    
+    axes[0].legend(
+        handles_sorted,
+        labels_sorted,
+        loc='center left',
+        bbox_to_anchor=(1.01, 0.5),
+        borderaxespad=0.
+    )
+
+    plt.tight_layout(rect=[0, 0, 1, 1])
+
+    if save_path is not None:
+        os.makedirs(os.path.dirname(save_path), exist_ok=True) if os.path.dirname(save_path) else None
+        plt.savefig(save_path, dpi=300, bbox_inches='tight')
+
+    plt.show()
+
+# %% [code]
+if not test_only:
+    snapshot_download(repo_id=repo_name, local_dir="", repo_type="model", allow_patterns=["**/*entities*.json"])
+    get_ipython().system('rm -rf .cache .gitattributes')
+
+# %% [code]
+if not test_only:
+    experiments = {}
+    for experiment in os.listdir(pretrained_dir):
+        if '.virtual_documents' in experiment:
+            continue
+        experiment_logs = []
+        try:
+            for seed in SEEDS:
+                for fold_idx in range(nfolds):
+                    with open(f"{pretrained_dir}/{experiment}/logs/{experiment}_s{seed}_f{fold_idx}_logging.json", "r", encoding="utf-8") as f:
+                        experiment_log = json.load(f)
+                    experiment_logs.append(experiment_log)
+        except:
+            pass
+        experiments[experiment] = get_avg_log(experiment_logs, 1000)
+    experiments[state_dict_save_name] = get_avg_log(training_logs, 1000)
+
+# %% [code]
+if not test_only:
+    score = get_avg_best_score(training_logs)
+    state_dict_save_name, score
+
+# %% [code]
+if not test_only:
+    plot_training_logs(experiments, save_path=f'{checkpoints_dir}/logs/{state_dict_save_name}_log_plot.jpg', figsize=(18, 7.5))
+

4.2_add_span_rerank_branch_4.5/lasts/4.2_add_span_rerank_branch_4.5_s26092004_f0_last_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9366440dbc8ba6e599af7db1fe997834b33d18987933e945063fa8b3b0476d2
+size 554283006

4.2_add_span_rerank_branch_4.5/logs/4.2_add_span_rerank_branch_4.5_s26092004_f0_logging.json

@@ -0,0 +1 @@
+{"1": {"lr": [2e-05, 0.0005], "train_loss": 0.1468505859375, "total": 0.14686975964225824, "token_margin_loss": 0.05261319172722191, "span_loss": 0.09663219675796535, "start_margin": 0.024280324203465622, "end_margin": 0.02826299608719955}, "2": {"lr": [1.988303923565381e-05, 0.0004969282409784868], "train_loss": 0.11334228515625, "total": 0.11333147009502516, "token_margin_loss": 0.04049049748462828, "span_loss": 0.07022079373951928, "start_margin": 0.019179709334823925, "end_margin": 0.02199203465623253}, "3": {"lr": [1.9535036904803962e-05, 0.0004877886008156408], "train_loss": 0.1005859375, "total": 0.10061486864169927, "token_margin_loss": 0.03818474007825601, "span_loss": 0.06005449972051426, "start_margin": 0.01807923420905534, "end_margin": 0.02041992733370598}, "4": {"lr": [1.8964561979789496e-05, 0.00047280612778499774], "train_loss": 0.091064453125, "total": 0.09104248183342649, "token_margin_loss": 0.036053661263275576, "span_loss": 0.051669927333705984, "start_margin": 0.01755519843487982, "end_margin": 0.01916224147568474}, "5": {"lr": [1.8185661446562005e-05, 0.00045234974009654937], "train_loss": 0.08251953125, "total": 0.08251816657350475, "token_margin_loss": 0.035075461151481274, "span_loss": 0.0451369480156512, "start_margin": 0.017162171604248183, "end_margin": 0.018323784237003912}, "6": {"lr": [1.7217514421272206e-05, 0.00042692314190604356], "train_loss": 0.07568359375, "total": 0.07567076579094466, "token_margin_loss": 0.034219536053661265, "span_loss": 0.03902319731693683, "start_margin": 0.016769144773616546, "end_margin": 0.01759886808272778}, "7": {"lr": [1.60839598967785e-05, 0.00039715242044697206], "train_loss": 0.0699462890625, "total": 0.06997624371157071, "token_margin_loss": 0.03364309670206819, "span_loss": 0.03380030743432085, "start_margin": 0.016463457238680826, "end_margin": 0.01731938233650084}, "8": {"lr": [1.4812909747525698e-05, 0.00036377062968501693], "train_loss": 0.06414794921875, "total": 0.06417691447736165, "token_margin_loss": 0.032752235885969816, "span_loss": 0.03002724986025713, "start_margin": 0.01611410005589715, "end_margin": 0.016900153717160426, "val_score": 0.6863627350530408, "best_score": 0.6863627350530408, "new_best_model": true, "precision": 0.7682875104263381, "recall": 0.6226203064307612, "f1": 0.6863627350530408}, "9": {"lr": [1.3435661446562005e-05, 0.0003275997400965494], "train_loss": 0.0589599609375, "total": 0.058971492453884854, "token_margin_loss": 0.03236794298490777, "span_loss": 0.025171185019564002, "start_margin": 0.01585208216880939, "end_margin": 0.016550796534376747, "val_score": 0.6859065883350041, "best_score": 0.6863627350530408, "new_best_model": false, "precision": 0.7664244548681001, "recall": 0.6230380104041229, "f1": 0.6859065883350041}, "10": {"lr": [1.1986127417882198e-05, 0.00028953039902753766], "train_loss": 0.05474853515625, "total": 0.05474427054220235, "token_margin_loss": 0.031424678591391836, "span_loss": 0.021974566797093347, "start_margin": 0.015467789267747344, "end_margin": 0.01621017328116266, "val_score": 0.686874862767166, "best_score": 0.686874862767166, "new_best_model": true, "precision": 0.76764898675689, "recall": 0.6239278347187776, "f1": 0.686874862767166}, "11": {"lr": [1.0500000000000003e-05, 0.0002505], "train_loss": 0.050750732421875, "total": 0.05076159865846842, "token_margin_loss": 0.031145192845164895, "span_loss": 0.019267048630519843, "start_margin": 0.01526690888764673, "end_margin": 0.01590448574622694, "val_score": 0.6872180862141289, "best_score": 0.6872180862141289, "new_best_model": true, "precision": 0.766240335643193, "recall": 0.6254058784056634, "f1": 0.6872180862141289}, "12": {"lr": [9.013872582117811e-06, 0.00021146960097246246], "train_loss": 0.046966552734375, "total": 0.046953605366126326, "token_margin_loss": 0.030568753493571826, "span_loss": 0.016524594745667972, "start_margin": 0.01500489100055897, "end_margin": 0.015694871436556734, "val_score": 0.6905636463106959, "best_score": 0.6905636463106959, "new_best_model": true, "precision": 0.7693898045943346, "recall": 0.6288075361008783, "f1": 0.6905636463106959}, "13": {"lr": [7.564338553438001e-06, 0.00017340025990345064], "train_loss": 0.044830322265625, "total": 0.044822526551145894, "token_margin_loss": 0.030236864169927335, "span_loss": 0.015048560648406931, "start_margin": 0.01477780883174958, "end_margin": 0.015467789267747344, "val_score": 0.6903134720088675, "best_score": 0.6905636463106959, "new_best_model": false, "precision": 0.7691587246053712, "recall": 0.6284828908420117, "f1": 0.6903134720088675}, "14": {"lr": [6.1870902524743065e-06, 0.00013722937031498307], "train_loss": 0.041961669921875, "total": 0.04195779765231973, "token_margin_loss": 0.029380939072107322, "span_loss": 0.012882546115148127, "start_margin": 0.014358580212409168, "end_margin": 0.014987423141419787, "val_score": 0.6894675601432522, "best_score": 0.6905636463106959, "new_best_model": false, "precision": 0.7691616105349628, "recall": 0.6270536907329344, "f1": 0.6894675601432522}, "15": {"lr": [4.916040103221507e-06, 0.00010384757955302797], "train_loss": 0.039398193359375, "total": 0.03940749021799888, "token_margin_loss": 0.029206260480715483, "span_loss": 0.01079513694801565, "start_margin": 0.014210103409726104, "end_margin": 0.014961221352711012, "val_score": 0.6902389874528908, "best_score": 0.6905636463106959, "new_best_model": false, "precision": 0.7696662485785113, "recall": 0.6279449177801002, "f1": 0.6902389874528908}, "16": {"lr": [3.7824855787278e-06, 7.40768580939564e-05], "train_loss": 0.0382080078125, "total": 0.038219675796534375, "token_margin_loss": 0.02869969256567915, "span_loss": 0.009965413638904417, "start_margin": 0.01392188373392957, "end_margin": 0.014611864169927333, "val_score": 0.6909374894379251, "best_score": 0.6909374894379251, "new_best_model": true, "precision": 0.7695734193089366, "recall": 0.6293325769979943, "f1": 0.6909374894379251}, "17": {"lr": [2.814338553438001e-06, 4.865025990345063e-05], "train_loss": 0.036468505859375, "total": 0.03647288988261599, "token_margin_loss": 0.028315399664617106, "span_loss": 0.008567984907769704, "start_margin": 0.013703535494689771, "end_margin": 0.014428451648965902, "val_score": 0.6899378747368216, "best_score": 0.6909374894379251, "new_best_model": false, "precision": 0.7691208417645939, "recall": 0.6279288525760165, "f1": 0.6899378747368216}, "18": {"lr": [2.0354380202105066e-06, 2.8193872215002235e-05], "train_loss": 0.0355224609375, "total": 0.03551215762996087, "token_margin_loss": 0.028053381777529345, "span_loss": 0.007423840134153158, "start_margin": 0.013546324762437115, "end_margin": 0.014253773057574064, "val_score": 0.6889585967473153, "best_score": 0.6909374894379251, "new_best_model": false, "precision": 0.7676446066334408, "recall": 0.6272630717039508, "f1": 0.6889585967473153}}

4.2_add_span_rerank_branch_4.5/r1s/4.2_add_span_rerank_branch_4.5_s26092004_f0_r1_vs0.69094_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05dcaf378904d9084f13eec451dd3fab35306a10332baa3949c0a25374ad9b5c
+size 554284870

4.2_add_span_rerank_branch_4.5/results/4.2_add_span_rerank_branch_4.5_test.json

@@ -0,0 +1,26 @@
+{
+  "Best model": {
+    "Ent-I": {
+      "precision": 0.7962769051765022,
+      "recall": 0.636118598382158,
+      "f1": 0.7072439756342697
+    },
+    "Ent-C": {
+      "precision": 0.6529948599594534,
+      "recall": 0.5780481010301995,
+      "f1": 0.6132400698873992
+    }
+  },
+  "Last model": {
+    "Ent-I": {
+      "precision": 0.7959656851370321,
+      "recall": 0.6381633980847308,
+      "f1": 0.7083827652429374
+    },
+    "Ent-C": {
+      "precision": 0.6533472803340445,
+      "recall": 0.5799981428168075,
+      "f1": 0.6144916079837778
+    }
+  }
+}