Upload 0_act_175negs_hr_05_wr10_1's state dict

.gitattributes

@@ -95,3 +95,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 0_act_175negs_hr_05_wr20_1/results/0_act_175negs_hr_05_wr20_1_pred_test.json filter=lfs diff=lfs merge=lfs -text
 0_ent_175_negs_hr_05_wr25_no_all_1/logs/0_ent_175_negs_hr_05_wr25_no_all_1_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
 0_act_175negs_hr_09_1/results/0_act_175negs_hr_09_1_pred_test.json filter=lfs diff=lfs merge=lfs -text
+0_act_175negs_hr_05_wr10_1/results/0_act_175negs_hr_05_wr10_1_pred_test.json filter=lfs diff=lfs merge=lfs -text

0_act_175negs_hr_05_wr10_1/0_act_175negs_hr_05_wr10_1.py

@@ -0,0 +1,2157 @@
+# %% [code]
+get_ipython().system('pip install evaluate seqeval underthesea positional-encodings[pytorch]')
+
+# %% [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 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_actions' # vhe, bkee, casie, kltn/only_issues, kltn/only_actions, 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 = 16
+device = "cuda" if torch.cuda.is_available() else "cpu"
+# # Thêm biến toàn cục nào đó vào đây
+repo_name = 'SS3M/kltn-ien-experiments'
+state_dict_save_name = "0_act_175negs_hr_05_wr10_1"
+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 == "casie" else "vinai/phobert-base"
+word_tokenize = lambda text: [token.text for token in en_tokenize_tool(text)] if dataset == "casie" else vi_tokenize_tool(text)
+max_len_dict = {
+    'kltn/raw': 256, 
+    'kltn/only_issues': 52, 
+    'kltn/only_actions': 69, #69
+    'vhe': 51, 
+    'bkee': 62, 
+    'casie': 40, 
+}
+zero_events_rate_dict = {
+    'kltn/raw': 1000, 
+    'kltn/only_issues': 0, 
+    'kltn/only_actions': 0.2, 
+    'vhe': 1000, # mean keep all zero-events samples
+    'bkee': 1000, 
+    'casie': 1000, 
+}
+
+max_len = max_len_dict[dataset]
+max_n_parts = 3 if dataset in ['kltn/raw'] else 1
+max_span_len = 14
+zero_events_rate = zero_events_rate_dict[dataset]
+n_negs = 175
+
+# 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, 
+    'max_span_len': max_span_len, 
+    'n_negs': n_negs, 
+    'weight_sampling': True, 
+    'weight_rate': 1.0, 
+    'hard_rate': 0.5, 
+}
+val_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+    'max_span_len': max_span_len, 
+    'n_negs': n_negs, 
+    'weight_sampling': True, 
+    'weight_rate': 1.0, 
+    'hard_rate': 0.5, 
+}
+
+# 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 // 4, 
+    '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, 
+}
+
+# Loss Func
+loss_func_params = {
+    'lambda_trg_ce': 1.0,
+    'lambda_arg_ce': 1.0,
+}
+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_trg_ce=1.0, lambda_arg_ce=1.0):
+        super().__init__()
+
+        self.lambda_trg_ce = lambda_trg_ce
+        self.lambda_arg_ce = lambda_arg_ce
+
+        self.ce = nn.CrossEntropyLoss(ignore_index=-100, reduction='none')
+
+    def forward(
+        self,
+        trg_logits, trg_labels, trg_weights,
+        arg_logits, arg_labels, arg_weights,
+    ):
+        device = trg_logits.device
+
+        B, N, C = trg_logits.shape
+
+        trg_logits = trg_logits.reshape(-1, C)
+        trg_labels = trg_labels.reshape(-1)
+        trg_weights = trg_weights.reshape(-1)
+
+        valid_trg = trg_labels != -100
+
+        if valid_trg.any():
+            trg_ce = self.ce(trg_logits, trg_labels)[valid_trg]
+            trg_weights = trg_weights[valid_trg]
+
+            trg_loss = (trg_ce * trg_weights).sum() / trg_weights.sum().clamp(min=1e-8)
+        else:
+            trg_loss = torch.tensor(0.0, device=device)
+
+        B, K, N, C = arg_logits.shape
+
+        arg_logits = arg_logits.reshape(-1, C)
+        arg_labels = arg_labels.reshape(-1)
+        arg_weights = arg_weights.reshape(-1)
+
+        valid_arg = arg_labels != -100
+
+        if valid_arg.any():
+            arg_ce = self.ce(arg_logits, arg_labels)[valid_arg]
+            arg_weights = arg_weights[valid_arg]
+
+            arg_loss = (arg_ce * arg_weights).sum() / arg_weights.sum().clamp(min=1e-8)
+        else:
+            arg_loss = torch.tensor(0.0, device=device)
+
+        total_loss = self.lambda_trg_ce * trg_loss + self.lambda_arg_ce * arg_loss
+
+        return {
+            "total": total_loss,
+            "trg_loss": trg_loss,
+            "arg_loss": arg_loss,
+        }
+
+# %% [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
+
+    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)
+
+    def gather(self, outputs, output_device):
+        return self._custom_gather(outputs, output_device)
+
+    def _custom_gather(self, outputs, output_device):
+        first = outputs[0]
+        
+        if torch.is_tensor(first):
+            return self._gather_tensor(outputs, output_device)
+
+        if isinstance(first, tuple):
+            return tuple(
+                self._custom_gather(
+                    list(items),
+                    output_device
+                )
+                for items in zip(*outputs)
+            )
+
+        if isinstance(first, list):
+            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
+
+        if isinstance(first, dict):
+            return {
+                k: self._custom_gather(
+                    [o[k] for o in outputs],
+                    output_device
+                )
+                for k in first.keys()
+            }
+        return outputs
+
+    def _gather_tensor(self, tensors, output_device):
+        tensors = [
+            t.to(output_device)
+            for t in tensors
+        ]
+
+        try:
+            return torch.cat(tensors, dim=0)
+        except RuntimeError:
+            pass
+
+        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])
+
+        padded = []
+        for t in tensors:
+            pad = []
+            
+            for d in reversed(range(len(max_shape))):
+                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)
+
+    def _gather_tensor_list(self, outputs, output_device):
+        merged = []
+
+        for out in outputs:
+            merged.extend(out)
+
+        return self._gather_tensor(merged, output_device)
+
+# %% [code]
+def extract_trgs(trg_logits, all_reprs, all_spans, trg_spans=None):
+    """
+    Args:
+        trg_logits: (B, N, C)
+        all_reprs:  (B, N, H)
+        all_spans:  (B, N, 2)
+        trg_spans:  None or (B, K, 2)
+
+    Returns:
+        trg_reprs: (B, K, H)
+        trg_spans: (B, K, 2)
+    """
+
+    B, N, H = all_reprs.shape
+    
+    if trg_spans is None:
+
+        trg_labels = torch.argmax(trg_logits, dim=-1)
+
+        mask = trg_labels > 0
+        counts = mask.sum(dim=1)
+        K = counts.max().item()
+
+        out_reprs = all_reprs.new_zeros((B, K, H))
+        out_spans = all_spans.new_zeros((B, K, 2))
+
+        for b in range(B):
+            idx = mask[b].nonzero(as_tuple=False).squeeze(-1)
+
+            k = idx.numel()
+            if k > 0:
+                out_reprs[b, :k] = all_reprs[b, idx]
+                out_spans[b, :k] = all_spans[b, idx]
+        
+    else:
+        _, K, _ = trg_spans.shape
+
+        out_reprs = all_reprs.new_zeros((B, K, H))
+        out_spans = all_spans.new_zeros((B, K, 2))
+
+        for b in range(B):
+            for k in range(K):
+                span = trg_spans[b, k]
+                if torch.all(span == 0):
+                    continue
+                    
+                match = (all_spans[b] == span).all(dim=-1)
+                idx = match.nonzero(as_tuple=False).squeeze(-1)
+                if idx.numel() > 0:
+                    idx = idx[0]
+                    out_reprs[b, k] = all_reprs[b, idx]
+                    out_spans[b, k] = all_spans[b, idx]
+
+    return out_reprs, out_spans
+
+class SpanExtractor(nn.Module):
+    def __init__(self, hidden_size):
+        super().__init__()
+
+        self.start_proj = MLP(hidden_size, hidden_size, hidden_size)
+        self.end_proj = MLP(hidden_size, hidden_size, hidden_size)
+
+        self.span_attn = nn.Sequential(
+            nn.Linear(hidden_size, hidden_size),
+            nn.GELU(),
+            nn.Linear(hidden_size, 1)
+        )
+
+    def forward(self, hidden_states, spans):
+        B, L, H = hidden_states.shape
+        N = spans.size(1)
+
+        start_hidden = self.start_proj(hidden_states)
+        end_hidden = self.end_proj(hidden_states)
+
+        batch_idx = torch.arange(B, device=hidden_states.device).unsqueeze(1)
+        start_idx = spans[..., 0]
+        end_idx = spans[..., 1]
+
+        start_h = start_hidden[batch_idx, start_idx]
+        end_h = end_hidden[batch_idx, end_idx]
+
+        token_idx = torch.arange(L, device=hidden_states.device).view(1, 1, L)
+        span_mask = (token_idx >= start_idx.unsqueeze(-1)) & (token_idx <= end_idx.unsqueeze(-1))
+
+        attn_scores = self.span_attn(hidden_states).squeeze(-1).unsqueeze(1).expand(-1, N, -1)
+        attn_scores = attn_scores.masked_fill(~span_mask, float('-inf'))
+        attn_weights = torch.softmax(attn_scores, dim=-1)
+        span_context = torch.einsum("bnl,blh->bnh", attn_weights, hidden_states)
+
+        span_repr = torch.cat([start_h, end_h, end_h - start_h, end_h * start_h, span_context], dim=-1)
+        return span_repr
+
+class MLP(nn.Module):
+    def __init__(self, in_size, hid_size, out_size, dropout=0.1):
+        super().__init__()
+
+        self.input_proj = nn.Identity() if in_size == hid_size else nn.Linear(in_size, hid_size)
+
+        self.block = nn.Sequential(
+            nn.Linear(hid_size, hid_size),
+            nn.LayerNorm(hid_size),
+            nn.GELU(),
+            nn.Dropout(dropout),
+
+            nn.Linear(hid_size, hid_size),
+            nn.LayerNorm(hid_size),
+            nn.GELU(),
+            nn.Dropout(dropout),
+        )
+
+        self.out = nn.Linear(hid_size, out_size)
+
+    def forward(self, x):
+        x = self.input_proj(x)
+        x = x + self.block(x)   # residual
+        return self.out(x)
+
+class IEModel(nn.Module):
+    def __init__(self, backbone_model_name, num_trg_labels, num_arg_labels):
+        super().__init__()
+        self.num_arg_labels = num_arg_labels
+
+        self.encoder = AutoModel.from_pretrained(backbone_model_name)
+        hidden_size = self.encoder.config.hidden_size
+
+        self.trg_extractor = SpanExtractor(hidden_size)
+        self.trg_classifier = MLP(5 * hidden_size, hidden_size, num_trg_labels)
+        
+        self.trg_hidden_state_proj = MLP(6 * hidden_size, 3 * hidden_size, hidden_size)
+        self.arg_extractor = SpanExtractor(hidden_size)
+        self.arg_classifier = MLP(5 * hidden_size, hidden_size, num_arg_labels)
+
+    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
+
+        hidden_states = hidden_states.view(B, n_parts, L, -1).reshape(B, n_parts * L, -1)
+        return hidden_states
+
+    def get_trg_logits(self, trg_reprs):
+        return self.trg_classifier(trg_reprs)
+
+    def get_arg_logits(self, arg_reprs):
+        return self.arg_classifier(arg_reprs)
+
+    def make_trg_hidden_states(self, hidden_states, trg_reprs):
+        B, L, H = hidden_states.shape
+        _, N, _ = trg_reprs.shape
+        
+        hidden_exp = hidden_states.unsqueeze(1).expand(-1, N, -1, -1)
+        trgs_exp = trg_reprs.unsqueeze(2).expand(-1, -1, L, -1)
+        
+        x = torch.cat([hidden_exp, trgs_exp], dim=-1) # B, N, L, 6H
+        trg_hidden_states = self.trg_hidden_state_proj(x) # B, N, L, H
+        return trg_hidden_states
+
+    def forward(self, input_ids, attention_mask, all_spans, trg_spans=None):
+        hidden_states = self.encode(input_ids, attention_mask)
+
+        all_reprs = self.trg_extractor(hidden_states, all_spans)
+        trg_logits = self.get_trg_logits(all_reprs)
+
+        trg_reprs, trg_spans = extract_trgs(trg_logits, all_reprs, all_spans, trg_spans)
+        trg_hidden_states = self.make_trg_hidden_states(hidden_states, trg_reprs)  # B, K, L, H
+        B, K, L, H = trg_hidden_states.shape
+        trg_hidden_states = trg_hidden_states.reshape(-1, L, H)
+
+        N = all_spans.size(1)
+        expanded_all_spans = all_spans.unsqueeze(1)
+        expanded_all_spans = expanded_all_spans.expand(-1, K, -1, -1)
+        expanded_all_spans = expanded_all_spans.reshape(B * K, N, 2)  # (B*K, N, 2)
+        arg_reprs = self.arg_extractor(trg_hidden_states, expanded_all_spans)
+        arg_logits = self.get_arg_logits(arg_reprs).reshape(B, K, N, self.num_arg_labels)
+        
+        return trg_logits, arg_logits, trg_spans
+
+def test_model():
+    model = nn.DataParallel(IEModel(backbone_model_name, 17, 23)).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)
+    s = torch.ones(bz, 3, 2, dtype=torch.long).to(device)
+    gs = torch.ones(bz, 3, 2, dtype=torch.long).to(device)
+
+    with torch.no_grad():
+        r = model(i, a, s)
+
+    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(arg_obj, trg_gold_spans, pred_trg_spans):
+    """
+    Args:
+        arg_obj:         (B, N, M)
+        trg_gold_spans: (B, N, 2)
+        pred_trg_spans: (B, K, 2)
+
+    Returns:
+        arg_align: (B, K, M)
+    """
+
+    B, N, M = arg_obj.shape
+    K = pred_trg_spans.size(1)
+
+    arg_align = torch.zeros(B, K, M, dtype=arg_obj.dtype, device=arg_obj.device)
+    
+    match = (trg_gold_spans.unsqueeze(2) == pred_trg_spans.unsqueeze(1)).all(dim=-1)  # B,N,K    
+    b_idx, n_idx, k_idx = torch.nonzero(match, as_tuple=True)
+    arg_align[b_idx, k_idx] = arg_obj[b_idx, n_idx]
+
+    return arg_align
+
+def extract_events(
+    input_ids,         # (B, L)
+    all_spans,         # (B, N, 2)
+    trg_logits,        # (B, N, C_trg)
+    arg_logits,        # (B, K, N, C_arg)
+    pred_gold_trgs,    # (B, K, 2)
+    id2label
+):
+    """
+    Return: [(bidx, (trg_token_ids, trg_label_name), (arg_token_ids, arg_label_name)),...]
+    """
+
+    results = []
+
+    trg_labels = trg_logits.argmax(dim=-1)
+    arg_labels = arg_logits.argmax(dim=-1)
+
+    B, N, _ = all_spans.shape
+    K = pred_gold_trgs.size(1)
+
+    for bidx in range(B):
+
+        keep = trg_labels[bidx] > 0
+        trg_spans = all_spans[bidx, keep]       # (K', 2)
+        trg_lbs   = trg_labels[bidx, keep]      # (K',)
+        
+        trg_map = {}
+        for span, lb in zip(trg_spans, trg_lbs):
+            s, e = span.tolist()
+            if s == 0 or e == 0:
+                continue
+
+            trg_map[(s, e)] = lb.item()
+            
+        for k in range(K):
+            trg_s, trg_e = pred_gold_trgs[bidx, k].tolist()
+            if trg_s == 0 or trg_e == 0:
+                continue
+
+            key = (trg_s, trg_e)
+            if key not in trg_map:
+                continue
+
+            trg_lb = trg_map[key]
+            trg_token_ids = input_ids[bidx, trg_s:trg_e+1].tolist()
+
+            cur_arg_labels = arg_labels[bidx, k]   # (N,)
+            for n in range(N):
+                arg_lb = cur_arg_labels[n].item()
+                if arg_lb <= 0:
+                    continue
+
+                arg_s, arg_e = all_spans[bidx, n].tolist()
+                if arg_s == 0 or arg_e == 0:
+                    continue
+
+                arg_token_ids = input_ids[bidx, arg_s:arg_e+1].tolist()
+                results.append((bidx, (tuple(trg_token_ids), id2label['Trg'][trg_lb]), (tuple(arg_token_ids), id2label['Arg'][arg_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")
+                        
+                    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)
+        sampled_spans = batch['sampled_spans'].to(self.device) # B, M, 2
+        sampled_trg_labels = batch['sampled_trg_labels'].to(self.device) # B, N, 2
+        sampled_trg_weights = batch['sampled_trg_weights'].to(self.device) # B, N
+        sampled_arg_labels = batch['sampled_arg_labels'].to(self.device) # B, N, 2
+        sampled_arg_weights = batch['sampled_arg_weights'].to(self.device) # B, N
+        trg_gold_spans = batch['trg_gold_spans'].to(self.device)
+
+        choice = random.random()        
+        if choice < teaching_rate:
+            trg_logits, arg_logits, pred_trg_spans = network(input_ids, attention_mask, sampled_spans, trg_spans=trg_gold_spans)
+        else:
+            trg_logits, arg_logits, pred_trg_spans = network(input_ids, attention_mask, sampled_spans, trg_spans=None)
+
+        sampled_arg_labels = align(sampled_arg_labels, trg_gold_spans, pred_trg_spans)
+        sampled_arg_weights = align(sampled_arg_weights, trg_gold_spans, pred_trg_spans)
+        
+        loss_dict = loss_fn(
+            trg_logits, sampled_trg_labels, sampled_trg_weights, 
+            arg_logits, sampled_arg_labels, sampled_arg_weights, 
+        )
+        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)
+        all_spans = batch['all_spans'].to(self.device) # B, N, 2
+        gold_events = batch['gold_events']
+
+        B, _, _ = input_ids.shape
+        
+        trg_logits, arg_logits, trg_spans = network(input_ids, attention_mask, all_spans, trg_spans=None)
+                
+        pred_ids = extract_events(input_ids.reshape(B, -1), all_spans, trg_logits, arg_logits, trg_spans, id2label)
+        pred_ids = list_to_tuple(pred_ids)
+            
+        gold_ids = list_to_tuple(gold_events)
+        
+        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):
+        trigger_spans, arg_spans = [], []
+        
+        for event in sample['events']:
+            trigger_type = event["label"]
+            spans = [tuple(event["offset"])]
+            trigger_spans.append({
+                "spans": spans,
+                "label": trigger_type
+            })
+            event_arg_spans = []
+            for arg in event['arguments']:
+                arg_type = arg["role"]
+                spans = [tuple(arg["offset"])]
+                event_arg_spans.append({
+                    "spans": spans,
+                    "label": arg_type
+                })
+            arg_spans.append(event_arg_spans)
+        
+        return trigger_spans, arg_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"]
+        triggers, arguments = 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 =====
+        triggers_gold_spans = []
+        arguments_gold_spans = []
+    
+        for trg, args in zip(triggers, arguments):
+            label = trg["label"]
+            
+            sub_spans = self.span_to_subword(
+                word_offsets,
+                word_subword_map,
+                trg["spans"]
+            )
+            valid_spans = self.extract_valid_spans(sub_spans)
+            if len(valid_spans) == 0:
+                continue    
+            triggers_gold_spans.append((tuple(valid_spans), label))
+
+            trg_args_gold_spans = []
+            for arg in args:
+                label = arg["label"]
+                
+                sub_spans = self.span_to_subword(
+                    word_offsets,
+                    word_subword_map,
+                    arg["spans"]
+                )
+                valid_spans = self.extract_valid_spans(sub_spans)
+                if len(valid_spans) == 0:
+                    continue    
+                trg_args_gold_spans.append((tuple(valid_spans), label))
+            arguments_gold_spans.append(tuple(trg_args_gold_spans))
+                
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "triggers_gold_spans": triggers_gold_spans,
+            "arguments_gold_spans": arguments_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, max_span_len, n_negs, 
+        weight_sampling=False, weight_rate=0.0, hard_rate=0.5, 
+    ):
+        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
+        self.max_span_len = max_span_len
+
+        self.n_negs = n_negs
+        self.weight_sampling = weight_sampling
+        self.weight_rate = weight_rate
+        self.hard_rate = hard_rate
+
+    def __len__(self):
+        return len(self.using_idxes)
+
+    def compute_iou(self, spans1, spans2):
+        s1 = spans1[:, None, 0]
+        e1 = spans1[:, None, 1]
+
+        s2 = spans2[None, :, 0]
+        e2 = spans2[None, :, 1]
+
+        inter = (torch.minimum(e1, e2) - torch.maximum(s1, s2) + 1).clamp(min=0)
+
+        len1 = e1 - s1 + 1
+        len2 = e2 - s2 + 1
+
+        union = len1 + len2 - inter
+
+        return inter.float() / union.float()
+
+    def sample_spans(
+        self,
+        all_spans,
+        trg_all_labels,
+        trg_arg_all_labels,
+        gold_spans,
+        triggers_gold_spans,
+        arguments_gold_spans,
+    ):
+        M = len(all_spans)
+    
+        # ===== sample indices =====
+    
+        global_pos_mask = trg_all_labels != 0
+    
+        for i in range(len(trg_arg_all_labels)):
+            global_pos_mask |= (trg_arg_all_labels[i] != 0)
+    
+        global_neg_mask = ~global_pos_mask
+    
+        pos_indices = torch.nonzero(global_pos_mask, as_tuple=False).squeeze(-1)
+        neg_indices = torch.nonzero(global_neg_mask, as_tuple=False).squeeze(-1)
+    
+        n_negs = min(self.n_negs, len(neg_indices))
+    
+        if self.weight_sampling and n_negs > 0 and len(neg_indices) > 0:
+            neg_spans = all_spans[neg_indices]
+    
+            max_ious = (
+                self.compute_iou(neg_spans, gold_spans).max(dim=1).values
+                if len(gold_spans) > 0 else
+                torch.zeros(len(neg_indices))
+            )
+    
+            hard_mask = max_ious > 0
+            easy_mask = max_ious == 0
+    
+            hard_indices = neg_indices[hard_mask]
+            easy_indices = neg_indices[easy_mask]
+    
+            hard_ious = max_ious[hard_mask]
+    
+            hard_k = min(int(self.hard_rate * n_negs), len(hard_indices))
+            easy_k = min(n_negs - hard_k, len(easy_indices))
+    
+            sampled_hard = torch.empty(0, dtype=torch.long)
+            sampled_easy = torch.empty(0, dtype=torch.long)
+    
+            if hard_k > 0:
+                probs = hard_ious / hard_ious.sum()
+    
+                hard_ids = torch.multinomial(probs, hard_k, replacement=False)
+    
+                sampled_hard = hard_indices[hard_ids]
+    
+            if easy_k > 0:
+                sampled_easy = easy_indices[torch.randperm(len(easy_indices))[:easy_k]]
+    
+            neg_indices = torch.cat([sampled_hard, sampled_easy], dim=0)
+    
+        elif n_negs > 0:
+            neg_indices = neg_indices[torch.randperm(len(neg_indices))[:n_negs]]
+    
+        sampled_indices = torch.cat([pos_indices, neg_indices], dim=0)
+    
+        if len(sampled_indices) > 0:
+            sampled_indices = sampled_indices[torch.randperm(len(sampled_indices))]
+    
+        K = len(sampled_indices)
+    
+        # ===== trigger =====
+    
+        sampled_trg_spans = all_spans[sampled_indices]
+    
+        sampled_trg_labels = trg_all_labels[sampled_indices]
+    
+        sampled_trg_weights = torch.ones(K, dtype=torch.float)
+    
+        trg_gold_spans = [list(trg_spans[0]) for trg_spans, _ in triggers_gold_spans]
+    
+        trg_gold_spans = (
+            torch.tensor(trg_gold_spans, dtype=torch.long)
+            if len(trg_gold_spans) > 0 else
+            torch.empty(0, 2, dtype=torch.long)
+        )
+    
+        if len(trg_gold_spans) > 0:
+            trg_max_ious = self.compute_iou(sampled_trg_spans, trg_gold_spans).max(dim=1).values
+    
+            sampled_trg_weights = 1.0 + self.weight_rate * trg_max_ious
+    
+        # ===== argument =====
+    
+        T = len(triggers_gold_spans)
+    
+        sampled_arg_spans = sampled_trg_spans.unsqueeze(0).expand(T, -1, -1).clone()
+    
+        sampled_arg_labels = torch.zeros(T, K, dtype=torch.long)
+    
+        sampled_arg_weights = torch.ones(T, K, dtype=torch.float)
+    
+        for tidx, args in enumerate(arguments_gold_spans):
+            sampled_arg_labels[tidx] = trg_arg_all_labels[tidx][sampled_indices]
+    
+            arg_gold_spans = [list(arg_spans[0]) for arg_spans, _ in args]
+    
+            arg_gold_spans = (
+                torch.tensor(arg_gold_spans, dtype=torch.long)
+                if len(arg_gold_spans) > 0 else
+                torch.empty(0, 2, dtype=torch.long)
+            )
+    
+            if len(arg_gold_spans) > 0:
+                arg_max_ious = self.compute_iou(
+                    sampled_arg_spans[tidx],
+                    arg_gold_spans
+                ).max(dim=1).values
+    
+                sampled_arg_weights[tidx] = 1.0 + self.weight_rate * arg_max_ious
+    
+        return (
+            sampled_trg_spans,
+            sampled_trg_labels,
+            sampled_trg_weights,
+            sampled_arg_labels,
+            sampled_arg_weights,
+        )
+
+    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)    
+        triggers_gold_spans = result["triggers_gold_spans"]
+        arguments_gold_spans = result["arguments_gold_spans"]
+    
+        all_spans = torch.tensor(generate_spans(attention_mask, self.max_span_len), dtype=torch.long)
+    
+        trg_gold_spans = []
+        trg_gold_labels = []    
+        gold_spans = []    
+        for (trg_spans, trg_label), args in zip(triggers_gold_spans, arguments_gold_spans):    
+            trg_span = list(trg_spans[0])    
+            trg_gold_spans.append(trg_span)
+            gold_spans.append(trg_span)    
+            trg_gold_labels.append(self.label2id['Trg'][trg_label])
+    
+            for arg_spans, arg_label in args:    
+                arg_span = list(arg_spans[0])    
+                gold_spans.append(arg_span)
+    
+        trg_gold_spans = torch.tensor(trg_gold_spans, dtype=torch.long) if len(trg_gold_spans) > 0 else torch.empty(0, 2, dtype=torch.long)
+        trg_gold_labels = torch.tensor(trg_gold_labels, dtype=torch.long) if len(trg_gold_labels) > 0 else torch.empty(0, dtype=torch.long)
+        gold_spans = torch.tensor(gold_spans, dtype=torch.long) if len(gold_spans) > 0 else torch.empty(0, 2, dtype=torch.long)    
+        trg_all_labels = match_gold_labels(
+            trg_gold_spans,
+            trg_gold_labels,
+            all_spans,
+            default_label=0
+        )
+    
+        N, M = len(triggers_gold_spans), len(all_spans)
+        trg_arg_all_labels = torch.zeros(N, M, dtype=torch.long)        
+        for tidx, ((trg_spans, trg_label), args) in enumerate(zip(triggers_gold_spans, arguments_gold_spans)):
+            arg_gold_spans = [list(arg_spans[0]) for arg_spans, _ in args]
+            arg_gold_labels = [self.label2id['Arg'][arg_label] for _, arg_label in args]
+        
+            arg_gold_spans = torch.tensor(arg_gold_spans, dtype=torch.long) if len(arg_gold_spans) > 0 else torch.empty(0, 2, dtype=torch.long)
+            arg_gold_labels = torch.tensor(arg_gold_labels, dtype=torch.long) if len(arg_gold_labels) > 0 else torch.empty(0, dtype=torch.long)
+        
+            trg_arg_all_labels[tidx] = match_gold_labels(
+                arg_gold_spans,
+                arg_gold_labels,
+                all_spans,
+                default_label=0
+            )
+    
+        sampled_spans, sampled_trg_labels, sampled_trg_weights, sampled_arg_labels, sampled_arg_weights, = self.sample_spans(
+            all_spans,
+            trg_all_labels,
+            trg_arg_all_labels,
+            gold_spans,
+            triggers_gold_spans, 
+            arguments_gold_spans, 
+        )
+    
+        gold_events = []    
+        for (trg_spans, trg_label), args in zip(triggers_gold_spans, arguments_gold_spans):
+            s, e = trg_spans[0]
+            event = [(tuple(input_ids[s:e+1].tolist()), trg_label)]    
+            for arg_spans, arg_label in args:
+                s, e = arg_spans[0]    
+                event.append((tuple(input_ids[s:e+1].tolist()), arg_label))
+            gold_events.append(event)
+    
+        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]
+    
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+    
+            "sampled_spans": sampled_spans,
+            "sampled_trg_labels": sampled_trg_labels,
+            "sampled_trg_weights": sampled_trg_weights,
+            "sampled_arg_labels": sampled_arg_labels,
+            "sampled_arg_weights": sampled_arg_weights,
+    
+            "all_spans": all_spans,    
+            "trg_gold_spans": trg_gold_spans,
+    
+            "gold_events": gold_events,
+        }
+
+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_events = []
+    for bidx, b in enumerate(batch):
+        for event in b['gold_events']:
+            trg = event[0]
+            if len(event) > 1:
+                for arg in event[1:]:
+                    gold_events.append([bidx, trg, arg])
+            else:
+                gold_events.append([bidx, trg, (tuple([]), 0)])
+
+    input_ids = [b["input_ids"].unsqueeze(-1) for b in batch]
+    attention_mask = [b["attention_mask"].unsqueeze(-1) for b in batch]
+    
+    sampled_spans = [b["sampled_spans"].unsqueeze(-1) for b in batch]
+    sampled_trg_labels = [b["sampled_trg_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    sampled_trg_weights = [b["sampled_trg_weights"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    sampled_arg_labels = [b["sampled_arg_labels"].unsqueeze(-1) for b in batch]
+    sampled_arg_weights = [b["sampled_arg_weights"].unsqueeze(-1) for b in batch]
+    
+    all_spans = [b["all_spans"].unsqueeze(-1) for b in batch]    
+    trg_gold_spans = [b["trg_gold_spans"].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)
+    
+    sampled_spans = _pad_batch(sampled_spans, pad_value=0).squeeze(-1)
+    sampled_trg_labels = _pad_batch(sampled_trg_labels, pad_value=-100).squeeze(-1).squeeze(-1)
+    sampled_trg_weights = _pad_batch(sampled_trg_weights, pad_value=0).squeeze(-1).squeeze(-1)
+    sampled_arg_labels = _pad_batch(sampled_arg_labels, pad_value=-100).squeeze(-1)
+    sampled_arg_weights = _pad_batch(sampled_arg_weights, pad_value=0).squeeze(-1)
+    
+    all_spans = _pad_batch(all_spans, pad_value=0).squeeze(-1)
+    trg_gold_spans = _pad_batch(trg_gold_spans, pad_value=0).squeeze(-1)
+    
+    return {
+        "input_ids": input_ids,
+        "attention_mask": attention_mask,
+
+        "sampled_spans": sampled_spans,
+        "sampled_trg_labels": sampled_trg_labels,
+        "sampled_trg_weights": sampled_trg_weights,
+        "sampled_arg_labels": sampled_arg_labels,
+        "sampled_arg_weights": sampled_arg_weights,
+
+        "all_spans": all_spans,    
+        "trg_gold_spans": trg_gold_spans,
+
+        "gold_events": gold_events,
+    }
+
+# %% [code]
+def shift_bidx(spans, batch_idx):
+    shifted = []
+    for bidx, trg, arg in spans:
+        new_bidx = bidx + batch_idx * (batch_size // 4)
+        shifted.append((new_bidx, trg, arg))
+    return shifted
+
+def refactor_events(events, save_dict):
+    trg_i, trg_c, arg_i, arg_c, soft, strict_dict = [], [], [], [], [], {}
+    for bidx, (trg_ids, trg_lb), (arg_k_ids, arg_k_lb) in events:
+        if (bidx, trg_ids) not in trg_i:
+            trg_i.append((bidx, trg_ids))
+            
+        if (bidx, (trg_ids, trg_lb)) not in trg_c:
+            trg_c.append((bidx, (trg_ids, trg_lb)))
+            
+        if (bidx, trg_ids, arg_k_ids) not in arg_i:
+            arg_i.append((bidx, trg_ids, arg_k_ids))
+            
+        if (bidx, trg_ids, (arg_k_ids, arg_k_lb)) not in arg_c:
+            arg_c.append((bidx, trg_ids, (arg_k_ids, arg_k_lb)))
+
+        if (bidx, (trg_ids, trg_lb), (arg_k_ids, arg_k_lb)) not in soft:
+            soft.append((bidx, (trg_ids, trg_lb), (arg_k_ids, arg_k_lb)))
+
+        if bidx not in strict_dict:
+            strict_dict[bidx] = {}
+        if (trg_ids, trg_lb) not in strict_dict[bidx]:
+            strict_dict[bidx][(trg_ids, trg_lb)] = []
+        strict_dict[bidx][(trg_ids, trg_lb)].append((arg_k_ids, arg_k_lb))
+
+    strict = []
+    for bidx, trg_dict in strict_dict.items():
+        for trg, args in trg_dict.items():
+            strict.append((bidx, trg, frozenset(args)))
+
+    save_dict['Trg-I'].extend(trg_i)
+    save_dict['Trg-C'].extend(trg_c)
+    save_dict['Arg-I'].extend(arg_i)
+    save_dict['Arg-C'].extend(arg_c)
+    save_dict['Soft-Event'].extend(soft)
+    save_dict['Strict-Event'].extend(strict)
+
+def test(network, state_dicts, test_loader, eval_fn, analyzer, device, id2label, tokenizer):
+    network = network.to(device)
+    network.eval()
+
+    eval_types = ['Trg-I', 'Trg-C', 'Arg-I', 'Arg-C', 'Soft-Event', 'Strict-Event']
+    
+    all_pred = {eval_type: [] for eval_type in eval_types}
+    all_gold = {eval_type: [] for eval_type in eval_types}
+
+    list_input_ids = []
+
+    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) # B, N, 2
+            gold_events = batch['gold_events']
+
+            B, _, _ = input_ids.shape
+            list_input_ids.extend(input_ids.reshape(B, -1).tolist())
+            
+            list_trg_logits = []
+            list_arg_logits = []
+            list_hidden_states = []
+            list_all_reprs = []
+            for sd in state_dicts:
+                network.load_state_dict(sd)
+                
+                hidden_states = network.encode(input_ids, attention_mask)
+                all_reprs = network.trg_extractor(hidden_states, all_spans)
+                trg_logits = network.get_trg_logits(all_reprs)
+                
+                list_hidden_states.append(hidden_states)
+                list_all_reprs.append(all_reprs)
+                list_trg_logits.append(trg_logits)
+                
+            ensemble_trg_logits = torch.stack(list_trg_logits, dim=0).mean(dim=0)
+            
+            for sd, hidden_states, all_reprs in zip(state_dicts, list_hidden_states, list_all_reprs):
+                network.load_state_dict(sd)
+                    
+                trg_reprs, trg_spans = extract_trgs(ensemble_trg_logits, all_reprs, all_spans)
+                trg_hidden_states = network.make_trg_hidden_states(hidden_states, trg_reprs)  # B, K, L, H
+                B, K, L, H = trg_hidden_states.shape
+                trg_hidden_states = trg_hidden_states.reshape(-1, L, H)
+        
+                N = all_spans.size(1)
+                expanded_all_spans = all_spans.unsqueeze(1).expand(-1, K, -1, -1).reshape(B * K, N, 2)  # (B*K, N, 2)
+                arg_reprs = network.arg_extractor(trg_hidden_states, expanded_all_spans)
+                arg_logits = network.get_arg_logits(arg_reprs).reshape(B, K, N, network.num_arg_labels)
+                
+                list_arg_logits.append(arg_logits)
+            ensemble_arg_logits = torch.stack(list_arg_logits, dim=0).mean(dim=0)
+
+            pred_events = extract_events(input_ids.reshape(B, -1), all_spans, ensemble_trg_logits, ensemble_arg_logits, trg_spans, id2label)
+            pred_events = shift_bidx(pred_events, batch_idx)
+            refactor_events(pred_events, all_pred)
+                
+            gold_events = shift_bidx(gold_events, batch_idx)
+            refactor_events(gold_events, all_gold)
+
+    # ===== 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['Trg-I']),  list_to_tuple(all_gold['Trg-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 event in all_pred['Strict-Event']:
+        bidx = event[0]
+        trg = tokenizer.decode(event[1][0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
+        trg_lb = event[1][1]
+        predictions[bidx].append((trg, trg_lb))
+
+        for arg_infor in event[2]:
+            arg = tokenizer.decode(arg_infor[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
+            arg_lb = arg_infor[1]
+            predictions[bidx].append((arg, arg_lb))            
+
+    return final_score, analyze_result, predictions
+
+# %% [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]
+trigger_types = sorted(list(set([e['label'] for d in data_train + data_test for e in d['events']]))) # NBR : Neighbor relation
+# bio_trigger_types = [f'{prefix}-{trg}' for trg in trigger_types for prefix in ['B', 'I']]
+trigger_label2id = {l: i for i, l in enumerate(['O'] + trigger_types)}
+trigger_id2label = {i: l for l, i in trigger_label2id.items()}
+
+argument_types = sorted(list(set([a['role'] for d in data_train + data_test for e in d['events'] for a in e['arguments']])))
+# bio_argument_types = [f'{prefix}-{arg}' for arg in argument_types for prefix in ['B', 'I']]
+argument_label2id = {l: i for i, l in enumerate(['O'] + argument_types)}
+argument_id2label = {i: l for l, i in argument_label2id.items()}
+
+label2id = {
+    'Trg': trigger_label2id, 
+    'Arg': argument_label2id, 
+    'All': {l: i for i, l in enumerate(trigger_types + argument_types)}, 
+}
+
+id2label = {
+    'Trg': trigger_id2label, 
+    'Arg': argument_id2label, 
+}
+
+# %% [code]
+zero_events_idxes = []
+for idx, d in enumerate(data_train):
+    if len(d['events']) == 0:
+        zero_events_idxes.append(idx)
+
+n_zero_events_samples = len(zero_events_idxes)
+n_has_events_samples = len(data_train) - n_zero_events_samples
+
+random.seed(42)
+k = min(int(n_has_events_samples * zero_events_rate), len(zero_events_idxes))
+sampled_zero_events_idxes = random.sample(zero_events_idxes, k)
+
+new_data_train = []
+for idx, d in enumerate(data_train):
+    if len(d['events']) == 0:
+        if idx in sampled_zero_events_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[:20]
+    data_test = data_test[:20]
+
+    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]
+# trainset = KLTNDataset(data_train, np.array(range(len(data_train))), label2id, tokenizer, **train_memory_params)
+# train_loader = DataLoader(trainset, collate_fn=collate_fn, **train_loader_params)
+# for b in train_loader:
+#     break
+
+# %% [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_trg_labels=len(trigger_label2id), 
+                num_arg_labels=len(argument_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_trg_labels=len(trigger_label2id), 
+    num_arg_labels=len(argument_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()
+
+best_score, best_analyze_result, best_pred_test = 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]
+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
+

0_act_175negs_hr_05_wr10_1/lasts/0_act_175negs_hr_05_wr10_1_s26092004_f0_last_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:721f6bc39266f9a64201b51a2295f04538e289a1fe7ddc5fe9ecd3d55abc17a3
+size 698445650

0_act_175negs_hr_05_wr10_1/logs/0_act_175negs_hr_05_wr10_1_s26092004_f0_logging.json

@@ -0,0 +1 @@
+{"1": {"lr": [2e-05, 0.0005], "train_loss": 0.1683495044708252, "total": 0.168349502969885, "trg_loss": 0.0400422879050488, "arg_loss": 0.1283069823036471}, "2": {"lr": [1.988303923565381e-05, 0.0004969282409784868], "train_loss": 0.0864180326461792, "total": 0.08641803732218523, "trg_loss": 0.027659025030621027, "arg_loss": 0.05875896241416654}, "3": {"lr": [1.9535036904803962e-05, 0.0004877886008156408], "train_loss": 0.07958684116601944, "total": 0.07958684371689619, "trg_loss": 0.02556420758041862, "arg_loss": 0.054022641678410636}, "4": {"lr": [1.8964561979789496e-05, 0.00047280612778499774], "train_loss": 0.07644027471542358, "total": 0.07644027451337394, "trg_loss": 0.0240064607116847, "arg_loss": 0.0524337731608467}, "5": {"lr": [1.8185661446562005e-05, 0.00045234974009654937], "train_loss": 0.07008393108844757, "total": 0.07008393590155872, "trg_loss": 0.021873354045877157, "arg_loss": 0.048210613259968976}, "6": {"lr": [1.7217514421272206e-05, 0.00042692314190604356], "train_loss": 0.06427405774593353, "total": 0.06427406144777165, "trg_loss": 0.02012924305174599, "arg_loss": 0.044144910761576873}, "7": {"lr": [1.60839598967785e-05, 0.00039715242044697206], "train_loss": 0.05860750004649162, "total": 0.058607501383266494, "trg_loss": 0.018493135096663137, "arg_loss": 0.040114340424249016}, "8": {"lr": [1.4812909747525698e-05, 0.00036377062968501693], "train_loss": 0.05323947221040726, "total": 0.05323947416956719, "trg_loss": 0.017008672903582777, "arg_loss": 0.036230790182118264, "val_score": 0.3268183468976117, "best_score": 0.3268183468976117, "new_best_model": true, "precision": 0.24422334690785708, "recall": 0.5473807861978923, "f1": 0.3268183468976117}, "9": {"lr": [1.3435661446562005e-05, 0.0003275997400965494], "train_loss": 0.04819445312023163, "total": 0.048194452983126516, "trg_loss": 0.015561810879095415, "arg_loss": 0.03263260515781061, "val_score": 0.33054177588243067, "best_score": 0.33054177588243067, "new_best_model": true, "precision": 0.2491534300095187, "recall": 0.5446608899063186, "f1": 0.33054177588243067}, "10": {"lr": [1.1986127417882198e-05, 0.00028953039902753766], "train_loss": 0.04337947443127632, "total": 0.04337947743856878, "trg_loss": 0.013887331503066832, "arg_loss": 0.02949220412579922, "val_score": 0.33296016527280126, "best_score": 0.33296016527280126, "new_best_model": true, "precision": 0.2534141315431134, "recall": 0.537972140776664, "f1": 0.33296016527280126}, "11": {"lr": [1.0500000000000003e-05, 0.0002505], "train_loss": 0.03928280249238014, "total": 0.039282802923539646, "trg_loss": 0.012691269659822847, "arg_loss": 0.026591571133592805, "val_score": 0.3330303048057699, "best_score": 0.3330303048057699, "new_best_model": true, "precision": 0.2559073449901142, "recall": 0.5268826337139572, "f1": 0.3330303048057699}, "12": {"lr": [9.013872582117811e-06, 0.00021146960097246246], "train_loss": 0.035227641463279724, "total": 0.03522764118185343, "trg_loss": 0.011401450316496102, "arg_loss": 0.023826172392247085, "val_score": 0.3347087695549201, "best_score": 0.3347087695549201, "new_best_model": true, "precision": 0.26091450467470556, "recall": 0.5144526919959698, "f1": 0.3347087695549201}, "13": {"lr": [7.564338553438001e-06, 0.00017340025990345064], "train_loss": 0.03177306801080704, "total": 0.03177306981121368, "trg_loss": 0.010165936375357048, "arg_loss": 0.02160711958102395, "val_score": 0.33735392340829, "best_score": 0.33735392340829, "new_best_model": true, "precision": 0.2667827299488793, "recall": 0.5039107600446552, "f1": 0.33735392340829}, "14": {"lr": [6.1870902524743065e-06, 0.00013722937031498307], "train_loss": 0.028341669589281082, "total": 0.028341671110065452, "trg_loss": 0.009089728995039157, "arg_loss": 0.01925190886342785, "val_score": 0.33907492281281143, "best_score": 0.33907492281281143, "new_best_model": true, "precision": 0.2722737812916824, "recall": 0.4925109217079782, "f1": 0.33907492281281143}, "15": {"lr": [4.916040103221507e-06, 0.00010384757955302797], "train_loss": 0.025531968101859093, "total": 0.025531968308418772, "trg_loss": 0.008242211330312216, "arg_loss": 0.017289767138317192, "val_score": 0.34027683857494595, "best_score": 0.34027683857494595, "new_best_model": true, "precision": 0.2773332722398068, "recall": 0.4816330962326916, "f1": 0.34027683857494595}, "16": {"lr": [3.7824855787278e-06, 7.40768580939564e-05], "train_loss": 0.023026730865240097, "total": 0.02302673192058868, "trg_loss": 0.007497703884762078, "arg_loss": 0.015529035886898456, "val_score": 0.342742309690231, "best_score": 0.342742309690231, "new_best_model": true, "precision": 0.28474658890833937, "recall": 0.47179014515394285, "f1": 0.342742309690231}, "17": {"lr": [2.814338553438001e-06, 4.865025990345063e-05], "train_loss": 0.020863287150859833, "total": 0.020863286865825514, "trg_loss": 0.006822319584955026, "arg_loss": 0.014041007459885272, "val_score": 0.3452106045578796, "best_score": 0.3452106045578796, "new_best_model": true, "precision": 0.29130381182444015, "recall": 0.4630817782753243, "f1": 0.3452106045578796}, "18": {"lr": [2.0354380202105066e-06, 2.8193872215002235e-05], "train_loss": 0.019170347601175308, "total": 0.019170348234384457, "trg_loss": 0.0063104680317654736, "arg_loss": 0.012859872813374886, "val_score": 0.3469599747485892, "best_score": 0.3469599747485892, "new_best_model": true, "precision": 0.29732471297618224, "recall": 0.4550110103114192, "f1": 0.3469599747485892}}

0_act_175negs_hr_05_wr10_1/r1s/0_act_175negs_hr_05_wr10_1_s26092004_f0_r1_vs0.34696_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0136dcc633d9b5d9ab413e45ae530b0a8f51c80473211295dbb1bfa173480ae
+size 698448090

0_act_175negs_hr_05_wr10_1/results/0_act_175negs_hr_05_wr10_1_pred_test.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc0b3a3bba82a817f4ff77ced0e40d3b386345bf2fbdd065493ef8087b5b9d42
+size 13626876

0_act_175negs_hr_05_wr10_1/results/0_act_175negs_hr_05_wr10_1_test.json

@@ -0,0 +1,66 @@
+{
+  "Best model": {
+    "Trg-I": {
+      "precision": 0.21553863974829096,
+      "recall": 0.7472748135377302,
+      "f1": 0.33457487451072837
+    },
+    "Trg-C": {
+      "precision": 0.19649209894912179,
+      "recall": 0.6811012331497531,
+      "f1": 0.3049955018536409
+    },
+    "Arg-I": {
+      "precision": 0.1414511961816576,
+      "recall": 0.6162570888463513,
+      "f1": 0.23008934426280805
+    },
+    "Arg-C": {
+      "precision": 0.13469540694576002,
+      "recall": 0.5862660944200976,
+      "f1": 0.21906120901464823
+    },
+    "Soft-Event": {
+      "precision": 0.12338302571381067,
+      "recall": 0.5370815450639166,
+      "f1": 0.2006670698586206
+    },
+    "Strict-Event": {
+      "precision": 0.04451063125668527,
+      "recall": 0.15428735302508093,
+      "f1": 0.06908950467916551
+    }
+  },
+  "Last model": {
+    "Trg-I": {
+      "precision": 0.21553863974829096,
+      "recall": 0.7472748135377302,
+      "f1": 0.33457487451072837
+    },
+    "Trg-C": {
+      "precision": 0.19649209894912179,
+      "recall": 0.6811012331497531,
+      "f1": 0.3049955018536409
+    },
+    "Arg-I": {
+      "precision": 0.1414511961816576,
+      "recall": 0.6162570888463513,
+      "f1": 0.23008934426280805
+    },
+    "Arg-C": {
+      "precision": 0.13469540694576002,
+      "recall": 0.5862660944200976,
+      "f1": 0.21906120901464823
+    },
+    "Soft-Event": {
+      "precision": 0.12338302571381067,
+      "recall": 0.5370815450639166,
+      "f1": 0.2006670698586206
+    },
+    "Strict-Event": {
+      "precision": 0.04451063125668527,
+      "recall": 0.15428735302508093,
+      "f1": 0.06908950467916551
+    }
+  }
+}