Upload 6_4_actions_8's state dict

.gitattributes

@@ -66,3 +66,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 6_4_entities_8/results/6_4_entities_8_pred_test.json filter=lfs diff=lfs merge=lfs -text
 6_4_issues_8/logs/6_4_issues_8_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
 1_lr_add_info_5/logs/1_lr_add_info_5_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
+6_4_actions_8/logs/6_4_actions_8_log_plot.jpg filter=lfs diff=lfs merge=lfs -text

6_4_actions_8/6_4_actions_8.py

@@ -0,0 +1,2180 @@
+# %% [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 = 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 = "6_4_actions_8"
+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, 
+    '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]
+
+# 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, 
+}
+val_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+    'max_span_len': max_span_len, 
+}
+
+# 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, 
+    'keep_neighbor': 0, 
+}
+
+# Loss Func
+loss_func_params = {
+    'lambda_trg_ce': 1.0,
+    'lambda_arg_ce': 1.0,
+    'lambda_start_ce': 1.0, 
+    'lambda_end_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,
+        lambda_start_ce=1.0,
+        lambda_end_ce=1.0,
+    ):
+        super().__init__()
+
+        self.lambda_trg_ce = lambda_trg_ce
+        self.lambda_arg_ce = lambda_arg_ce
+        self.lambda_start_ce = lambda_start_ce
+        self.lambda_end_ce = lambda_end_ce
+
+        self.ce = nn.CrossEntropyLoss(ignore_index=-100)
+
+    def forward(
+        self,
+        trg_logits, trg_labels,
+        arg_logits, arg_labels,
+        start_logits, start_labels, 
+        end_logits, end_labels,
+    ):
+        device = trg_logits.device
+
+        # ===== TRG CE =====
+        B, N, C = trg_logits.shape
+
+        flat_logits = trg_logits.reshape(-1, C)
+        flat_labels = trg_labels.reshape(-1)
+        
+        trg_loss = (
+            self.ce(flat_logits, flat_labels)
+            if (flat_labels != -100).any()
+            else torch.tensor(0.0, device=device)
+        )
+
+        # ===== ARG CE =====
+        B, K, N, C = arg_logits.shape
+
+        arg_logits_flat = arg_logits.view(B * K * N, C)
+        arg_labels_flat = arg_labels.view(-1)
+
+        valid_arg = (arg_labels_flat != -100)
+
+        if valid_arg.any():
+            arg_loss = self.ce(
+                arg_logits_flat,
+                arg_labels_flat
+            )
+        else:
+            arg_loss = torch.tensor(0.0, device=device)
+
+        # ===== TRG CE =====
+        B, L, C = start_logits.shape
+        
+        start_logits_flat = start_logits.view(B * L, C)
+        start_labels_flat = start_labels.view(-1)
+        start_loss = self.ce(start_logits_flat, start_labels_flat)  # (B*N,)
+        
+        end_logits_flat = end_logits.view(B * L, C)
+        end_labels_flat = end_labels.view(-1)
+        end_loss = self.ce(end_logits_flat, end_labels_flat)  # (B*N,)
+
+        # ===== TOTAL =====
+        total_loss = (
+            self.lambda_trg_ce * trg_loss +
+            self.lambda_arg_ce * arg_loss +
+            self.lambda_start_ce * start_loss +
+            self.lambda_end_ce * end_loss
+        )
+
+        return {
+            "total": total_loss,
+            "trg_loss": trg_loss,
+            "arg_loss": arg_loss,
+            "start_loss": start_loss,
+            "end_loss": end_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]
+def get_span_reprs(hidden, spans):
+    """
+    Args:
+        hidden: (B, L, H)
+        spans:  (B, N, 2)
+
+    Return:
+        span_repr: (B, N, 4*H)
+    """
+
+    B, N, _ = spans.shape
+    H = hidden.size(-1)
+
+    batch_idx = torch.arange(B, device=hidden.device).unsqueeze(1)
+
+    start_idx = spans[..., 0]  # (B, N)
+    end_idx   = spans[..., 1]  # (B, N)
+    start_h = hidden[batch_idx, start_idx]
+    end_h   = hidden[batch_idx, end_idx]
+
+    span_repr = torch.cat(
+        [start_h, end_h, end_h - start_h, end_h * start_h],
+        dim=-1
+    )
+
+    return span_repr
+
+def extract_pred_pos_trgs(trg_logits, spans):
+    """
+    Args:
+        trg_logits: (B, N, C)
+        spans:      (B, N, 2)
+
+    Return:
+        gold_spans: (B, K, 2)
+            - chỉ giữ span có label > 0
+            - bỏ span (0, 0)
+            - padding bằng (0, 0)
+    """
+    B, N, _ = trg_logits.shape
+    device = spans.device
+
+    pred_labels = trg_logits.argmax(dim=-1)
+
+    # span hợp lệ
+    valid_span = ~(
+        (spans[:, :, 0] == 0)
+        &
+        (spans[:, :, 1] == 0)
+    )
+
+    keep = (pred_labels > 0) & valid_span
+
+    K = keep.sum(dim=1).max().item()
+
+    # rank trong từng batch
+    rank = keep.cumsum(dim=1) - 1
+
+    # output
+    gold_spans = torch.zeros(
+        (B, K, 2),
+        dtype=spans.dtype,
+        device=device
+    )
+
+    # index hợp lệ
+    b_idx, n_idx = torch.where(keep)
+    k_idx = rank[b_idx, n_idx]
+
+    gold_spans[b_idx, k_idx] = spans[b_idx, n_idx]
+
+    return gold_spans
+
+def filter_spans(
+    start_logits,   # (B, L, C)
+    end_logits,     # (B, L, C)
+    spans,          # (B, N, 2)
+    keep_neighbor=1
+):
+    """
+    Return:
+        filtered_spans: (B, N, 2)
+
+    Span bị loại sẽ được thay bằng (0, 0)
+    """
+
+    # (B, L)
+    start_pred = start_logits.argmax(dim=-1)
+    end_pred = end_logits.argmax(dim=-1)
+
+    s = spans[..., 0]  # (B, N)
+    e = spans[..., 1]  # (B, N)
+
+    # start/end hợp lệ trực tiếp
+    valid_s = start_pred.gather(1, s) > 0
+    valid_e = end_pred.gather(1, e) > 0
+
+    keep_mask = valid_s & valid_e
+
+    if keep_neighbor > 0:
+
+        B = spans.size(0)
+
+        for b in range(B):
+
+            good_s = torch.where(start_pred[b] > 0)[0]
+            good_e = torch.where(end_pred[b] > 0)[0]
+
+            if len(good_s) == 0 or len(good_e) == 0:
+                continue
+
+            cur_s = s[b]
+            cur_e = e[b]
+
+            # khoảng cách tới nearest positive start/end
+            dist_s = (
+                cur_s[:, None] - good_s[None, :]
+            ).abs().min(dim=1).values
+
+            dist_e = (
+                cur_e[:, None] - good_e[None, :]
+            ).abs().min(dim=1).values
+
+            near_s = dist_s <= keep_neighbor
+            near_e = dist_e <= keep_neighbor
+
+            keep_mask[b] = (
+                (valid_s[b] | near_s)
+                &
+                (valid_e[b] | near_e)
+            )
+
+    filtered_spans = spans.clone()
+
+    # thay span bị loại bằng (0, 0)
+    filtered_spans[~keep_mask] = 0
+
+    return filtered_spans
+
+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, num_trg_labels, num_arg_labels, keep_neighbor):
+        super().__init__()
+        self.keep_neighbor = keep_neighbor
+        
+        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.trg_classifier = MLP(4*hidden_size, hidden_size, num_trg_labels)
+        
+        self.span_repr_proj = MLP(4*hidden_size, hidden_size, hidden_size)
+        self.arg_classifier = MLP(2*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 # B * n_parts, L, H
+        
+        hidden_states = hidden_states.view(B, n_parts, L, -1).reshape(B, n_parts*L, -1) # B, L, H
+        return hidden_states
+
+    def get_token_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_trg_logits(self, span_reprs):
+        return self.trg_classifier(span_reprs)      # N, classes
+
+    def proj_span_repr(self, span_reprs):
+        return self.span_repr_proj(span_reprs)      # N, classes
+
+    def get_arg_logits(self, span_reprs, gold_reprs):
+        B, N, H = span_reprs.shape
+        K = gold_reprs.size(1)
+    
+        gold_expand = gold_reprs[:, :, None, :].expand(B, K, N, H)
+        span_expand = span_reprs[:, None, :, :].expand(B, K, N, H)
+    
+        x = torch.cat([gold_expand, span_expand], dim=-1)    
+        logits = self.arg_classifier(x)    
+        return logits
+
+    def forward(self, input_ids, attention_mask, spans, gold_spans=None):
+        hidden_states = self.encode(input_ids, attention_mask)
+        start_logits, end_logits = self.get_token_logits(hidden_states)
+        spans = filter_spans(start_logits, end_logits, spans, self.keep_neighbor)
+        
+        span_reprs = get_span_reprs(hidden_states, spans)
+        trg_logits = self.get_trg_logits(span_reprs)
+
+        if gold_spans is None:
+            gold_spans = extract_pred_pos_trgs(trg_logits, spans)
+        gold_reprs = get_span_reprs(hidden_states, gold_spans)
+
+        span_reprs = self.proj_span_repr(span_reprs)
+        gold_reprs = self.proj_span_repr(gold_reprs)
+
+        arg_logits = self.get_arg_logits(span_reprs, gold_reprs)
+        
+        return trg_logits, arg_logits, gold_spans
+
+def test_model():
+    model = nn.DataParallel(IEModel(backbone_model_name, 17, 7, 10, 0)).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)
+            
+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)
+
+        outputs = self.parallel_apply(
+            [getattr(replica, method_name) for replica in replicas],
+            inputs_scattered,
+            kwargs_scattered
+        )
+
+        return self.gather(outputs, self.output_device)
+
+def align_labels(gold_trgs, all_arg_labels, pred_gold_trgs):
+    """
+    Args:
+        gold_trgs:      (B, M, 2)
+        all_arg_labels: (B, M, N)
+        pred_gold_trgs: (B, K, 2)
+
+    Return:
+        pred_arg_labels: (B, K, N)
+            - match được -> copy label
+            - không match -> -100
+    """
+
+    B, M, _ = gold_trgs.shape
+    K = pred_gold_trgs.size(1)
+    N = all_arg_labels.size(-1)
+
+    device = gold_trgs.device
+
+    # (B, K, M, 2)
+    pred_expand = pred_gold_trgs[:, :, None, :]
+    gold_expand = gold_trgs[:, None, :, :]
+
+    # (B, K, M)
+    matched = (pred_expand == gold_expand).all(dim=-1)
+
+    # output
+    pred_arg_labels = torch.full(
+        (B, K, N),
+        fill_value=-100,
+        dtype=all_arg_labels.dtype,
+        device=device
+    )
+
+    # lấy vị trí match
+    b_idx, k_idx, m_idx = torch.where(matched)
+
+    # copy labels
+    pred_arg_labels[b_idx, k_idx] = all_arg_labels[b_idx, m_idx]
+
+    return pred_arg_labels
+
+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):
+
+        # ===== lấy trg positive =====
+        keep = trg_labels[bidx] > 0
+
+        trg_spans = all_spans[bidx, keep]       # (K', 2)
+        trg_lbs   = trg_labels[bidx, keep]      # (K',)
+
+        # map:
+        # span tuple -> trg label
+        trg_map = {}
+
+        for span, lb in zip(trg_spans, trg_lbs):
+            s, e = span.tolist()
+
+            # bỏ padding
+            if s == 0 or e == 0:
+                continue
+
+            trg_map[(s, e)] = lb.item()
+
+        # ===== duyệt pred_gold_trgs =====
+        for k in range(K):
+
+            trg_s, trg_e = pred_gold_trgs[bidx, k].tolist()
+
+            # padding
+            if trg_s == 0 or trg_e == 0:
+                continue
+
+            key = (trg_s, trg_e)
+
+            # không match trg prediction
+            if key not in trg_map:
+                continue
+
+            trg_lb = trg_map[key]
+
+            trg_token_ids = input_ids[bidx, trg_s:trg_e+1].tolist()
+
+            # arg của trigger này
+            cur_arg_labels = arg_labels[bidx, k]   # (N,)
+
+            for n in range(N):
+
+                arg_lb = cur_arg_labels[n].item()
+
+                # bỏ non-arg
+                if arg_lb <= 0:
+                    continue
+
+                arg_s, arg_e = all_spans[bidx, n].tolist()
+
+                # bỏ padding
+                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")
+                    
+                    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_trgs = batch['gold_trgs'].to(self.device) # B, M, 2
+        all_spans = batch['all_spans'].to(self.device) # B, N, 2
+        all_trg_labels = batch['all_trg_labels'].to(self.device) # B, N
+        all_arg_labels = batch['all_arg_labels'].to(self.device) # B, M, N
+        start_labels = batch['start_labels'].to(self.device)
+        end_labels = batch['end_labels'].to(self.device)
+        
+        hidden_states = network.encode(input_ids, attention_mask)
+        start_logits, end_logits = network.get_token_logits(hidden_states)
+        pred_spans = filter_spans(start_logits, end_logits, all_spans, network.keep_neighbor)
+        
+        span_reprs = get_span_reprs(hidden_states, pred_spans)
+        trg_logits = network.get_trg_logits(span_reprs)
+
+        choice = random.random()        
+        if choice < teaching_rate:
+            pred_gold_trgs = gold_trgs
+        else:
+            pred_gold_trgs = extract_pred_pos_trgs(trg_logits, pred_spans)
+            # pred_gold_trgs = gold_trgs
+        gold_reprs = get_span_reprs(hidden_states, pred_gold_trgs)
+
+        span_reprs = network.proj_span_repr(span_reprs)
+        gold_reprs = network.proj_span_repr(gold_reprs)
+        arg_logits = network.get_arg_logits(span_reprs, gold_reprs)
+
+        pred_gold_trg_arg_labels = align_labels(gold_trgs, all_arg_labels, pred_gold_trgs)
+
+        all_trg_labels[(pred_spans == 0).all(dim=-1)] = -100
+        
+        invalid_span_mask = (pred_spans == 0).all(dim=-1)  # (B, N)
+        pred_gold_trg_arg_labels[
+            invalid_span_mask.unsqueeze(1).expand_as(pred_gold_trg_arg_labels)
+        ] = -100
+        
+        loss_dict = loss_fn(
+            trg_logits, all_trg_labels, 
+            arg_logits, pred_gold_trg_arg_labels, 
+            start_logits, start_labels, 
+            end_logits, end_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_trgs = batch['gold_trgs'].to(self.device) # B, M, 2
+        all_spans = batch['all_spans'].to(self.device) # B, N, 2
+        gold_events = batch['gold_events']
+
+        B, _, _ = input_ids.shape
+        
+        hidden_states = network.encode(input_ids, attention_mask)
+        start_logits, end_logits = network.get_token_logits(hidden_states)
+        pred_spans = filter_spans(start_logits, end_logits, all_spans, network.keep_neighbor)
+        
+        span_reprs = get_span_reprs(hidden_states, pred_spans)
+        trg_logits = network.get_trg_logits(span_reprs)
+
+        pred_gold_trgs = extract_pred_pos_trgs(trg_logits, pred_spans)
+        # pred_gold_trgs = gold_trgs
+        gold_reprs = get_span_reprs(hidden_states, pred_gold_trgs)
+
+        span_reprs = network.proj_span_repr(span_reprs)
+        gold_reprs = network.proj_span_repr(gold_reprs)
+        arg_logits = network.get_arg_logits(span_reprs, gold_reprs)
+                
+        pred_ids = extract_events(input_ids.reshape(B, -1), pred_spans, trg_logits, arg_logits, pred_gold_trgs, 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):
+        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
+
+    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)
+        triggers_gold_spans = result["triggers_gold_spans"]
+        arguments_gold_spans = result["arguments_gold_spans"]
+
+        # Get all spans & labels
+        all_spans = torch.tensor(generate_spans(attention_mask, self.max_span_len))
+        gold_trgs = torch.tensor([list(trg_spans[0]) for trg_spans, _ in triggers_gold_spans], dtype=torch.long) if triggers_gold_spans else torch.empty(0, 2, dtype=torch.long)
+        gold_trg_labels = torch.tensor([self.label2id['Trg'][label] for _, label in triggers_gold_spans], dtype=torch.long) if triggers_gold_spans else torch.empty(0, dtype=torch.long)
+        all_trg_labels = match_gold_labels(
+            gold_trgs,      # (N, 2)
+            gold_trg_labels,     # (N,)
+            all_spans,      # (M, 2)
+            default_label=0
+        )
+        all_arg_labels = []
+        for idx, args in enumerate(arguments_gold_spans):
+            gold_args = torch.tensor([list(spans[0]) for spans, _ in args], dtype=torch.long) if args else torch.empty(0, 2, dtype=torch.long)
+            gold_arg_labels = torch.tensor([self.label2id['Arg'][label] for _, label in args], dtype=torch.long) if args else torch.empty(0, dtype=torch.long)
+            arg_labels = match_gold_labels(
+                gold_args,      # (N, 2)
+                gold_arg_labels,     # (N,)
+                all_spans,      # (M, 2)
+                default_label=0
+            )
+            all_arg_labels.append(arg_labels)
+        all_arg_labels = torch.stack(all_arg_labels, dim=0) if all_arg_labels else torch.empty(0, len(all_spans), dtype=torch.long)
+        
+        # Get event label
+        gold_events = []
+        start_labels = torch.ones_like(input_ids) * (1-attention_mask) * (-100)
+        end_labels = torch.ones_like(input_ids) * (1-attention_mask) * (-100)
+        all_arg_start_labels, all_arg_end_labels = [], []
+        for (trg_spans, trg_label), args in zip(triggers_gold_spans, arguments_gold_spans):
+            s, e = trg_spans[0]
+            
+            start_labels[s] = self.label2id['All'][f'{trg_label}']
+            end_labels[e] = self.label2id['All'][f'{trg_label}']
+            
+            event = [(tuple(input_ids[s:e+1].tolist()), trg_label)]
+            
+            for arg_spans, arg_label in args:
+                s, e = arg_spans[0]
+            
+                start_labels[s] = self.label2id['All'][f'{arg_label}']
+                end_labels[e] = self.label2id['All'][f'{arg_label}']
+            
+                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]
+        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,
+            
+            "gold_trgs": gold_trgs, 
+            "all_spans": all_spans, 
+            "all_trg_labels": all_trg_labels, 
+            "all_arg_labels": all_arg_labels, 
+            
+            "start_labels": start_labels, 
+            "end_labels": end_labels, 
+            
+            "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]
+    
+    gold_trgs = [b["gold_trgs"].unsqueeze(-1) for b in batch]
+    all_spans = [b["all_spans"].unsqueeze(-1) for b in batch]
+    all_trg_labels = [b["all_trg_labels"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    all_arg_labels = [b["all_arg_labels"].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)
+    
+    gold_trgs = _pad_batch(gold_trgs, pad_value=0).squeeze(-1)
+    all_spans = _pad_batch(all_spans, pad_value=0).squeeze(-1)
+    all_trg_labels = _pad_batch(all_trg_labels, pad_value=0).squeeze(-1).squeeze(-1)
+    all_arg_labels = _pad_batch(all_arg_labels, pad_value=0).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,
+        
+        "gold_trgs": gold_trgs, 
+        "all_spans": all_spans, 
+        "all_trg_labels": all_trg_labels, 
+        "all_arg_labels": all_arg_labels, 
+        
+        "start_labels": start_labels,
+        "end_labels": end_labels,
+        
+        "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
+        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):
+    if torch.cuda.device_count() > 1:
+        network = DataParallelProxy(network)
+    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)
+            gold_trgs = batch['gold_trgs'].to(device) # B, M, 2
+            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_span_reprs = []
+            list_hidden_states = []
+            list_arg_logits = []
+
+            list_hidden_states = []
+            list_logits = []
+            list_start_logits = []
+            list_end_logits = []
+            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 = network.encode(input_ids, attention_mask)
+                start_logits, end_logits = network.get_token_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 = filter_spans(start_logits, end_logits, all_spans, network.keep_neighbor)
+            
+            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)
+                span_reprs = get_span_reprs(hidden_states, spans)
+                trg_logits = network.get_trg_logits(span_reprs)
+                
+                list_span_reprs.append(span_reprs)
+                list_trg_logits.append(trg_logits)
+                list_hidden_states.append(hidden_states)
+
+            ensemble_trg_logits = torch.stack(list_trg_logits, dim=0).mean(dim=0)
+            pred_gold_trgs = extract_pred_pos_trgs(ensemble_trg_logits, spans)
+            # pred_gold_trgs = gold_trgs
+            
+            for sd, span_reprs, hidden_states in zip(state_dicts, list_span_reprs, list_hidden_states):
+                if torch.cuda.device_count() > 1:
+                    network.module.load_state_dict(sd)
+                else:
+                    network.load_state_dict(sd)
+                    
+                gold_reprs = get_span_reprs(hidden_states, pred_gold_trgs)
+            
+                span_reprs = network.proj_span_repr(span_reprs)
+                gold_reprs = network.proj_span_repr(gold_reprs)
+                arg_logits = network.get_arg_logits(span_reprs, gold_reprs)
+
+                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, pred_gold_trgs, 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_labels=len(label2id['All']), 
+                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_labels=len(label2id['All']), 
+    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
+
+# %% [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=["**/*.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))
+

6_4_actions_8/lasts/6_4_actions_8_s26092004_f0_last_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2974568d06c2be26e04d36f588d2775984784659592dbb92b02d00f02fe4264
+size 570950757

6_4_actions_8/logs/6_4_actions_8_s26092004_f0_logging.json

@@ -0,0 +1 @@
+{"1": {"lr": [2e-05, 0.0005], "train_loss": 0.9112924337387085, "total": 0.9112924020034184, "trg_loss": 0.07789018747305708, "arg_loss": 0.22461442679657248, "start_loss": 0.3163917740916929, "end_loss": 0.2923965971098288}, "2": {"lr": [1.988303923565381e-05, 0.0004969282409784868], "train_loss": 0.6298381090164185, "total": 0.629838084897598, "trg_loss": 0.08514611095565101, "arg_loss": 0.1286854776269815, "start_loss": 0.22154730679566492, "end_loss": 0.19445914520529556}, "3": {"lr": [1.9535036904803962e-05, 0.0004877886008156408], "train_loss": 0.5736811757087708, "total": 0.5736811915764157, "trg_loss": 0.08199325812636132, "arg_loss": 0.11042006456609617, "start_loss": 0.20467525160901198, "end_loss": 0.17659297917265399}, "4": {"lr": [1.8964561979789496e-05, 0.00047280612778499774], "train_loss": 0.5348196029663086, "total": 0.5348195817325145, "trg_loss": 0.07715586009584115, "arg_loss": 0.10276086148080152, "start_loss": 0.19157987227167533, "end_loss": 0.16332279585684142}, "5": {"lr": [1.8185661446562005e-05, 0.00045234974009654937], "train_loss": 0.503925621509552, "total": 0.5039256299915296, "trg_loss": 0.07317104875060504, "arg_loss": 0.09738894053604111, "start_loss": 0.18092403799612777, "end_loss": 0.1524411964970429}, "6": {"lr": [1.7217514421272206e-05, 0.00042692314190604356], "train_loss": 0.4682256579399109, "total": 0.46822564714991227, "trg_loss": 0.06936488193094673, "arg_loss": 0.08883334637149837, "start_loss": 0.16930921910923888, "end_loss": 0.14071823666656583}, "7": {"lr": [1.60839598967785e-05, 0.00039715242044697206], "train_loss": 0.43897730112075806, "total": 0.4389772812140761, "trg_loss": 0.06430294483698118, "arg_loss": 0.08520043992580772, "start_loss": 0.15908538729731364, "end_loss": 0.13038853869664357}, "8": {"lr": [1.4812909747525698e-05, 0.00036377062968501693], "train_loss": 0.4074500799179077, "total": 0.40745008234132984, "trg_loss": 0.059605309485465954, "arg_loss": 0.07867683491886873, "start_loss": 0.148988854850481, "end_loss": 0.12017938959171558, "val_score": 0.3616719150623694, "best_score": 0.3616719150623694, "new_best_model": true, "precision": 0.47052724521901934, "recall": 0.2971123137999204, "f1": 0.3616719150623694}, "9": {"lr": [1.3435661446562005e-05, 0.0003275997400965494], "train_loss": 0.3764086365699768, "total": 0.37640862404896236, "trg_loss": 0.05471176561058318, "arg_loss": 0.07237482209404117, "start_loss": 0.1389253658800596, "end_loss": 0.11039682187046225, "val_score": 0.3626506153498854, "best_score": 0.3626506153498854, "new_best_model": true, "precision": 0.462872735629236, "recall": 0.3015237400325962, "f1": 0.3626506153498854}, "10": {"lr": [1.1986127417882198e-05, 0.00028953039902753766], "train_loss": 0.35091128945350647, "total": 0.35091129092486995, "trg_loss": 0.05014919565354058, "arg_loss": 0.06931237722066343, "start_loss": 0.12990872142160273, "end_loss": 0.10154122189192204, "val_score": 0.36140226999963887, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.45625165643847315, "recall": 0.30261878694950484, "f1": 0.36140226999963887}, "11": {"lr": [1.0500000000000003e-05, 0.0002505], "train_loss": 0.3243034780025482, "total": 0.32430347837760165, "trg_loss": 0.04599585076153913, "arg_loss": 0.06362410597972001, "start_loss": 0.12151197819372807, "end_loss": 0.09317170962013326, "val_score": 0.36055086143634113, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.4509854722471751, "recall": 0.303726902110411, "f1": 0.36055086143634113}, "12": {"lr": [9.013872582117811e-06, 0.00021146960097246246], "train_loss": 0.303201824426651, "total": 0.3032018286676398, "trg_loss": 0.043012295895601405, "arg_loss": 0.059327541916321926, "start_loss": 0.11440922606949193, "end_loss": 0.08645287187840331, "val_score": 0.3594854857951942, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.44762960425345516, "recall": 0.30389110028768396, "f1": 0.3594854857951942}, "13": {"lr": [7.564338553438001e-06, 0.00017340025990345064], "train_loss": 0.28374871611595154, "total": 0.2837487190298282, "trg_loss": 0.03949165898163344, "arg_loss": 0.055837005290265, "start_loss": 0.10808725717106955, "end_loss": 0.08033296007154525, "val_score": 0.35703540846698795, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.44127810908738513, "recall": 0.3032473054883594, "f1": 0.35703540846698795}, "14": {"lr": [6.1870902524743065e-06, 0.00013722937031498307], "train_loss": 0.2665584087371826, "total": 0.2665584006591088, "trg_loss": 0.0376466970822241, "arg_loss": 0.05080971066972524, "start_loss": 0.1027717996543744, "end_loss": 0.07533008616060624, "val_score": 0.35392963444139386, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.436462740630476, "recall": 0.30104307491018295, "f1": 0.35392963444139386}, "15": {"lr": [4.916040103221507e-06, 0.00010384757955302797], "train_loss": 0.2507762312889099, "total": 0.25077623365463153, "trg_loss": 0.03463690463247973, "arg_loss": 0.047353122264348756, "start_loss": 0.09794941668588909, "end_loss": 0.0708366017373926, "val_score": 0.3537248420113581, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.437100476192545, "recall": 0.3004536336606692, "f1": 0.3537248420113581}, "16": {"lr": [3.7824855787278e-06, 7.40768580939564e-05], "train_loss": 0.23973706364631653, "total": 0.23973705369297554, "trg_loss": 0.03257899916760573, "arg_loss": 0.044744870092608, "start_loss": 0.0944964546432569, "end_loss": 0.06791683318884999, "val_score": 0.35221362027535, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.4364102291849418, "recall": 0.29852800399343604, "f1": 0.35221362027535}, "17": {"lr": [2.814338553438001e-06, 4.865025990345063e-05], "train_loss": 0.22852948307991028, "total": 0.22852948051223682, "trg_loss": 0.030845821192280715, "arg_loss": 0.04169298303092669, "start_loss": 0.09129442065406053, "end_loss": 0.06469627594555451, "val_score": 0.3495525262982319, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.43276043357566313, "recall": 0.29658257646725456, "f1": 0.3495525262982319}, "18": {"lr": [2.0354380202105066e-06, 2.8193872215002235e-05], "train_loss": 0.22239969670772552, "total": 0.2223996866822589, "trg_loss": 0.03005715745924057, "arg_loss": 0.04002173203120273, "start_loss": 0.08925520831363822, "end_loss": 0.06306549101808295, "val_score": 0.348364210968453, "best_score": 0.3626506153498854, "new_best_model": false, "precision": 0.4317700226039152, "recall": 0.29518386258792473, "f1": 0.348364210968453}}

6_4_actions_8/r1s/6_4_actions_8_s26092004_f0_r1_vs0.36265_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c2b24dff438715e69adaa740efe6ebd3db1d03f41d3a0feb8128708083b42e0
+size 570952557

6_4_actions_8/results/6_4_actions_8_test.json

@@ -0,0 +1,66 @@
+{
+  "Best model": {
+    "Trg-I": {
+      "precision": 0.32390146471317594,
+      "recall": 0.5582329317253063,
+      "f1": 0.409943117323735
+    },
+    "Trg-C": {
+      "precision": 0.2974367509981734,
+      "recall": 0.5124749067952037,
+      "f1": 0.376408631475956
+    },
+    "Arg-I": {
+      "precision": 0.23471573208711316,
+      "recall": 0.4143323595115876,
+      "f1": 0.29967061871254663
+    },
+    "Arg-C": {
+      "precision": 0.22593205490108734,
+      "recall": 0.39845493562197554,
+      "f1": 0.28835879771724426
+    },
+    "Soft-Event": {
+      "precision": 0.20685291540922474,
+      "recall": 0.36480686695247655,
+      "f1": 0.26400794667986344
+    },
+    "Strict-Event": {
+      "precision": 0.07872836218362396,
+      "recall": 0.13564668769677188,
+      "f1": 0.09963138029170508
+    }
+  },
+  "Last model": {
+    "Trg-I": {
+      "precision": 0.29956896551678025,
+      "recall": 0.5582329317253063,
+      "f1": 0.3899018187357665
+    },
+    "Trg-C": {
+      "precision": 0.2727832512311072,
+      "recall": 0.5081732147963631,
+      "f1": 0.3550035014136682
+    },
+    "Arg-I": {
+      "precision": 0.24573885448111493,
+      "recall": 0.41252792576008546,
+      "f1": 0.30800320301927603
+    },
+    "Arg-C": {
+      "precision": 0.23584809089966433,
+      "recall": 0.3955364806863557,
+      "f1": 0.29549826388448025
+    },
+    "Soft-Event": {
+      "precision": 0.21481216091707706,
+      "recall": 0.36025751072930445,
+      "f1": 0.269141973003389
+    },
+    "Strict-Event": {
+      "precision": 0.0728140394087549,
+      "recall": 0.13564668769677188,
+      "f1": 0.0947610893136188
+    }
+  }
+}