Upload 5_lr_up_lr_6's state dict

.gitattributes

@@ -68,3 +68,4 @@ saved_model/**/* 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
 8_entities_add_weight_keep_2_learn_all_10/logs/8_entities_add_weight_keep_2_learn_all_10_log_plot.jpg filter=lfs diff=lfs merge=lfs -text
+5_lr_up_lr_6/logs/5_lr_up_lr_6_log_plot.jpg filter=lfs diff=lfs merge=lfs -text

5_lr_up_lr_6/5_lr_up_lr_6.py

@@ -0,0 +1,1558 @@
+# %% [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/raw' # 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 = "5_lr_up_lr_6"
+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 = 2
+max_span_len = 14
+n_negs = 5 * 20
+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, 
+    'n_negs': n_negs,
+}
+val_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+    'n_negs': n_negs,
+}
+corpus_memory_params = {
+    'max_len': max_len, 
+    'max_n_parts': max_n_parts, 
+}
+
+# Data Loader
+def seed_worker(worker_id):
+    worker_seed = torch.initial_seed() % 2**32
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+    
+train_loader_params = {
+    'batch_size': batch_size, 
+    'shuffle': True,
+    'pin_memory':True, 
+    'num_workers': 2, 
+    'drop_last': False, 
+    'worker_init_fn': seed_worker, 
+    'persistent_workers': False, 
+}
+val_loader_params = {
+    'batch_size': batch_size, 
+    'shuffle': False,
+    'pin_memory':True, 
+    'num_workers': 1, 
+    'drop_last': False, 
+    'worker_init_fn': seed_worker, 
+    'persistent_workers': False, 
+}
+
+# Model
+model_params = {
+    'backbone_name': backbone_model_name, 
+    'projection_dim': 256, 
+    'normalize': True, 
+}
+
+# Loss Func
+loss_func_params = {
+    'lambda_contrastive': 1.0,
+    'lambda_triplet': 5.0,
+}
+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,
+        temperature=0.05,
+        margin=0.2,
+        lambda_contrastive=1.0,
+        lambda_triplet=0.5,
+    ):
+        super().__init__()
+
+        self.temperature = temperature
+        self.margin = margin
+
+        self.lambda_contrastive = lambda_contrastive
+        self.lambda_triplet = lambda_triplet
+
+    def forward(
+        self, 
+        encoded_text, encoded_pos, encoded_neg, pos_mask, 
+    ):
+        loss_contrastive = self.multi_pos_contrastive_loss(encoded_text, encoded_pos, encoded_neg, pos_mask)
+        loss_triplet = self.hardest_triplet_loss(encoded_text, encoded_pos, encoded_neg, pos_mask)
+
+        total_loss = (
+            self.lambda_contrastive * loss_contrastive + 
+            self.lambda_triplet * loss_triplet
+        )
+
+        return {
+            "total": total_loss,
+            "contrastive_loss": loss_contrastive,
+            "triplet_loss": loss_triplet,
+        }
+
+    def multi_pos_contrastive_loss(self, q, pos, neg, pos_mask):
+        B, P, D = pos.shape
+        N = neg.shape[1]
+
+        # ===== concat docs =====
+        docs = torch.cat([pos, neg], dim=1)  # [B, P+N, D]
+
+        # ===== similarity =====
+        logits = torch.matmul(q.unsqueeze(1), docs.transpose(1, 2)).squeeze(1)
+        logits = logits / self.temperature  # [B, P+N]
+
+        # ===== labels =====
+        labels = torch.zeros_like(logits)
+        labels[:, :P] = pos_mask  # chỉ pos hợp lệ
+
+        # ===== log-softmax =====
+        log_prob = logits - torch.logsumexp(logits, dim=1, keepdim=True)
+
+        # ===== normalize theo số pos thật =====
+        pos_count = pos_mask.sum(dim=1).clamp(min=1)
+
+        loss = -(labels * log_prob).sum(dim=1) / pos_count
+
+        return loss.mean()
+
+    def hardest_triplet_loss(self, q, pos, neg, pos_mask):
+        # ===== similarity =====
+        pos_sim = torch.matmul(q.unsqueeze(1), pos.transpose(1, 2)).squeeze(1)  # [B, P]
+        neg_sim = torch.matmul(q.unsqueeze(1), neg.transpose(1, 2)).squeeze(1)  # [B, N]
+
+        # ===== mask pos =====
+        pos_sim_masked = pos_sim.clone()
+        pos_sim_masked[pos_mask == 0] = float('inf')  # loại pad
+
+        # ===== hardest =====
+        hardest_pos = pos_sim_masked.min(dim=1).values
+        hardest_neg = neg_sim.max(dim=1).values
+
+        # ===== loss =====
+        loss = F.relu(self.margin + hardest_neg - hardest_pos)
+
+        return loss.mean()
+
+# %% [code]
+class CustomEvalFn(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, pred_topk, real_topk):
+        """
+        pred_topk: List[List[int]]  shape [B, K]
+        real_topk: List[List[int]]  shape [B, Ki]
+        """
+
+        B = len(pred_topk)
+
+        total_recall = 0.0
+        total_precision = 0.0
+        total_mrr = 0.0
+        total_f2 = 0.0
+
+        for i in range(B):
+            preds = pred_topk[i]
+            pred_set = set(preds)
+
+            gts = set(real_topk[i])
+
+            # ===== Recall@K =====
+            hit = any(p in gts for p in preds)
+            total_recall += 1.0 if hit else 0.0
+
+            # ===== MRR =====
+            rr = 0.0
+            for rank, p in enumerate(preds, start=1):
+                if p in gts:
+                    rr = 1.0 / rank
+                    break
+
+            total_mrr += rr
+
+            # ===== Precision / Recall / F2 =====
+            tp = len(pred_set & gts)
+
+            precision = tp / len(pred_set) if len(pred_set) > 0 else 0.0
+            recall_f = tp / len(gts) if len(gts) > 0 else 0.0
+
+            total_precision += precision
+
+            beta2 = 2 ** 2
+
+            if precision + recall_f > 0:
+                f2 = (1 + beta2) * precision * recall_f / (
+                    beta2 * precision + recall_f
+                )
+            else:
+                f2 = 0.0
+
+            total_f2 += f2
+
+        recall = total_recall / B
+        precision = total_precision / B
+        mrr = total_mrr / B
+        f2 = total_f2 / B
+
+        return {
+            "precision": precision,
+            "recall": recall,
+            "f2": f2, 
+            "mrr": mrr,
+        }
+
+# %% [code]
+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 EncodeModel(nn.Module):
+    def __init__(self, backbone_name, projection_dim, normalize):
+        super().__init__()
+
+        self.encoder = AutoModel.from_pretrained(backbone_name)
+        hidden_size = self.encoder.config.hidden_size
+
+        self.proj = MLP(hidden_size, hidden_size, projection_dim)
+        self.info_proj = MLP(hidden_size, hidden_size, projection_dim)
+
+        self.normalize = normalize
+
+    def embed_query(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 = outputs.last_hidden_state
+        hidden = hidden.view(B, n_parts, L, -1)
+
+        cls = hidden[:, :, 0].mean(dim=1)
+
+        embed = self.proj(cls)
+
+        return embed
+
+    def embed_info(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 = outputs.last_hidden_state
+        hidden = hidden.view(B, n_parts, L, -1)
+
+        cls = hidden[:, :, 0].mean(dim=1)
+
+        embed = self.info_proj(cls)
+
+        return embed
+
+    def embed_doc(self, input_ids, attention_mask):
+        B, K, 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 = outputs.last_hidden_state
+        hidden = hidden.view(B, K, n_parts, L, -1)
+
+        cls = hidden[:, :, :, 0].mean(dim=2)
+
+        embed = self.proj(cls)
+
+        return embed
+
+    def forward(
+        self,
+        text_input_ids,
+        text_attn_mask,
+        info_input_ids,
+        info_attn_mask,
+        doc_input_ids,
+        doc_attn_mask
+    ):
+        q_emb = self.embed_query(text_input_ids, text_attn_mask)
+        i_emb = self.embed_info(info_input_ids, info_attn_mask)
+        d_emb = self.embed_doc(doc_input_ids, doc_attn_mask)
+
+        if self.normalize:
+            q_emb = F.normalize(q_emb, dim=-1)
+            i_emb = F.normalize(i_emb, dim=-1)
+            d_emb = F.normalize(d_emb, dim=-1)
+
+        q_emb = q_emb + 0.1 * i_emb
+
+        if self.normalize:
+            q_emb = F.normalize(q_emb, dim=-1)
+
+        return q_emb, d_emb
+
+def test_model():
+    model = nn.DataParallel(EncodeModel('vinai/phobert-base', 256, True)).to(device)
+    model.eval()
+
+    bz = 32
+    vocab_size = 1000
+    qi = torch.randint(0, vocab_size, (bz, 1, 256)).to(device)
+    qa = torch.ones(bz, 1, 256).to(device)
+    di = torch.randint(0, vocab_size, (bz, 5, 2, 256)).to(device)
+    da = torch.ones(bz, 5, 2, 256).to(device)
+
+    st = time.time()
+
+    with torch.no_grad():
+        r = model(qi, qa, qi, qa, di, da)
+
+    if type(r) == tuple:
+        print([r[i].shape for i in range(len(r))])
+    else:
+        print(r.shape)
+    print(time.time() - st)
+
+    del model, qi, qa, di, da
+    torch.cuda.empty_cache()
+    gc.collect()
+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)
+    
+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, corpus_loader=None, eval_fn=None, start_epoch=1, start_training_time=None, refresh_every=3):
+        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:
+                eval_net = self.ema_net if (self.use_ema and self.ema_net is not None) else network
+                if (epoch - start_epoch) % refresh_every == 0:
+                    encoded_docs = self._get_encoded_docs(eval_net, corpus_loader)
+                    print(f"[Trainer CallBack] 📢 Epoch {epoch}/{epochs}: Refresh Encoded Doc (refresh_every={refresh_every})!")
+                elif (epoch - start_epoch - start_eval) % self.eval_every == 0 and epoch - start_epoch >= start_eval:
+                    encoded_docs = self._get_encoded_docs(eval_net, corpus_loader)
+                    print(f"[Trainer CallBack] 📢 Epoch {epoch}/{epochs}: Refresh Encoded Doc (eval_every={self.eval_every})!")
+                    
+                train_loss_epoch, train_loss_epoch_dict = self._train_epoch(network, train_loader, optimizer, scheduler, loss_fn, encoded_docs)
+                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:                    
+                    val_score, val_score_dict, _ = self._eval_epoch(eval_net, val_loader, eval_fn, encoded_docs)
+                    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, encoded_docs):
+        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, encoded_docs)
+
+                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, encoded_docs):
+        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, encoded_docs)
+                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 _get_encoded_docs(self, network, corpus_loader):
+        network.eval()
+        with torch.no_grad():
+            encoded_docs = []
+            for batch_idx, batch in enumerate(corpus_loader):
+                input_ids = batch['input_ids'].to(self.device)
+                attn_mask = batch['attn_mask'].to(self.device)
+                encoded_doc = network.embed_doc(input_ids, attn_mask)
+                if network.normalize:
+                    encoded_doc = F.normalize(encoded_doc, dim=-1)
+                encoded_docs.append(encoded_doc)
+            encoded_docs = torch.concat(encoded_docs, dim=0).squeeze(1)
+        return encoded_docs
+
+    def _cal_loss(self, network, batch, batch_idx, loss_fn, encoded_docs):
+        # Bạn cần override _cal_loss để tính loss
+        text_input_ids = batch['text_input_ids'].to(self.device)
+        text_attn_mask = batch['text_attn_mask'].to(self.device)
+        info_input_ids = batch['info_input_ids'].to(self.device)
+        info_attn_mask = batch['info_attn_mask'].to(self.device)
+        pos_idxes = batch['pos_idxes'].to(self.device)
+        pos_mask = batch['pos_mask'].to(self.device)
+        neg_idxes = batch['neg_idxes'].to(self.device)
+
+        encoded_text = network.embed_query(text_input_ids, text_attn_mask)
+        encoded_info = network.embed_info(info_input_ids, info_attn_mask)
+        encoded_text = encoded_text + encoded_info
+
+        if network.normalize:
+            encoded_text = F.normalize(encoded_text, dim=-1)
+            
+        encoded_pos = encoded_docs[pos_idxes]
+        encoded_neg = encoded_docs[neg_idxes]
+
+        loss_dict = loss_fn(
+            encoded_text, encoded_pos, encoded_neg, pos_mask, 
+        )
+        return loss_dict['total'], loss_dict
+
+    def _cal_val_score(self, network, batch, batch_idx, eval_fn, encoded_docs):
+        # Bạn cần override _cal_val_score để tính val score, list bên cạnh là để trả về y hay pred gì đó (nếu cần)
+        text_input_ids = batch['text_input_ids'].to(self.device)
+        text_attn_mask = batch['text_attn_mask'].to(self.device)
+        info_input_ids = batch['info_input_ids'].to(self.device)
+        info_attn_mask = batch['info_attn_mask'].to(self.device)
+        gt_pos_idxes = batch['gt_pos_idxes']
+
+        encoded_text = network.embed_query(text_input_ids, text_attn_mask)
+        encoded_info = network.embed_info(info_input_ids, info_attn_mask)
+        encoded_text = encoded_text + encoded_info
+
+        if network.normalize:
+            encoded_text = F.normalize(encoded_text, dim=-1)
+        scores = torch.matmul(encoded_text, encoded_docs.T)
+        
+        topk_scores, topk_indices = torch.topk(scores, k=10, dim=-1)  # [B, K]
+        pred_topk = [idx[logit > 0].tolist() for idx, logit in zip(topk_indices, topk_scores)]
+
+        pred_topk = list_to_tuple(pred_topk)
+        gt_pos_idxes = list_to_tuple(gt_pos_idxes)
+        score_dict = eval_fn(pred_topk, gt_pos_idxes)
+        return score_dict['f2'], score_dict, []
+
+# %% [code]
+def tokenize_to_parts(text, tokenizer, max_len, max_n_parts):
+    # Tokenize với overflow để chia thành nhiều đoạn
+    enc = tokenizer(
+        text,
+        max_length=max_len*max_n_parts,
+        truncation=True,
+        padding="max_length",
+        return_overflowing_tokens=True,
+        return_tensors="pt"
+    )
+
+    input_ids = enc["input_ids"].reshape(max_n_parts, max_len)          # (n_parts, max_len)
+    attn_mask = enc["attention_mask"].reshape(max_n_parts, max_len)     # (n_parts, max_len)
+
+    return input_ids, attn_mask
+
+class LawRetrievalDataset(Dataset):
+    def __init__(self, all_data, using_idxes, corpus_dict, tokenizer, max_len, max_n_parts, n_negs):
+        super().__init__()
+
+        self.all_data = all_data
+        self.using_idxes = using_idxes
+        self.tokenizer = tokenizer
+        self.max_len = max_len
+        self.max_n_parts = max_n_parts
+        self.n_negs = n_negs
+
+        # ===== BUILD CORPUS =====
+        idx = 0
+        self.corpus_list = []
+        self.corpus_dict = {}
+
+        for doc_name, articles_dict in corpus_dict.items():
+            self.corpus_dict[doc_name] = {}
+            for article_idx, content in articles_dict.items():
+                self.corpus_list.append([doc_name, article_idx, content])
+                self.corpus_dict[doc_name][article_idx] = {
+                    'content': content,
+                    'idx': idx
+                }
+                idx += 1
+
+    def __len__(self):
+        return len(self.using_idxes)
+
+    def __getitem__(self, idx):
+        ridx = self.using_idxes[idx]
+        data = self.all_data[ridx]
+        
+        query_text = data['text']
+
+        text_input_ids, text_attn_mask = tokenize_to_parts(
+            query_text, self.tokenizer, self.max_len, 1
+        )
+
+        info = ['[entities]']
+        for entity in data['entities']:
+            info.extend([f'[{entity["label"]}]', entity["name"]])
+        info += ['[actions]']
+        for action in data['actions']:
+            info.extend([f'[{action["label"]}]', action["trigger"]])
+            for arg in action['arguments']:
+                info.extend([f'[{arg["role"]}]', arg["argument"]])
+        info += ['[issues]']
+        for issue in data['issues']:
+            info.extend([f'[{issue["label"]}]', issue["trigger"]])
+            for arg in issue['arguments']:
+                info.extend([f'[{arg["role"]}]', arg["argument"]])
+        info_input_ids, info_attn_mask = tokenize_to_parts(
+            ' '.join(info), self.tokenizer, self.max_len, 1
+        )
+
+        # ===== POS =====
+        gt_pos_idxes = []
+        hard_names = []
+        for law in data['relevant_law']:
+            name = law['doc']
+            art = law['art']
+            gt_pos_idxes.append(self.corpus_dict[name][art]['idx'])
+            if name not in hard_names:
+                hard_names.append(name)
+
+        pos_idxes = torch.tensor(gt_pos_idxes, dtype=torch.long)
+        pos_mask = torch.ones(len(pos_idxes))
+
+        # ===== NEG =====
+        hard_neg_idxes = []
+        for name in hard_names:
+            for content in self.corpus_dict[name].values():
+                if content['idx'] in gt_pos_idxes:
+                    continue
+                hard_neg_idxes.append(content['idx'])
+
+        easy_neg_idxes = list(range(len(self.corpus_list)))
+        for i in gt_pos_idxes + hard_neg_idxes:
+            if i in easy_neg_idxes:
+                easy_neg_idxes.remove(i)
+
+        n_hards = min(len(hard_neg_idxes), self.n_negs // 2)
+        neg_idxes = random.sample(hard_neg_idxes, n_hards) + random.sample(easy_neg_idxes, self.n_negs - n_hards)
+        neg_idxes = torch.tensor(neg_idxes, dtype=torch.long)
+
+        return {
+            'text_input_ids': text_input_ids,
+            'text_attn_mask': text_attn_mask,
+            'info_input_ids': text_input_ids,
+            'info_attn_mask': text_attn_mask,
+            'gt_pos_idxes': gt_pos_idxes,
+            'pos_idxes': pos_idxes,
+            'pos_mask': pos_mask,
+            'neg_idxes': neg_idxes,
+        }
+
+class CorpusDataset(Dataset):
+    def __init__(self, corpus_dict, tokenizer, max_len, max_n_parts):
+        super().__init__()
+        self.tokenizer = tokenizer
+        self.max_len = max_len
+        self.max_n_parts = max_n_parts
+
+        idx = 0
+        self.corpus_list = []
+        self.corpus_dict = {}
+        for doc_name, articles_dict in corpus_dict.items():
+            self.corpus_dict[doc_name] = {}
+            for article_idx, content in articles_dict.items():
+                self.corpus_list.append([doc_name, article_idx, content])
+                self.corpus_dict[doc_name][article_idx] = {'content': content, 'idx': idx}
+                idx += 1
+
+    def __len__(self):
+        return len(self.corpus_list)
+        
+    def _encode_contexts(self, idxes):
+        all_input_ids, all_attn_mask = [], []
+        for idx in idxes:
+            name = self.corpus_list[idx][0]
+            art  = self.corpus_list[idx][1]
+            corpus = self.corpus_dict[name][art]
+            if 'content_input_ids' in corpus and 'content_attn_mask' in corpus:
+                content_input_ids = corpus['content_input_ids']
+                content_attn_mask = corpus['content_attn_mask']
+            else:
+                content = corpus['content']
+                content_input_ids, content_attn_mask = tokenize_to_parts(content, self.tokenizer, self.max_len, self.max_n_parts)
+                corpus['content_input_ids'] = content_input_ids
+                corpus['content_attn_mask'] = content_attn_mask
+                
+            all_input_ids.append(content_input_ids)
+            all_attn_mask.append(content_attn_mask)
+            
+        all_input_ids = torch.stack(all_input_ids)
+        all_attn_mask = torch.stack(all_attn_mask)
+        return all_input_ids, all_attn_mask
+
+    def __getitem__(self, idx):
+        input_ids, attn_mask = self._encode_contexts([idx])
+        
+        return {
+            'input_ids': input_ids, 
+            'attn_mask': attn_mask, 
+        }    
+
+def _pad_batch(tensor_list, pad_value=0):
+    """
+    tensor_list: list of tensors, mỗi tensor shape (Nk, max_n_parts, max_len)
+    return: tensor shape (B, max_Nk, max_n_parts, max_len)
+    """
+    max_Nk = max(t.size(0) for t in tensor_list)
+
+    padded = []
+    for t in tensor_list:
+        Nk = t.size(0)
+
+        if Nk < max_Nk:
+            pad_shape = (max_Nk - Nk, *t.shape[1:])
+            pad_tensor = t.new_full(pad_shape, 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):
+    text_input_ids = torch.stack([b["text_input_ids"] for b in batch])
+    text_attn_mask = torch.stack([b["text_attn_mask"] for b in batch])
+    info_input_ids = torch.stack([b["info_input_ids"] for b in batch])
+    info_attn_mask = torch.stack([b["info_attn_mask"] for b in batch])
+    gt_pos_idxes = [b["gt_pos_idxes"] for b in batch]
+    neg_idxes = torch.stack([b["neg_idxes"] for b in batch])
+
+    pos_idxes = [b["pos_idxes"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+    pos_mask = [b["pos_mask"].unsqueeze(-1).unsqueeze(-1) for b in batch]
+
+    # pad theo Nk
+    pos_idxes = _pad_batch(pos_idxes, pad_value=0).squeeze(-1).squeeze(-1)
+    pos_mask = _pad_batch(pos_mask, pad_value=0).squeeze(-1).squeeze(-1)
+
+    return {
+        'text_input_ids': text_input_ids,
+        'text_attn_mask': text_attn_mask,
+        'info_input_ids': info_input_ids,
+        'info_attn_mask': info_attn_mask,
+        'gt_pos_idxes': gt_pos_idxes,
+        'pos_idxes': pos_idxes,
+        'pos_mask': pos_mask,
+        'neg_idxes': neg_idxes,
+    }
+
+# %% [code]
+def encode_corpus(state_dicts, network, corpus_loader, device):
+    if torch.cuda.device_count() > 1:
+        network = DataParallelProxy(network)
+    network.to(device)
+    network.eval()
+
+    all_model_embs = []
+    for i, state_dict in enumerate(state_dicts):
+        # ===== load model =====
+        if torch.cuda.device_count() > 1:
+            network.module.load_state_dict(state_dict)
+        else:
+            network.load_state_dict(state_dict)
+
+        encoded_docs = []
+
+        with torch.no_grad():
+            for batch in corpus_loader:
+                input_ids = batch['input_ids'].to(device)
+                attn_mask = batch['attn_mask'].to(device)
+
+                encoded_doc = network.embed_doc(input_ids, attn_mask)
+                if network.normalize:
+                    encoded_doc = F.normalize(encoded_doc, dim=-1)
+
+                encoded_docs.append(encoded_doc)
+
+        encoded_docs = torch.concat(encoded_docs, dim=0).squeeze(1)  # [N, D]
+        all_model_embs.append(encoded_docs)
+
+    # ===== ensemble =====
+    # stack → [M, N, D]
+    all_model_embs = torch.stack(all_model_embs, dim=0)
+    final_embs = all_model_embs.mean(dim=0)  # [N, D]
+
+    return final_embs
+
+def test(state_dicts, network, test_loader, device, eval_fn, encoded_docs, topks=[5, 10, 15]):
+    if torch.cuda.device_count() > 1:
+        network = DataParallelProxy(network)
+    network.to(device)
+    network.eval()
+
+    per_model_scores = []
+    max_k = max(topks)
+    
+    all_scores = []
+    all_logits = []
+    all_gt_pos_idxes = []
+    with torch.no_grad():
+        for batch in test_loader:
+            text_input_ids = batch['text_input_ids'].to(device)
+            text_attn_mask = batch['text_attn_mask'].to(device)
+            info_input_ids = batch['info_input_ids'].to(device)
+            info_attn_mask = batch['info_attn_mask'].to(device)
+            gt_pos_idxes = batch['gt_pos_idxes']
+            all_gt_pos_idxes.extend(gt_pos_idxes)
+
+            list_encoded_texts = []
+            list_logits = []
+
+            for state_dict in state_dicts:
+                # ===== load model =====
+                if torch.cuda.device_count() > 1:
+                    network.module.load_state_dict(state_dict)
+                else:
+                    network.load_state_dict(state_dict)
+
+                encoded_text = network.embed_query(text_input_ids, text_attn_mask)
+                encoded_info = network.embed_info(info_input_ids, info_attn_mask)
+                
+                if network.normalize:
+                    encoded_text = F.normalize(encoded_text, dim=-1)
+                    encoded_info = F.normalize(encoded_info, dim=-1)
+                encoded_text = encoded_text + 0.1 * encoded_info
+        
+                if network.normalize:
+                    encoded_text = F.normalize(encoded_text, dim=-1)
+                
+                list_encoded_texts.append(encoded_text)
+                
+            ensemble_encoded_text = torch.stack(list_encoded_texts, dim=0).mean(dim=0)
+            scores = torch.matmul(ensemble_encoded_text, encoded_docs.T) # B, M
+            all_scores.append(scores)
+
+    all_scores = torch.concat(all_scores, dim=0) # N, M
+                
+    topk_scores, topk_indices = torch.topk(all_scores, k=10, dim=-1)  # [B, K]
+    pred_topk_full = [idx[logit > 0].tolist() for idx, logit in zip(topk_indices, topk_scores)]
+
+    pred_topk_full = list_to_tuple(pred_topk_full)
+    all_gt_pos_idxes = list_to_tuple(all_gt_pos_idxes)
+
+    final_score = {}
+    for k in topks:
+        pred_topk_k = [p[:k] for p in pred_topk_full]
+        final_score[k] = eval_fn(pred_topk_k, all_gt_pos_idxes)
+
+    return final_score
+
+# %% [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)
+    
+with open(f'{test_dir}/corpus.json', "r", encoding="utf-8") as f:
+    data_corpus = json.load(f)
+
+print('Train:', len(data_train))
+print('Test:', len(data_test))
+print('Corpus:', len(data_corpus))
+
+# %% [code]
+entity_types = sorted(list(set([e['label'] for d in data_train + data_test for e in d['entities']])))
+action_trigger_types = sorted(list(set([e['label'] for d in data_train + data_test for e in d['actions']])))
+issue_trigger_types = sorted(list(set([e['label'] for d in data_train + data_test for e in d['issues']])))
+argument_types = sorted(list(set([a['role'] for d in data_train + data_test for e in d['issues'] for a in e['arguments']])))
+
+# %% [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)
+special_tokens = ['[entities]', '[actions]', '[issues]'] + [f'[{t}]' for t in entity_types + action_trigger_types + issue_trigger_types + argument_types]
+tokenizer.add_special_tokens({'additional_special_tokens': sorted(list(set(special_tokens)))})
+
+# %% [code]
+print('Experiment name:', state_dict_save_name)
+
+# %% [code]
+if not test_only:
+    full_idxes = np.array(range(len(data_train)))
+    training_logs, best_models, last_models = [], [], []
+    start_training_time = time.time()
+    for seed in SEEDS:
+        kf = KFold(n_splits=nfolds, shuffle=True, random_state=seed)
+        generator = torch.Generator()
+        generator.manual_seed(seed)
+    
+        corpusset = CorpusDataset(data_corpus, tokenizer, **corpus_memory_params)
+        corpus_loader = DataLoader(corpusset, generator=generator, **val_loader_params)
+        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 = LawRetrievalDataset(data_train, train_idxes, data_corpus, tokenizer, **train_memory_params)
+            valset = LawRetrievalDataset(data_train, val_idxes, data_corpus, tokenizer, **val_memory_params)
+            
+            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 = EncodeModel(
+                **model_params
+            )
+            my_model.encoder.resize_token_embeddings(len(tokenizer))
+            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": 5e-4},
+                {"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, corpus_loader, eval_fn, 
+                start_epoch=1, start_training_time=start_training_time, refresh_every=2, 
+            )
+            
+            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 = LawRetrievalDataset(data_test, range(len(data_test)), data_corpus, 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)
+my_model = EncodeModel(
+    **model_params
+)
+my_model.encoder.resize_token_embeddings(len(tokenizer))
+total_params = sum(p.numel() for p in my_model.parameters())
+print(f"Total params: {total_params:,}")
+
+# %% [code]
+start_time = time.time()
+encoded_docs = encode_corpus(best_models, my_model, corpus_loader, device)
+best_score = test(best_models, my_model, test_loader, device, eval_fn, encoded_docs)
+
+encoded_docs = encode_corpus(last_models, my_model, corpus_loader, device)
+last_score = test(last_models, my_model, test_loader, device, eval_fn, encoded_docs)
+
+result_test = {"Best model": best_score, "Last model": last_score}
+
+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)
+    
+print('Test:', time.time() - start_time, 's --> Done!')
+
+# %% [code]
+def dict_to_df(data):
+    row_tuples = []
+    row_values = []
+
+    # ===== lấy model đầu tiên =====
+    first_model = next(iter(data.values()))
+
+    # ===== eval keys =====
+    eval_keys = list(first_model.keys())
+
+    # ===== tự lấy metrics =====
+    first_eval = next(iter(first_model.values()))
+    metrics = list(first_eval.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 = []
+
+    for model_name in data.keys():
+        for metric in ["f1", "f2", "mrr", "recall", "precision"]:
+            key = (model_name, metric)
+
+            if key in df.columns:
+                sort_keys.append(key)
+
+    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", "f2")
+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=["**/*lr*.json"], ignore_patterns=[])
+    get_ipython().system('rm -rf .cache .gitattributes')
+
+# %% [code]
+if not test_only:
+    experiments = {}
+    for experiment in os.listdir(pretrained_dir):
+        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))
+

5_lr_up_lr_6/lasts/5_lr_up_lr_6_s26092004_f0_last_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba8d56e59d51e1ad70eef1aeaf490feab6a1e3444be712ad5d101a64051c922e
+size 546476760

5_lr_up_lr_6/logs/5_lr_up_lr_6_s26092004_f0_logging.json

@@ -0,0 +1 @@
+{"1": {"lr": [0.0005, 0.0005], "train_loss": 5.792922019958496, "total": 5.792922208141722, "contrastive_loss": 4.634641513377926, "triplet_loss": 0.24289297658862877}, "2": {"lr": [0.0004969282409784868, 0.0004969282409784868], "train_loss": 5.726807594299316, "total": 5.726807457148829, "contrastive_loss": 4.611471424932065, "triplet_loss": 0.2334866220735786}, "3": {"lr": [0.0004877886008156408, 0.0004877886008156408], "train_loss": 5.654202938079834, "total": 5.654202974759615, "contrastive_loss": 4.65382329117893, "triplet_loss": 0.19147157190635453}, "4": {"lr": [0.00047280612778499774, 0.00047280612778499774], "train_loss": 5.654203414916992, "total": 5.65420338302153, "contrastive_loss": 4.6538686082514635, "triplet_loss": 0.19147157190635453}, "5": {"lr": [0.00045234974009654937, 0.00045234974009654937], "train_loss": 5.654292106628418, "total": 5.654291975857023, "contrastive_loss": 4.65389514527592, "triplet_loss": 0.19147157190635453}, "6": {"lr": [0.00042692314190604356, 0.00042692314190604356], "train_loss": 5.654357433319092, "total": 5.6543577060252925, "contrastive_loss": 4.653764093201296, "triplet_loss": 0.19147157190635453}, "7": {"lr": [0.00039715242044697206, 0.00039715242044697206], "train_loss": 5.654600620269775, "total": 5.654600621864549, "contrastive_loss": 4.653791855011496, "triplet_loss": 0.19147157190635453}, "8": {"lr": [0.00036377062968501693, 0.00036377062968501693], "train_loss": 5.654788970947266, "total": 5.654789238869147, "contrastive_loss": 4.653832272941053, "triplet_loss": 0.19147157190635453, "val_score": 0.0016072584822584824, "best_score": 0.0016072584822584824, "new_best_model": true, "precision": 0.0006250000000000002, "recall": 0.00625, "f2": 0.0016072584822584824, "mrr": 0.0025850694444444437}, "9": {"lr": [0.0003275997400965494, 0.0003275997400965494], "train_loss": 5.654257774353027, "total": 5.654257681856187, "contrastive_loss": 4.653806552440426, "triplet_loss": 0.19147157190635453, "val_score": 0.0003422040922040922, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.00016666666666666666, "recall": 0.0016666666666666668, "f2": 0.0003422040922040922, "mrr": 0.00033556547619047623}, "10": {"lr": [0.00028953039902753766, 0.00028953039902753766], "train_loss": 5.653966426849365, "total": 5.653966591110995, "contrastive_loss": 4.653811451583403, "triplet_loss": 0.1913670568561873, "val_score": 0.0006956006956006955, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.0003333333333333334, "recall": 0.003125, "f2": 0.0006956006956006955, "mrr": 0.0005031415343915344}, "11": {"lr": [0.0002505, 0.0002505], "train_loss": 5.653940200805664, "total": 5.653940054086538, "contrastive_loss": 4.653813492892977, "triplet_loss": 0.19147157190635453, "val_score": 0.0009017487142487143, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.00039583333333333343, "recall": 0.00375, "f2": 0.0009017487142487143, "mrr": 0.001205109126984127}, "12": {"lr": [0.00021146960097246246, 0.00021146960097246246], "train_loss": 5.654030799865723, "total": 5.654030688231606, "contrastive_loss": 4.653771441915761, "triplet_loss": 0.19147157190635453, "val_score": 0.0013901492026492028, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.0006458333333333337, "recall": 0.00625, "f2": 0.0013901492026492028, "mrr": 0.0015205026455026457}, "13": {"lr": [0.00017340025990345064, 0.00017340025990345064], "train_loss": 5.654173374176025, "total": 5.654173579901756, "contrastive_loss": 4.653814717678721, "triplet_loss": 0.1913670568561873, "val_score": 0.0008313330188330187, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.0003958333333333335, "recall": 0.003958333333333334, "f2": 0.0008313330188330187, "mrr": 0.000707010582010582}, "14": {"lr": [0.00013722937031498307, 0.00013722937031498307], "train_loss": 5.654489517211914, "total": 5.654489574623746, "contrastive_loss": 4.653799203725962, "triplet_loss": 0.1913670568561873, "val_score": 0.0015971875346875346, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.0007291666666666671, "recall": 0.006875, "f2": 0.0015971875346875346, "mrr": 0.002682787698412698}, "15": {"lr": [0.00010384757955302797, 0.00010384757955302797], "train_loss": 5.654581546783447, "total": 5.654581433554557, "contrastive_loss": 4.653864117370401, "triplet_loss": 0.19147157190635453, "val_score": 0.000866355866355866, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.00041666666666666675, "recall": 0.004166666666666667, "f2": 0.000866355866355866, "mrr": 0.0021487268518518518}, "16": {"lr": [7.40768580939564e-05, 7.40768580939564e-05], "train_loss": 5.654132843017578, "total": 5.654132753710284, "contrastive_loss": 4.653907801395276, "triplet_loss": 0.1913670568561873, "val_score": 0.000542119917119917, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.00025000000000000006, "recall": 0.0025, "f2": 0.000542119917119917, "mrr": 0.0012953869047619049}, "17": {"lr": [4.865025990345063e-05, 4.865025990345063e-05], "train_loss": 5.654181480407715, "total": 5.6541813368781355, "contrastive_loss": 4.653891062656773, "triplet_loss": 0.19147157190635453, "val_score": 0.0009113802863802865, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.0004166666666666667, "recall": 0.004166666666666667, "f2": 0.0009113802863802865, "mrr": 0.0019849537037037036}, "18": {"lr": [2.8193872215002235e-05, 2.8193872215002235e-05], "train_loss": 5.65456485748291, "total": 5.654565103077968, "contrastive_loss": 4.653933930157818, "triplet_loss": 0.19147157190635453, "val_score": 0.0007158813408813409, "best_score": 0.0016072584822584824, "new_best_model": false, "precision": 0.00029166666666666675, "recall": 0.002916666666666667, "f2": 0.0007158813408813409, "mrr": 0.0012632275132275135}}

5_lr_up_lr_6/r1s/5_lr_up_lr_6_s26092004_f0_r1_vs0.00161_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b75a09ed812dd5994be37e3ece925da21c66a9d2aa3f3579ec254bdf8b1ce5e4
+size 546478464

5_lr_up_lr_6/results/5_lr_up_lr_6_test.json

@@ -0,0 +1,42 @@
+{
+  "Best model": {
+    "5": {
+      "precision": 0.00030120481927710846,
+      "recall": 0.0015060240963855422,
+      "f2": 0.00046890860143872194,
+      "mrr": 0.00048318273092369475
+    },
+    "10": {
+      "precision": 0.0002635542168674699,
+      "recall": 0.0026355421686746986,
+      "f2": 0.0006595855616939954,
+      "mrr": 0.0006557479919678715
+    },
+    "15": {
+      "precision": 0.0002635542168674699,
+      "recall": 0.0026355421686746986,
+      "f2": 0.0006595855616939954,
+      "mrr": 0.0006557479919678715
+    }
+  },
+  "Last model": {
+    "5": {
+      "precision": 0.0004518072289156626,
+      "recall": 0.002259036144578313,
+      "f2": 0.0008085455374612,
+      "mrr": 0.0006337851405622489
+    },
+    "10": {
+      "precision": 0.000602409638554217,
+      "recall": 0.006024096385542169,
+      "f2": 0.0014563331581403868,
+      "mrr": 0.0011450564161407533
+    },
+    "15": {
+      "precision": 0.000602409638554217,
+      "recall": 0.006024096385542169,
+      "f2": 0.0014563331581403868,
+      "mrr": 0.0011450564161407533
+    }
+  }
+}