Code/Train-Transformer.py

import os, sys, copy, joblib
import numpy as np
import pandas as pd
from tqdm import tqdm
import MetaTrader5 as mt5
import torch
import torch.nn as nn
import torch.optim as optim
from sklearn.model_selection import TimeSeriesSplit
from sklearn.preprocessing import MinMaxScaler, RobustScaler, PowerTransformer
from torch.utils.data import DataLoader
from pathlib import Path

# --- Config ---
LOOKBACK = 5
BATCH_SIZE = 4
EPOCHS = 50
LR = 1e-2
PATIENCE = 10
N_SPLITS = 3
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
datestr = pd.Timestamp.now().strftime("%d%m%Y")

# # --- Utility Class ---
# # sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '../../../common')))
# common_path = Path(__file__).resolve().parent.parent.parent.parent.parent / "common"

# # เพิ่มเข้า sys.path
# sys.path.append(str(common_path))


from AIBaseClass import *
from ClassUtils import *

# ================= Dynamic Beta SmoothL1Loss =================
class DynamicBetaSmoothL1Loss(nn.Module):
    def __init__(self, init_beta=0.5, alpha=0.9, reduction='mean', eps=1e-6):
        super().__init__()
        self.register_buffer('beta', torch.tensor(init_beta))
        self.alpha = alpha
        self.reduction = reduction
        self.eps = eps

    def forward(self, pred, target):
        diff = torch.abs(pred - target)
        batch_beta = diff.mean().detach() + self.eps
        self.beta = self.alpha * self.beta + (1 - self.alpha) * batch_beta
        beta = self.beta
        loss = torch.where(diff < beta, 0.5 * diff**2 / beta, diff - 0.5*beta)
        if self.reduction=='mean':
            return loss.mean()
        elif self.reduction=='sum':
            return loss.sum()
        else:
            return loss

# ================= Main Loop =================
symbollist = ['DXYm','EURUSDm','GBPUSDm','USDJPYm','USDCADm']
for symbol in tqdm(symbollist):
    SAVE_DIR = Path(__file__).resolve().parent.parent / f"Models/"
    
    os.makedirs(SAVE_DIR, exist_ok=True)

    CHECKPOINT_DIR = Path(__file__).resolve().parent.parent / f"Checkpoint/"
    os.makedirs(CHECKPOINT_DIR, exist_ok=True)
    
    try:
        
        data_dir = Path(__file__).resolve().parent.parent / "Datasource"
        # ====== Config ======
        if not data_dir.exists():
            raise ValueError(f"Folder {data_dir} ไม่พบ")

        csv_file = os.path.join(data_dir,f'{symbol}.csv')

        df = pd.read_csv(csv_file)

        # ตรวจสอบ columns ที่เราต้องใช้
        required_cols = ["time", "open", "high", "low", "close", "tick_volume"]
        for col in required_cols:
            if col not in df.columns:
                raise ValueError(f"{col} ไม่พบใน {csv_file}")

        # แปลง time เป็น datetime
        df['time'] = pd.to_datetime(df['time'])

        # ===== sort by time =====
        df = df.sort_values(by='time')

        # ===== set time เป็น index =====
        df.set_index('time', inplace=True)

        # ===== เพิ่ม feature =====
        df['hour']    = df.index.hour / 23.0
        df['weekday'] = df.index.weekday / 6.0
        df['month']   = (df.index.month - 1) / 11.0

        df                          = df.copy()
        N = 3
        # Shift targets to next day
        df['target_high']           = df['high'].shift(-1).ewm(span=N).mean()
        df['target_low']            = df['low'].shift(-1).ewm(span=N).mean()
        df.dropna(inplace=True)
        df.reset_index(drop=True, inplace=True)

        price_data                  = df[['open','high','low','close']].values
        target_high_data            = df[['target_high']].values
        target_low_data             = df[['target_low']].values
        tick_data                   = df[['tick_volume']].values
        time_data                   = df[['hour','weekday','month']].values

        # Initialize scalers
        scaler_price                = MinMaxScaler(feature_range=(-1,1))
        scaler_target_high          = MinMaxScaler(feature_range=(-1,1))
        scaler_target_low           = MinMaxScaler(feature_range=(-1,1))
        scaler_tick                 = RobustScaler()

        tscv = TimeSeriesSplit(n_splits=N_SPLITS)
        for fold, (train_idx, val_idx) in enumerate(tscv.split(price_data)):
            print(f"[{symbol}] Fold {fold+1}/{N_SPLITS}")

            # --- Split train/val ---
            train_price, val_price       = price_data[train_idx], price_data[val_idx]
            train_target_high, val_target_high = target_high_data[train_idx], target_high_data[val_idx]
            train_target_low, val_target_low   = target_low_data[train_idx], target_low_data[val_idx]
            train_tick, val_tick         = tick_data[train_idx], tick_data[val_idx]
            train_time, val_time         = time_data[train_idx], time_data[val_idx]

            # --- Scale train/val separately ---
            scaled_train_price       = scaler_price.fit_transform(train_price)
            scaled_val_price         = scaler_price.transform(val_price)
            scaled_train_target_high = scaler_target_high.fit_transform(train_target_high)
            scaled_val_target_high   = scaler_target_high.transform(val_target_high)
            scaled_train_target_low  = scaler_target_low.fit_transform(train_target_low)
            scaled_val_target_low    = scaler_target_low.transform(val_target_low)
            scaled_train_tick        = scaler_tick.fit_transform(train_tick)
            scaled_val_tick          = scaler_tick.transform(val_tick)

            # --- Combine features ---
            X_train_scaled = np.concatenate([
                scaled_train_price, scaled_train_tick, train_time
            ], axis=1)
            X_val_scaled = np.concatenate([
                scaled_val_price, scaled_val_tick, val_time
            ], axis=1)

            y_train_high, y_train_low = scaled_train_target_high, scaled_train_target_low
            y_val_high, y_val_low     = scaled_val_target_high, scaled_val_target_low

            # --- Dataset / DataLoader ---
            train_dataset = SinglestepDataset(X_train_scaled, y_train_high, y_train_low)
            val_dataset   = SinglestepDataset(X_val_scaled, y_val_high, y_val_low)
            train_loader  = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=False)
            val_loader    = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False)

            # --- Model / Optimizer / Scheduler / Loss ---
            model = ConvGRUTransformerHLV10(input_dim=8, seq_len=LOOKBACK, kernel_size=LOOKBACK).to(DEVICE)
            optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=0.001)
            scheduler = torch.optim.lr_scheduler.OneCycleLR(
                optimizer,
                max_lr=LR,
                steps_per_epoch=len(train_loader),
                epochs=EPOCHS,
                pct_start=0.3,
                anneal_strategy='cos'
            )
            loss_fn = DynamicBetaSmoothL1Loss(init_beta=0.5, alpha=0.1)
            early_stopping = EarlyStopping(patience=PATIENCE, verbose=True)

            best_val_loss = float('inf')
            best_model_weights = None

            # --- Training Loop ---
            for epoch in range(EPOCHS):
                model.train()
                train_loss = 0
                for xb, yb_high, yb_low in train_loader:
                    xb, yb_high, yb_low = xb.to(DEVICE), yb_high.to(DEVICE), yb_low.to(DEVICE)
                    optimizer.zero_grad()
                    pred_high, pred_low = model(xb)
                    loss = (loss_fn(pred_high, yb_high) + loss_fn(pred_low, yb_low)) / 2
                    loss.backward()
                    torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
                    optimizer.step()
                    train_loss += loss.item()

                train_loss /= len(train_loader)

                # --- Validation ---
                model.eval()
                val_loss = 0
                with torch.no_grad():
                    for xb, yb_high, yb_low in val_loader:
                        xb, yb_high, yb_low = xb.to(DEVICE), yb_high.to(DEVICE), yb_low.to(DEVICE)
                        pred_high, pred_low = model(xb)
                        loss = (loss_fn(pred_high, yb_high) + loss_fn(pred_low, yb_low)) / 2
                        val_loss += loss.item()

                val_loss /= len(val_loader)
                print(f"[{symbol}] Epoch {epoch+1} | TrainLoss: {train_loss:.6f} | ValLoss: {val_loss:.6f}")

                scheduler.step()
                model_path = os.path.join(f'{CHECKPOINT_DIR}',f'{symbol}_transformer_checkpoint_fold{fold+1}.pth')
                early_stopping(val_loss, model, model_path)

                if val_loss < best_val_loss:
                    best_val_loss = val_loss
                    best_model_weights = copy.deepcopy(model.state_dict())

                if early_stopping.early_stop:
                    print("Early stopping triggered")
                    break

            # --- Save best model ---
            if best_model_weights is not None:
                torch.save(best_model_weights, os.path.join(f'{SAVE_DIR}',f'{symbol}_best_fold{fold+1}.pth'))
                print(f"[{symbol}] ✅ Fold {fold+1} saved with val_loss: {best_val_loss:.6f}")

            model.load_state_dict(best_model_weights)
            torch.save({
                    'model_state_dict': best_model_weights,
                    'scalers': {
                        'scaler_price': scaler_price,
                        'scaler_tick': scaler_tick,
                        'scaler_target_high': scaler_target_high,
                        'scaler_target_low': scaler_target_low
                    }
                }, os.path.join(SAVE_DIR, f"{symbol}_fold{fold+1}.pth"))

    except Exception as e:
        print(f"[{symbol}] ❌ Error: {e}")