Model1.py

# model_train_eval.py
"""
Trains a Transformer model on saved data, evaluates predictability, and saves models to mar2025/models/.
Run after data_download.py in Google Colab (GPU recommended).
Includes time features and observed mask for TimeSeriesTransformerForPrediction with None checks.
"""

from google.colab import drive
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler
import torch
from transformers import TimeSeriesTransformerForPrediction, TimeSeriesTransformerConfig
import os

# Mount Google Drive
drive.mount('/content/drive')

# Configuration
BASE_DIR = '/content/drive/MyDrive/mar2025'
DATA_PATH = os.path.join(BASE_DIR, 'data')
MODEL_PATH = os.path.join(BASE_DIR, 'models')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load data
def load_data(symbol):
    file_path = os.path.join(DATA_PATH, f"{symbol}.csv")
    if os.path.exists(file_path):
        return pd.read_csv(file_path, index_col=0, parse_dates=True)
    raise FileNotFoundError(f"Data for {symbol} not found at {file_path}")

# Prepare sequences with time features
def create_sequences(data, context_length, prediction_length):
    scaler = MinMaxScaler(feature_range=(0, 1))
    value_scaled = scaler.fit_transform(data[['value']])
    time_features = data[['day', 'month', 'year', 'day_of_year']].values  # Ensure time features exist
    
    if len(data) < context_length + prediction_length:
        raise ValueError(f"Data length {len(data)} too short for context_length {context_length} and prediction_length {prediction_length}")
    
    X_values, y_values, X_time = [], [], []
    for i in range(len(data) - context_length - prediction_length + 1):
        X_values.append(value_scaled[i:i + context_length])
        y_values.append(value_scaled[i + context_length:i + context_length + prediction_length])
        X_time.append(time_features[i:i + context_length])
    
    X_values = np.array(X_values)
    y_values = np.array(y_values)
    X_time = np.array(X_time)
    
    if X_values.size == 0 or y_values.size == 0 or X_time.size == 0:
        raise ValueError(f"Empty sequence arrays for {symbol}")
    
    return (X_values, y_values, X_time, scaler)

# Train and predict
def train_and_predict(data, symbol, context_length=20, prediction_length=5):
    try:
        X_values, y_values, X_time, scaler...

Something went wrong, please try again.