testing.py

import os, json, math, pickle
from datetime import datetime, timedelta
import numpy as np
import yfinance as yf
from sklearn.metrics import mean_absolute_error, mean_squared_error
import torch

from models import StockLSTM

os.environ["CUDA_VISIBLE_DEVICES"] = ""
ARTIFACTS_DIR = "artifacts"

@torch.no_grad()
def evaluate(symbol: str):
    base = os.path.join(ARTIFACTS_DIR, symbol.upper())
    model = StockLSTM(input_dim=1, hidden_dim=64, num_layers=2, dropout=0.2)
    model.load_state_dict(torch.load(os.path.join(base, "model.pt"), map_location="cpu"))
    model.eval()
    with open(os.path.join(base, "scaler.pkl"), "rb") as f:
        scaler = pickle.load(f)
    with open(os.path.join(base, "meta.json"), "r") as f:
        meta = json.load(f)

    seq_len = meta["seq_len"]

    end = datetime.utcnow().date()
    start = end - timedelta(days=5*365)
    df = yf.download(symbol, start=start.isoformat(), end=end.isoformat(), progress=False, auto_adjust=True)
    data = df[["Close"]].dropna().values

    scaled = scaler.transform(data)
    split_idx = int(len(scaled) * 0.8)
    test_scaled = scaled[split_idx - seq_len:]  # include tail of train for continuity

    # build sequences
    X, y = [], []
    for i in range(seq_len, len(test_scaled)):
        X.append(test_scaled[i-seq_len:i])
        y.append(test_scaled[i])
    X = np.array(X, dtype=np.float32)
    y = np.array(y, dtype=np.float32)

    X_t = torch.from_numpy(X)       # [N, T, 1]
    pred_scaled = model(X_t).numpy()
    pred = scaler.inverse_transform(pred_scaled)
    y_true = scaler.inverse_transform(y)

    rmse = math.sqrt(mean_squared_error(y_true, pred))
    mae = mean_absolute_error(y_true, pred)
    return {"symbol": symbol.upper(), "rmse": rmse, "mae": mae, "n": len(y_true)}