inference.py

import os, json, pickle
from datetime import datetime, timedelta
import numpy as np
import yfinance as yf
import torch

from models import StockLSTM

os.environ["CUDA_VISIBLE_DEVICES"] = ""
ARTIFACTS_DIR = "artifacts"
device = torch.device("cpu")

def _paths(symbol: str):
    base = os.path.join(ARTIFACTS_DIR, symbol.upper())
    return {
        "model": os.path.join(base, "model.pt"),
        "scaler": os.path.join(base, "scaler.pkl"),
        "meta": os.path.join(base, "meta.json"),
    }

def _load_artifacts(symbol: str):
    p = _paths(symbol)
    if not (os.path.exists(p["model"]) and os.path.exists(p["scaler"])):
        raise FileNotFoundError(f"Model/scaler not found for {symbol}. Train first.")
    with open(p["meta"], "r") as f:
        meta = json.load(f)
    with open(p["scaler"], "rb") as f:
        scaler = pickle.load(f)

    model = StockLSTM(input_dim=1, hidden_dim=64, num_layers=2, dropout=0.2)
    model = model.to(device)
    model.load_state_dict(torch.load(p["model"], map_location="cpu"))    
    model.eval()
    return model, scaler, meta

def _last_close_series(symbol: str, days: int = 400):
    end = datetime.utcnow().date()
    start = end - timedelta(days=days)
    df = yf.download(symbol, start=start.isoformat(), end=end.isoformat(), progress=False, auto_adjust=True)
    if df.empty:
        raise ValueError(f"No data for {symbol}")
    return df["Close"].values.reshape(-1, 1)

@torch.no_grad()
def predict_next(symbol: str, n_days: int = 1):
    model, scaler, meta = _load_artifacts(symbol)
    seq_len = meta["seq_len"]

    closes = _last_close_series(symbol, days=max(400, seq_len*5))
    scaled = scaler.transform(closes)

    # seed window
    window = scaled[-seq_len:].reshape(1, seq_len, 1).astype(np.float32)
    window_t = torch.from_numpy(window)

    preds_scaled = []
    for _ in range(n_days):
        yhat = model(window_t).numpy()          # [1,1] in scaled space
        preds_scaled.append(yhat[0, 0])
        # roll
        next_window = np.concatenate([window[:, 1:, :], yhat.reshape(1, 1, 1)], axis=1)
        window = next_window
        window_t = torch.from_numpy(window.astype(np.float32))

    preds_scaled = np.array(preds_scaled, dtype=np.float32).reshape(-1, 1)
    preds_unscaled = scaler.inverse_transform(preds_scaled).flatten().tolist()
    return {"symbol": symbol.upper(), "days": n_days, "predictions": preds_unscaled, "seq_len": seq_len, "meta": meta}