التسلسلات: RNN و LSTM

import numpy as np
import torch
from torch import nn
import matplotlib.pyplot as plt

torch.manual_seed(0)
rng = np.random.default_rng(0)

# ---- 1 year of daily sales: trend + weekly pattern + noise ----
days = np.arange(365)
sales = 200 + 0.3 * days + 40 * np.sin(2 * np.pi * days / 7) + rng.normal(0, 8, 365)

mean, std = sales[:300].mean(), sales[:300].std()      # scale using train part only
s = (sales - mean) / std

WINDOW = 14
X = np.array([s[i:i + WINDOW] for i in range(len(s) - WINDOW)])
y = s[WINDOW:]
X = torch.tensor(X, dtype=torch.float32).unsqueeze(-1)  # (samples, 14 steps, 1 feature)
y = torch.tensor(y, dtype=torch.float32).unsqueeze(-1)
split = 300 - WINDOW
X_tr, y_tr, X_te, y_te = X[:split], y[:split], X[split:], y[split:]

class SalesLSTM(nn.Module):
    def __init__(self):
        super().__init__()
        self.lstm = nn.LSTM(input_size=1, hidden_size=32, batch_first=True)
        self.head = nn.Linear(32, 1)
    def forward(self, x):
        out, _ = self.lstm(x)           # out: output at every step
        return self.head(out[:, -1])    # use the last step to predict tomorrow

model = SalesLSTM()
opt = torch.optim.Adam(model.parameters(), lr=0.01)
for epoch in range(150):
    opt.zero_grad()
    loss = nn.functional.mse_loss(model(X_tr), y_tr)
    loss.backward()
    opt.step()

model.eval()
with torch.no_grad():
    pred = model(X_te).squeeze().numpy() * std + mean   # back to real units
actual = y_te.squeeze().numpy() * std + mean
print(f"average error on unseen days: {np.abs(pred - actual).mean():.1f} units")

plt.figure(figsize=(7, 3.2))
plt.plot(actual, label="actual", color="#14213D")
plt.plot(pred, label="LSTM forecast", color="#E0A800", linewidth=2)
plt.title("Next-day sales forecast (65 unseen days)")
plt.legend(); plt.grid(alpha=0.3)
plt.show()