#!/usr/bin/env python3
# train_seq.py
import os, json, argparse
import numpy as np
import torch, torch.nn as nn
from torch.utils.data import Dataset, DataLoader

def get_device():
    return torch.device("mps") if torch.backends.mps.is_available() else torch.device("cpu")

class SeqDataset(Dataset):
    def __init__(self, X, y, augment=False):
        self.X = X.astype(np.float32)   # (Nclip, T, 63)
        self.y = y.astype(np.int64)
        self.augment = augment

    def __len__(self): return len(self.y)

    def _augment(self, seq):  # seq: (T,63)
        T = seq.shape[0]
        pts = seq.reshape(T, 21, 3).copy()
        # small 2D rotation (±7°) + scale (±10%) + Gaussian noise (σ=0.01)
        ang = np.deg2rad(np.random.uniform(-7, 7))
        c, s = np.cos(ang), np.sin(ang)
        R = np.array([[c,-s],[s,c]], np.float32)
        scale = np.random.uniform(0.9, 1.1)
        pts[:, :, :2] = (pts[:, :, :2] @ R.T) * scale
        pts += np.random.normal(0, 0.01, size=pts.shape).astype(np.float32)
        return pts.reshape(T, 63)

    def __getitem__(self, i):
        xi = self.X[i]
        if self.augment:
            xi = self._augment(xi)
        return torch.from_numpy(xi).float(), int(self.y[i])

class SeqGRU(nn.Module):
    def __init__(self, input_dim=63, hidden=128, num_classes=26):
        super().__init__()
        self.gru = nn.GRU(input_dim, hidden, batch_first=True, bidirectional=True)
        self.head = nn.Sequential(
            nn.Linear(hidden*2, 128),
            nn.ReLU(),
            nn.Dropout(0.2),
            nn.Linear(128, num_classes),
        )
    def forward(self, x):           # x: (B,T,63)
        h,_ = self.gru(x)           # (B,T,2H)
        h_last = h[:, -1, :]        # or mean over time: h.mean(1)
        return self.head(h_last)

def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--landmarks", default="landmarks_seq32", help="Folder from prep_sequence_resampled.py")
    ap.add_argument("--epochs", type=int, default=40)
    ap.add_argument("--batch", type=int, default=64)
    ap.add_argument("--lr", type=float, default=1e-3)
    ap.add_argument("--out", default="asl_seq32_gru.pt")
    args = ap.parse_args()

    # Load dataset
    trX = np.load(os.path.join(args.landmarks,"train_X.npy"))  # (N, T, 63)
    trY = np.load(os.path.join(args.landmarks,"train_y.npy"))
    vaX = np.load(os.path.join(args.landmarks,"val_X.npy"))
    vaY = np.load(os.path.join(args.landmarks,"val_y.npy"))
    classes = json.load(open(os.path.join(args.landmarks,"class_names.json")))
    meta = json.load(open(os.path.join(args.landmarks,"meta.json")))
    T = int(meta["frames"])

    print(f"Loaded: train {trX.shape}  val {vaX.shape}  classes={classes}")

    # Global mean/std over train (time+batch)
    X_mean = trX.reshape(-1, trX.shape[-1]).mean(axis=0, keepdims=True).astype(np.float32)  # (1,63)
    X_std  = trX.reshape(-1, trX.shape[-1]).std(axis=0, keepdims=True).astype(np.float32) + 1e-6
    trXn   = (trX - X_mean) / X_std
    vaXn   = (vaX - X_mean) / X_std

    tr_ds = SeqDataset(trXn, trY, augment=True)
    va_ds = SeqDataset(vaXn, vaY, augment=False)
    tr_dl = DataLoader(tr_ds, batch_size=args.batch, shuffle=True)
    va_dl = DataLoader(va_ds, batch_size=args.batch, shuffle=False)

    device = get_device()
    model = SeqGRU(input_dim=63, hidden=128, num_classes=len(classes)).to(device)
    crit = nn.CrossEntropyLoss()
    opt = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=1e-4)
    sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=args.epochs)

    best_acc, best_state = 0.0, None
    for epoch in range(1, args.epochs+1):
        # Train
        model.train()
        tot, correct, loss_sum = 0, 0, 0.0
        for xb, yb in tr_dl:
            xb, yb = xb.to(device), yb.to(device)
            opt.zero_grad(set_to_none=True)
            logits = model(xb)
            loss = crit(logits, yb)
            loss.backward()
            opt.step()
            loss_sum += loss.item() * yb.size(0)
            correct += (logits.argmax(1)==yb).sum().item()
            tot += yb.size(0)
        tr_loss = loss_sum / max(1, tot)
        tr_acc = correct / max(1, tot)

        # Validate
        model.eval()
        vtot, vcorrect = 0, 0
        with torch.no_grad():
            for xb, yb in va_dl:
                xb, yb = xb.to(device), yb.to(device)
                logits = model(xb)
                vcorrect += (logits.argmax(1)==yb).sum().item()
                vtot += yb.size(0)
        va_acc = vcorrect / max(1, vtot)
        sch.step()

        print(f"Epoch {epoch:02d}: train_loss={tr_loss:.4f} train_acc={tr_acc:.3f} val_acc={va_acc:.3f}")

        if va_acc > best_acc:
            best_acc = va_acc
            best_state = {
                "model": model.state_dict(),
                "classes": classes,
                "frames": T,
                "X_mean": torch.from_numpy(X_mean),   # tensors → future-proof
                "X_std":  torch.from_numpy(X_std),
            }
            torch.save(best_state, args.out)
            print(f"  ✅ Saved best → {args.out} (val_acc={best_acc:.3f})")

    print("Done. Best val_acc:", best_acc)

if __name__ == "__main__":
    main()