import os
import json
import math
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

from torch.utils.data import Dataset, DataLoader
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelEncoder, StandardScaler
from multiprocessing import Pool, cpu_count
from functools import partial
from tqdm import tqdm
from collections import Counter


def load_kaggle_asl_data(base_path):
    train_df = pd.read_csv(os.path.join(base_path, "train.csv"))
    with open(os.path.join(base_path, "sign_to_prediction_index_map.json")) as f:
        sign_to_idx = json.load(f)
    return train_df, sign_to_idx


def extract_hand_landmarks_from_parquet(path):
    try:
        df = pd.read_parquet(path)
        left = df[df["type"] == "left_hand"]
        right = df[df["type"] == "right_hand"]

        hand = left if len(left) >= len(right) else right

        if len(hand) == 0:
            return None

        frames = sorted(hand['frame'].unique())
        landmarks_seq = []

        for frame in frames:
            lm_frame = hand[hand['frame'] == frame]
            lm_list = []
            for i in range(21):
                lm = lm_frame[lm_frame['landmark_index'] == i]
                if len(lm) == 0:
                    lm_list.append([0.0, 0.0, 0.0])
                else:
                    lm_list.append([
                        float(lm['x'].iloc[0]),
                        float(lm['y'].iloc[0]),
                        float(lm['z'].iloc[0])
                    ])
            landmarks_seq.append(lm_list)

        return np.array(landmarks_seq, dtype=np.float32)
    except:
        return None


def get_features_sequence(landmarks_seq, max_frames=100):
    if landmarks_seq is None or len(landmarks_seq) == 0:
        return None

    # Center on wrist
    landmarks_seq -= landmarks_seq[:, 0:1, :]

    # Scale using index → middle finger tip distance (more stable than single point)
    scale = np.linalg.norm(landmarks_seq[:, 8] - landmarks_seq[:, 12], axis=1, keepdims=True)
    scale = np.maximum(scale, 1e-6)
    landmarks_seq /= scale

    # Flatten
    seq = landmarks_seq.reshape(landmarks_seq.shape[0], -1)

    # Pad / truncate
    if len(seq) < max_frames:
        pad = np.zeros((max_frames - len(seq), seq.shape[1]), dtype=np.float32)
        seq = np.concatenate([seq, pad])
    else:
        seq = seq[:max_frames]

    return seq


def process_row(row, base_path, max_frames=100):
    path = os.path.join(base_path, row["path"])
    if not os.path.exists(path):
        return None, None
    try:
        lm = extract_hand_landmarks_from_parquet(path)
        if lm is None:
            return None, None
        feat = get_features_sequence(lm, max_frames)
        if feat is None:
            return None, None
        return feat, row["sign"]
    except:
        return None, None


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_len=128):
        super().__init__()
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        self.register_buffer('pe', pe.unsqueeze(0))

    def forward(self, x):
        return x + self.pe[:, :x.size(1)]


class TransformerASL(nn.Module):
    def __init__(self, input_dim=63, num_classes=250, d_model=192, nhead=6, num_layers=4):
        super().__init__()
        self.proj = nn.Linear(input_dim, d_model)
        self.norm_in = nn.LayerNorm(d_model)
        self.pos = PositionalEncoding(d_model)

        enc_layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=nhead,
            dim_feedforward=d_model * 4,
            dropout=0.15,
            activation='gelu',
            batch_first=True,
            norm_first=True
        )
        self.encoder = nn.TransformerEncoder(enc_layer, num_layers=num_layers)

        self.head = nn.Sequential(
            nn.LayerNorm(d_model),
            nn.Dropout(0.25),
            nn.Linear(d_model, num_classes)
        )

    def forward(self, x, key_padding_mask=None):
        x = self.proj(x)
        x = self.norm_in(x)
        x = self.pos(x)
        x = self.encoder(x, src_key_padding_mask=key_padding_mask)
        x = x.mean(dim=1)  # global average pooling
        return self.head(x)


def create_padding_mask(lengths, max_len):
    return torch.arange(max_len, device=lengths.device)[None, :] >= lengths[:, None]


def main():
    # ===============================
    # DEVICE
    # ===============================
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")
    if device.type == "cuda":
        print("GPU:", torch.cuda.get_device_name(0))

    # ===============================
    # CONFIG
    # ===============================
    base_path = "asl_kaggle"  # ← CHANGE THIS TO YOUR ACTUAL PATH
    max_frames = 100
    MIN_SAMPLES_PER_CLASS = 6  # ← important! prevents stratified split crash

    # ===============================
    # DATA LOADING & PROCESSING
    # ===============================
    print("Loading metadata...")
    train_df, sign_to_idx = load_kaggle_asl_data(base_path)

    print(f"Processing {len(train_df)} videos...")
    rows = [row for _, row in train_df.iterrows()]

    with Pool(cpu_count()) as pool:
        results = list(tqdm(
            pool.imap(
                partial(process_row, base_path=base_path, max_frames=max_frames),
                rows
            ),
            total=len(rows),
            desc="Extracting landmarks"
        ))

    X_list, y_list = [], []
    for feat, sign in results:
        if feat is not None:
            X_list.append(feat)
            y_list.append(sign)

    if not X_list:
        print("No valid sequences found. Check parquet files / paths.")
        return

    X = np.stack(X_list)
    print(f"Loaded {len(X)} valid sequences | shape: {X.shape}")

    # Global normalization (very important for stability)
    print("Before global norm → mean:", X.mean(), "std:", X.std())
    X = np.clip(X, -5.0, 5.0)
    mean = X.mean(axis=(0, 1), keepdims=True)
    std = X.std(axis=(0, 1), keepdims=True) + 1e-8
    X = (X - mean) / std
    print("After global norm  → mean:", X.mean(), "std:", X.std())

    # ===============================
    # LABELS
    # ===============================
    le = LabelEncoder()
    y = le.fit_transform(y_list)

    # Remove classes with too few samples (prevents stratify error)
    counts = Counter(y)
    valid_classes = [cls for cls, cnt in counts.items() if cnt >= MIN_SAMPLES_PER_CLASS]

    mask = np.isin(y, valid_classes)
    X = X[mask]
    y = y[mask]

    # Re-encode labels consecutively (0,1,2,... no gaps)
    le = LabelEncoder()
    y = le.fit_transform(y)

    print(f"After filtering: {len(X)} samples remain | {len(le.classes_)} classes")

    # ===============================
    # SPLIT
    # ===============================
    X_train, X_test, y_train, y_test = train_test_split(
        X, y,
        test_size=0.15,
        stratify=y,  # should be safe now
        random_state=42
    )

    # ===============================
    # DATASET & LOADERS
    # ===============================
    class ASLSequenceDataset(Dataset):
        def __init__(self, X, y):
            self.X = torch.from_numpy(X).float()
            self.y = torch.from_numpy(y).long()

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

        def __getitem__(self, idx):
            return self.X[idx], self.y[idx]

    train_loader = DataLoader(
        ASLSequenceDataset(X_train, y_train),
        batch_size=64,
        shuffle=True,
        num_workers=4,
        pin_memory=True
    )

    test_loader = DataLoader(
        ASLSequenceDataset(X_test, y_test),
        batch_size=96,
        shuffle=False,
        num_workers=4,
        pin_memory=True
    )

    # ===============================
    # MODEL
    # ===============================
    model = TransformerASL(
        input_dim=63,
        num_classes=len(le.classes_),
        d_model=192,
        nhead=6,
        num_layers=4
    ).to(device)

    print(f"Model parameters: {sum(p.numel() for p in model.parameters()):,}")

    # ===============================
    # TRAINING SETUP
    # ===============================
    criterion = nn.CrossEntropyLoss(label_smoothing=0.05)
    optimizer = optim.AdamW(model.parameters(), lr=5e-4, weight_decay=1e-4)
    scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0=10)

    # ===============================
    # TRAIN / EVAL
    # ===============================
    def train_epoch():
        model.train()
        total_loss = 0
        correct = total = 0

        for x, y in tqdm(train_loader, desc="Train"):
            x, y = x.to(device), y.to(device)
            lengths = (x.abs().sum(dim=2) > 1e-5).sum(dim=1)
            mask = create_padding_mask(lengths, x.size(1))

            optimizer.zero_grad(set_to_none=True)
            logits = model(x, key_padding_mask=mask)

            loss = criterion(logits, y)
            loss.backward()

            grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=0.8)

            optimizer.step()

            total_loss += loss.item()
            correct += (logits.argmax(-1) == y).sum().item()
            total += y.size(0)

        return total_loss / len(train_loader), correct / total * 100

    @torch.no_grad()
    def evaluate():
        model.eval()
        correct = total = 0
        for x, y in test_loader:
            x, y = x.to(device), y.to(device)
            lengths = (x.abs().sum(dim=2) > 1e-5).sum(dim=1)
            mask = create_padding_mask(lengths, x.size(1))

            logits = model(x, key_padding_mask=mask)
            correct += (logits.argmax(-1) == y).sum().item()
            total += y.size(0)
        return correct / total * 100 if total > 0 else 0.0

    # ===============================
    # TRAINING LOOP
    # ===============================
    best_acc = 0.0
    patience = 15
    wait = 0
    epochs = 70

    for epoch in range(epochs):
        loss, train_acc = train_epoch()
        test_acc = evaluate()

        print(f"[{epoch + 1:2d}/{epochs}]  loss: {loss:.4f}  |  train: {train_acc:.2f}%  |  test: {test_acc:.2f}%")

        scheduler.step()

        if test_acc > best_acc:
            best_acc = test_acc
            wait = 0
            torch.save({
                'model': model.state_dict(),
                'optimizer': optimizer.state_dict(),
                'label_encoder_classes': le.classes_,
                'acc': best_acc,
                'epoch': epoch
            }, "best_asl_transformer.pth")
            print("   → New best saved")
        else:
            wait += 1
            if wait >= patience:
                print("Early stopping")
                break

    print(f"\nBest test accuracy reached: {best_acc:.2f}%")


if __name__ == '__main__':
    main()