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
from multiprocessing import Pool, cpu_count
from functools import partial
from tqdm import tqdm
from collections import Counter

# ===============================
# GPU CONFIGURATION
# ===============================
print("=" * 60)
print("GPU CONFIGURATION")
print("=" * 60)

if torch.cuda.is_available():
    print(f"✓ CUDA available!")
    print(f"✓ GPU: {torch.cuda.get_device_name(0)}")
    device = torch.device('cuda:0')
    torch.backends.cudnn.benchmark = True
else:
    print("✗ CUDA not available, using CPU")
    device = torch.device('cpu')

print("=" * 60)


# ===============================
# DATA LOADING WITH NaN HANDLING
# ===============================
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):
    """Extract hand landmarks, handling NaN values properly"""
    try:
        df = pd.read_parquet(path)

        # Get hand data
        left = df[df["type"] == "left_hand"]
        right = df[df["type"] == "right_hand"]

        # Choose hand with more non-NaN data
        left_valid = left[['x', 'y', 'z']].notna().all(axis=1).sum()
        right_valid = right[['x', 'y', 'z']].notna().all(axis=1).sum()

        if left_valid == 0 and right_valid == 0:
            return None  # No valid hand data

        hand = left if left_valid >= right_valid else right

        if len(hand) == 0:
            return None

        # Get frames with valid data
        frames = sorted(hand['frame'].unique())
        landmarks_seq = []

        for frame in frames:
            lm_frame = hand[hand['frame'] == frame]

            # Check if this frame has valid data
            valid_rows = lm_frame[['x', 'y', 'z']].notna().all(axis=1)
            if valid_rows.sum() < 10:  # Need at least 10 valid landmarks
                continue

            lm_list = []
            frame_has_data = False

            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:
                    x = lm['x'].iloc[0]
                    y = lm['y'].iloc[0]
                    z = lm['z'].iloc[0]

                    # Check for NaN
                    if pd.isna(x) or pd.isna(y) or pd.isna(z):
                        lm_list.append([0.0, 0.0, 0.0])
                    else:
                        lm_list.append([float(x), float(y), float(z)])
                        frame_has_data = True

            if frame_has_data:
                landmarks_seq.append(lm_list)

        if len(landmarks_seq) == 0:
            return None

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


def get_features_sequence(landmarks_seq, max_frames=100):
    """Extract features from landmark sequence"""
    if landmarks_seq is None or len(landmarks_seq) == 0:
        return None, None

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

    # Scale using wrist → middle finger MCP distance
    scale = np.linalg.norm(landmarks_seq[:, 0] - landmarks_seq[:, 9], axis=1, keepdims=True)
    scale = np.maximum(scale, 1e-6)
    landmarks_seq = landmarks_seq / scale[:, np.newaxis, :]

    # Replace any remaining NaN/Inf with 0
    landmarks_seq = np.nan_to_num(landmarks_seq, nan=0.0, posinf=0.0, neginf=0.0)

    # Finger curl distances
    tips = [4, 8, 12, 16, 20]
    bases = [1, 5, 9, 13, 17]
    curl_features = []
    for b, t in zip(bases, tips):
        curl = np.linalg.norm(landmarks_seq[:, t] - landmarks_seq[:, b], axis=1)
        curl_features.append(curl)
    curl_features = np.stack(curl_features, axis=1)

    # Temporal deltas
    deltas = np.zeros_like(landmarks_seq)
    deltas[1:] = landmarks_seq[1:] - landmarks_seq[:-1]

    # Flatten features
    seq = np.concatenate([landmarks_seq, deltas, curl_features[:, :, np.newaxis]], axis=2)
    seq = seq.reshape(seq.shape[0], -1)

    # Pad or truncate
    T, F = seq.shape
    if T < max_frames:
        pad = np.zeros((max_frames - T, F), dtype=np.float32)
        seq = np.concatenate([seq, pad], axis=0)
    elif T > max_frames:
        seq = seq[:max_frames, :]

    # Create mask
    valid_mask = np.zeros(max_frames, dtype=bool)
    valid_mask[:min(T, max_frames)] = True

    return seq.astype(np.float32), valid_mask


def process_row(row, base_path, max_frames=100):
    """Process a single row"""
    path = os.path.join(base_path, row["path"])
    if not os.path.exists(path):
        return None, None, None

    try:
        lm = extract_hand_landmarks_from_parquet(path)
        if lm is None:
            return None, None, None

        feat, mask = get_features_sequence(lm, max_frames)
        if feat is None:
            return None, None, None

        # Final NaN check
        if np.isnan(feat).any() or np.isinf(feat).any():
            return None, None, None

        return feat, mask, row["sign"]
    except:
        return None, None, None


# ===============================
# TRANSFORMER MODEL
# ===============================
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=68, num_classes=250, d_model=256, nhead=8, 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)
        return self.head(x)


# ===============================
# MAIN TRAINING
# ===============================
def main():
    base_path = "asl_kaggle"
    max_frames = 100
    MIN_SAMPLES_PER_CLASS = 6

    print("\nLoading metadata...")
    train_df, sign_to_idx = load_kaggle_asl_data(base_path)
    print(f"Total sequences: {len(train_df)}")

    rows = [row for _, row in train_df.iterrows()]

    print("\nProcessing sequences with NaN handling...")
    with Pool(cpu_count()) as pool:
        results = list(tqdm(
            pool.imap(
                partial(process_row, base_path=base_path, max_frames=max_frames),
                rows,
                chunksize=100
            ),
            total=len(rows),
            desc="Extracting landmarks"
        ))

    # Filter valid results
    X_list, masks_list, y_list = [], [], []
    for feat, mask, sign in results:
        if feat is not None and feat.shape[0] == max_frames:
            X_list.append(feat)
            masks_list.append(mask)
            y_list.append(sign)

    print(f"\n✓ Valid sequences: {len(X_list)} out of {len(train_df)}")

    if not X_list:
        print("❌ No valid sequences found!")
        print("\nPossible issues:")
        print("  1. Most files contain only NaN hand landmarks")
        print("  2. Hand detection failed in most videos")
        print("  3. Dataset might be corrupted")
        return

    X = np.stack(X_list)
    masks = np.stack(masks_list)
    print(f"Data shape: {X.shape}")

    # Global normalization
    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

    # Encode labels
    le = LabelEncoder()
    y = le.fit_transform(y_list)

    # Filter classes with too few samples
    counts = Counter(y)
    valid_classes = [cls for cls, cnt in counts.items() if cnt >= MIN_SAMPLES_PER_CLASS]
    mask_valid = np.isin(y, valid_classes)
    X = X[mask_valid]
    masks = masks[mask_valid]
    y = y[mask_valid]

    # Re-encode
    le = LabelEncoder()
    y = le.fit_transform(y)

    print(f"Final dataset: {len(X)} samples | {len(le.classes_)} classes")

    # Train-test split
    X_train, X_test, masks_train, masks_test, y_train, y_test = train_test_split(
        X, masks, y, test_size=0.15, stratify=y, random_state=42
    )

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

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

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

    batch_size = 128 if device.type == 'cuda' else 64

    train_loader = DataLoader(
        ASLSequenceDataset(X_train, masks_train, y_train),
        batch_size=batch_size, shuffle=True, num_workers=4, pin_memory=True
    )
    test_loader = DataLoader(
        ASLSequenceDataset(X_test, masks_test, y_test),
        batch_size=batch_size * 2, shuffle=False, num_workers=4, pin_memory=True
    )

    # Model
    model = TransformerASL(
        input_dim=X.shape[2],
        num_classes=len(le.classes_),
        d_model=256,
        nhead=8,
        num_layers=4
    ).to(device)

    total_params = sum(p.numel() for p in model.parameters())
    print(f"\nModel parameters: {total_params:,}")

    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)

    # Training
    best_acc = 0.0
    patience = 15
    wait = 0
    epochs = 70

    print("\nStarting training...")
    print("=" * 60)

    for epoch in range(epochs):
        model.train()
        total_loss = 0
        correct = total = 0

        for x, mask, yb in tqdm(train_loader, desc=f"Epoch {epoch + 1}/{epochs}"):
            x, mask, yb = x.to(device), mask.to(device), yb.to(device)
            key_mask = ~mask

            optimizer.zero_grad(set_to_none=True)
            logits = model(x, key_padding_mask=key_mask)
            loss = criterion(logits, yb)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=0.8)
            optimizer.step()

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

        train_acc = correct / total * 100

        # Eval
        model.eval()
        correct = total = 0
        with torch.no_grad():
            for x, mask, yb in test_loader:
                x, mask, yb = x.to(device), mask.to(device), yb.to(device)
                key_mask = ~mask
                logits = model(x, key_padding_mask=key_mask)
                correct += (logits.argmax(-1) == yb).sum().item()
                total += yb.size(0)

        test_acc = correct / total * 100

        print(f"[{epoch + 1:2d}/{epochs}] Loss: {total_loss / len(train_loader):.4f} | "
              f"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,
                'input_dim': X.shape[2],
                'num_classes': len(le.classes_)
            }, "best_asl_transformer.pth")
            print(f"  → New best: {best_acc:.2f}%")
        else:
            wait += 1
            if wait >= patience:
                print("Early stopping")
                break

    print("=" * 60)
    print(f"\n✓ Training complete!")
    print(f"✓ Best test accuracy: {best_acc:.2f}%")
    print(f"✓ Model saved: best_asl_transformer.pth")


if __name__ == "__main__":
    main()