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
    torch.backends.cudnn.enabled = True
else:
    print("✗ CUDA not available, using CPU")
    device = torch.device('cpu')

print("=" * 60)


# ===============================
# DATA LOADING - HANDLES PARTIAL NaN
# ===============================
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 - ONLY uses frames with valid (non-NaN) data"""
    try:
        df = pd.read_parquet(path)

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

        if len(left) == 0 and len(right) == 0:
            return None

        # Count valid (non-NaN) rows for each hand
        left_valid = 0
        right_valid = 0

        if len(left) > 0:
            left_valid = left[['x', 'y', 'z']].notna().all(axis=1).sum()
        if len(right) > 0:
            right_valid = right[['x', 'y', 'z']].notna().all(axis=1).sum()

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

        # Choose hand with more valid data
        hand = left if left_valid >= right_valid else right

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

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

            # Count how many valid landmarks this frame has
            valid_count = lm_frame[['x', 'y', 'z']].notna().all(axis=1).sum()

            # Skip frames with too few valid landmarks
            if valid_count < 10:
                continue

            # Extract landmarks for this frame
            frame_landmarks = []
            valid_landmarks_in_frame = 0

            for i in range(21):
                lm = lm_frame[lm_frame['landmark_index'] == i]

                if len(lm) == 0:
                    frame_landmarks.append([0.0, 0.0, 0.0])
                else:
                    x = float(lm['x'].iloc[0])
                    y = float(lm['y'].iloc[0])
                    z = float(lm['z'].iloc[0])

                    # Check if valid
                    if pd.notna(x) and pd.notna(y) and pd.notna(z):
                        frame_landmarks.append([x, y, z])
                        valid_landmarks_in_frame += 1
                    else:
                        frame_landmarks.append([0.0, 0.0, 0.0])

            # Only add frame if it has enough valid landmarks
            if valid_landmarks_in_frame >= 10:
                landmarks_seq.append(frame_landmarks)

        # Need at least 3 valid frames
        if len(landmarks_seq) < 3:
            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 (landmark 0)
    wrist = landmarks_seq[:, 0:1, :].copy()
    landmarks_seq = landmarks_seq - wrist

    # Scale normalization using wrist to middle finger MCP (landmark 9)
    scale = np.linalg.norm(landmarks_seq[:, 9, :] - np.zeros(3), axis=1, keepdims=True)
    scale = np.maximum(scale, 1e-6)  # Avoid division by zero
    landmarks_seq = landmarks_seq / scale[:, np.newaxis, :]

    # Clean up any remaining NaN/Inf
    landmarks_seq = np.nan_to_num(landmarks_seq, nan=0.0, posinf=0.0, neginf=0.0)

    # Clip extreme values
    landmarks_seq = np.clip(landmarks_seq, -10, 10)

    # Calculate finger curl features
    tips = [4, 8, 12, 16, 20]  # Thumb, index, middle, ring, pinky tips
    bases = [1, 5, 9, 13, 17]  # Corresponding base joints

    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)  # (T, 5)

    # Temporal deltas (motion)
    deltas = np.zeros_like(landmarks_seq)
    if len(landmarks_seq) > 1:
        deltas[1:] = landmarks_seq[1:] - landmarks_seq[:-1]

    # Flatten each component separately, then concatenate
    landmarks_flat = landmarks_seq.reshape(landmarks_seq.shape[0], -1)  # (T, 63)
    deltas_flat = deltas.reshape(deltas.shape[0], -1)  # (T, 63)
    # curl_features is already (T, 5)

    # Combine: 63 + 63 + 5 = 131 features per frame
    seq = np.concatenate([
        landmarks_flat,
        deltas_flat,
        curl_features
    ], axis=1)

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

    # Create attention mask (True for valid positions)
    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 - worker function for multiprocessing"""
    path = os.path.join(base_path, row["path"])

    if not os.path.exists(path):
        return None, None, None

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

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

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

        return feat, mask, row["sign"]

    except Exception as e:
        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, num_classes, d_model=256, nhead=8, num_layers=4):
        super().__init__()

        # Input projection
        self.proj = nn.Linear(input_dim, d_model)
        self.norm_in = nn.LayerNorm(d_model)
        self.pos = PositionalEncoding(d_model, max_len=128)

        # Transformer encoder
        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)

        # Classification head
        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: (batch, seq_len, input_dim)
        # key_padding_mask: (batch, seq_len) - True for padding positions

        x = self.proj(x)
        x = self.norm_in(x)
        x = self.pos(x)
        x = self.encoder(x, src_key_padding_mask=key_padding_mask)

        # Global average pooling over valid positions
        x = x.mean(dim=1)

        return self.head(x)


# ===============================
# MAIN TRAINING
# ===============================
def main():
    base_path = "asl_kaggle"
    max_frames = 100
    MIN_SAMPLES_PER_CLASS = 5

    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 (this will take a few minutes)...")
    print("Expected: ~36,000 valid sequences based on diagnostic")

    # Process with multiprocessing
    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 mask is not None and sign is not None:
            if feat.shape[0] == max_frames:
                X_list.append(feat)
                masks_list.append(mask)
                y_list.append(sign)

    print(f"\n✓ Successfully extracted: {len(X_list)} valid sequences")
    print(f"  Success rate: {len(X_list) / len(train_df) * 100:.1f}%")

    if len(X_list) < 100:
        print("❌ Too few valid sequences found!")
        print("   This shouldn't happen - please share this output for debugging")
        return

    # Stack into arrays
    X = np.stack(X_list)
    masks = np.stack(masks_list)

    print(f"\nData shape: {X.shape}")
    print(f"Feature dimension: {X.shape[2]}")

    # Global normalization
    print("Normalizing features...")
    X = np.clip(X, -10.0, 10.0)
    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 after filtering
    le = LabelEncoder()
    y = le.fit_transform(y)

    print(f"\nFinal dataset after filtering:")
    print(f"  Samples: {len(X):,}")
    print(f"  Classes: {len(le.classes_)}")
    print(f"  Sign examples: {list(le.classes_[:10])}")

    # 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
    )

    print(f"\nTrain set: {len(X_train):,} samples")
    print(f"Test set: {len(X_test):,} samples")

    # Dataset wrapper
    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]

    # DataLoaders
    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 if device.type == 'cuda' else False
    )

    test_loader = DataLoader(
        ASLSequenceDataset(X_test, masks_test, y_test),
        batch_size=batch_size * 2,
        shuffle=False,
        num_workers=4,
        pin_memory=True if device.type == 'cuda' else False
    )

    # Initialize model
    print("\nInitializing 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"Model parameters: {total_params:,}")

    # 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, T_mult=2)

    # Training loop
    best_acc = 0.0
    patience = 15
    wait = 0
    epochs = 60

    print("\n" + "=" * 60)
    print("STARTING TRAINING")
    print("=" * 60)

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

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

            # Invert mask: True for padding positions
            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=1.0)
            optimizer.step()

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

        train_acc = correct / total * 100

        # Evaluate
        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 progress
        print(f"[{epoch + 1:2d}/{epochs}] Loss: {total_loss / len(train_loader):.4f} | "
              f"Train: {train_acc:.2f}% | Test: {test_acc:.2f}%", end="")

        scheduler.step()

        # Save best model
        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_),
                'd_model': 256,
                'nhead': 8,
                'num_layers': 4
            }, "best_asl_transformer.pth")
            print(f" → New best: {best_acc:.2f}% ✓")
        else:
            wait += 1
            print()

            if wait >= patience:
                print(f"\nEarly stopping triggered at epoch {epoch + 1}")
                break

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


if __name__ == "__main__":
    main()