# ===============================
# IMPORTS
# ===============================
import os
import json
import math
import time
import pickle
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

# ===============================
# GPU SETUP
# ===============================
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))
    torch.backends.cudnn.benchmark = True

# ===============================
# DATA LOADING
# ===============================
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):
    df = pd.read_parquet(path)

    left = df[df["type"] == "left_hand"]
    right = df[df["type"] == "right_hand"]

    if len(left) > 0:
        hand = left
    elif len(right) > 0:
        hand = right
    else:
        return None

    landmarks = []
    for i in range(21):
        lm = hand[hand["landmark_index"] == i]
        if len(lm) == 0:
            landmarks.append([0.0, 0.0, 0.0])
        else:
            landmarks.append([
                lm["x"].mean(),
                lm["y"].mean(),
                lm["z"].mean()
            ])

    return np.array(landmarks, dtype=np.float32)


def get_features(landmarks):
    if landmarks is None:
        return None

    wrist = landmarks[0]
    points = landmarks - wrist

    scale = np.linalg.norm(points[9])
    if scale < 1e-6:
        scale = 1.0
    points /= scale

    mean = points.mean(axis=0)
    std = points.std(axis=0) + 1e-6
    points = (points - mean) / std

    features = points.flatten()

    tips = [4, 8, 12, 16, 20]
    bases = [1, 5, 9, 13, 17]

    tip_dist = []
    curl = []

    for b, t in zip(bases, tips):
        curl.append(np.linalg.norm(points[t] - points[b]))

    for i in range(len(tips) - 1):
        tip_dist.append(np.linalg.norm(points[tips[i]] - points[tips[i+1]]))

    return np.concatenate([features, tip_dist, curl]).astype(np.float32)


def process_row(row, base_path):
    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)
        feat = get_features(lm)
        return feat, row["sign"]
    except:
        return None, None


# ===============================
# LOAD + PROCESS DATA
# ===============================
base_path = "asl_kaggle"
train_df, sign_to_idx = load_kaggle_asl_data(base_path)

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

with Pool(cpu_count()) as pool:
    func = partial(process_row, base_path=base_path)
    for feat, sign in pool.map(func, rows):
        if feat is not None:
            X.append(feat)
            y.append(sign)

X = np.array(X, dtype=np.float32)
y = np.array(y)

print("Samples:", len(X))
print("Feature dim:", X.shape[1])

# ===============================
# LABEL ENCODING
# ===============================
le = LabelEncoder()
y = le.fit_transform(y)
num_classes = len(le.classes_)

# ===============================
# SPLIT
# ===============================
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

# ===============================
# DATASET
# ===============================
class ASLDataset(Dataset):
    def __init__(self, X, y):
        self.X = torch.tensor(X, dtype=torch.float32)
        self.y = torch.tensor(y, dtype=torch.long)

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

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


train_loader = DataLoader(
    ASLDataset(X_train, y_train),
    batch_size=256,
    shuffle=True,
    pin_memory=True
)

test_loader = DataLoader(
    ASLDataset(X_test, y_test),
    batch_size=256,
    shuffle=False,
    pin_memory=True
)

# ===============================
# MODEL (FIXED)
# ===============================
class TransformerASL(nn.Module):
    def __init__(self, input_dim, num_classes):
        super().__init__()

        self.proj = nn.Linear(input_dim, 256)
        self.norm = nn.LayerNorm(256)

        encoder_layer = nn.TransformerEncoderLayer(
            d_model=256,
            nhead=8,
            dim_feedforward=1024,
            dropout=0.1,
            activation="gelu",
            batch_first=True,
            norm_first=True
        )

        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=4)

        self.fc1 = nn.Linear(256, 512)
        self.bn1 = nn.BatchNorm1d(512)
        self.drop1 = nn.Dropout(0.4)

        self.fc2 = nn.Linear(512, 256)
        self.bn2 = nn.BatchNorm1d(256)
        self.drop2 = nn.Dropout(0.3)

        self.out = nn.Linear(256, num_classes)

    def forward(self, x):
        x = self.proj(x)
        x = self.norm(x)

        x = x.unsqueeze(1)      # (B, 1, 256)
        x = self.encoder(x)
        x = x.squeeze(1)

        x = F.gelu(self.bn1(self.fc1(x)))
        x = self.drop1(x)

        x = F.gelu(self.bn2(self.fc2(x)))
        x = self.drop2(x)

        return self.out(x)


model = TransformerASL(X.shape[1], num_classes).to(device)
print("Parameters:", sum(p.numel() for p in model.parameters()))

# ===============================
# TRAINING SETUP
# ===============================
criterion = nn.CrossEntropyLoss(label_smoothing=0.1)
optimizer = optim.AdamW(model.parameters(), lr=3e-4, weight_decay=1e-4)
scheduler = optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, 10)

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

    for x, y in train_loader:
        x, y = x.to(device), y.to(device)

        optimizer.zero_grad(set_to_none=True)
        logits = model(x)
        loss = criterion(logits, y)
        loss.backward()

        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()

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

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


@torch.no_grad()
def evaluate():
    model.eval()
    total, correct = 0, 0

    for x, y in test_loader:
        x, y = x.to(device), y.to(device)
        logits = model(x)
        correct += (logits.argmax(1) == y).sum().item()
        total += y.size(0)

    return 100 * correct / total


# ===============================
# TRAIN LOOP
# ===============================
best_acc = 0
patience = 15
wait = 0
epochs = 50

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

    print(f"Epoch {epoch+1}/{epochs} | "
          f"Loss {loss:.4f} | "
          f"Train {train_acc:.2f}% | "
          f"Test {test_acc:.2f}%")

    if test_acc > best_acc:
        best_acc = test_acc
        wait = 0
        torch.save({
            "model": model.state_dict(),
            "label_encoder": le
        }, "asl_transformer_fixed.pth")
    else:
        wait += 1

    if wait >= patience:
        print("Early stopping")
        break

print("Best accuracy:", best_acc)