# ===============================
# IMPORTS
# ===============================
import os
import json
import math
import time
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 & FEATURE EXTRACTION
# ===============================
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"]

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

    # Keep all frames
    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([
                    lm['x'].mean(),
                    lm['y'].mean(),
                    lm['z'].mean()
                ])
        landmarks_seq.append(lm_list)

    return np.array(landmarks_seq, dtype=np.float32)  # (T, 21, 3)

def get_features_sequence(landmarks_seq, max_frames=100):
    if landmarks_seq is None:
        return None
    # Center on wrist
    points = landmarks_seq - landmarks_seq[:, 0:1, :]
    scale = np.linalg.norm(points[:, 9, :], axis=1, keepdims=True)
    scale[scale < 1e-6] = 1.0
    points /= scale[:, np.newaxis, :]
    # Flatten per frame
    frames = points.reshape(points.shape[0], -1)
    # Pad or truncate
    if frames.shape[0] < max_frames:
        pad = np.zeros((max_frames - frames.shape[0], frames.shape[1]), dtype=np.float32)
        frames = np.vstack([frames, pad])
    else:
        frames = frames[:max_frames]
    return frames  # (max_frames, 63)

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_seq = extract_hand_landmarks_from_parquet(path)
        feat_seq = get_features_sequence(lm_seq, max_frames)
        return feat_seq, 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 = [], []

func = partial(process_row, base_path=base_path, max_frames=100)
with Pool(cpu_count()) as pool:
    for feat_seq, sign in pool.map(func, rows):
        if feat_seq is not None:
            X.append(feat_seq)
            y.append(sign)

X = np.stack(X)  # (N, T, 63)
y = np.array(y)
print("Samples:", len(X))
print("Sequence shape:", X.shape[1:])

# ===============================
# LABEL ENCODING
# ===============================
le = LabelEncoder()
y = le.fit_transform(y)
num_classes = len(le.classes_)
print("Num classes:", num_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 ASLSequenceDataset(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(ASLSequenceDataset(X_train, y_train), batch_size=64, shuffle=True, pin_memory=True)
test_loader = DataLoader(ASLSequenceDataset(X_test, y_test), batch_size=64, shuffle=False, pin_memory=True)

# ===============================
# TRANSFORMER MODEL
# ===============================
class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_len=100):
        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__()
        self.proj = nn.Linear(input_dim, d_model)
        self.norm = nn.LayerNorm(d_model)
        self.pos = PositionalEncoding(d_model)

        encoder_layer = nn.TransformerEncoderLayer(d_model=d_model, nhead=nhead, dim_feedforward=1024,
                                                   dropout=0.1, activation='gelu', batch_first=True, norm_first=True)
        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)

        self.fc = nn.Sequential(
            nn.Linear(d_model, 512),
            nn.BatchNorm1d(512),
            nn.GELU(),
            nn.Dropout(0.3),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        x = self.proj(x)
        x = self.norm(x)
        x = self.pos(x)
        x = self.encoder(x)           # (B, T, d_model)
        x = x.mean(dim=1)             # temporal average
        x = self.fc(x)
        return x

model = TransformerASL(input_dim=X.shape[2], num_classes=num_classes).to(device)
print("Parameters:", sum(p.numel() for p in model.parameters()))

# ===============================
# TRAIN 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, T_0=10)

# ===============================
# TRAIN / EVAL FUNCTIONS
# ===============================
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} | Loss {loss:.4f} | Train {train_acc:.2f}% | 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_full.pth")
    else:
        wait += 1
        if wait >= patience:
            print("Early stopping")
            break

print("Best accuracy:", best_acc)