import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torchvision.datasets import ImageFolder
import torchvision.transforms as transforms
from codecarbon import EmissionsTracker
from carbontracker.tracker import CarbonTracker
from fvcore.nn import FlopCountAnalysis
from sklearn.metrics import precision_recall_fscore_support, accuracy_score
from einops import rearrange
from tqdm import tqdm
import pandas as pd
import numpy as np
import os
import time
import logging
import warnings
import gc

# --- Environment & Logging Optimization ---
warnings.filterwarnings("ignore", category=UserWarning) 
# Hard-mute CodeCarbon terminal spam
logging.getLogger("codecarbon").setLevel(logging.CRITICAL)
logging.getLogger("codecarbon").disabled = True 

# --- Configurations ---
DATA_DIR = r"C:\Users\shanm\Dataset Download\custom image net"
LOG_FILE = "eden_optimized_custom_imagenet_mobilevitv3.csv"
MODEL_SAVE_PATH = "eden_optimized_custom_mobilevitv3_imagenet.pth"

BATCH_SIZE = 32
ACCUMULATION_STEPS = 4
LEARNING_RATE = 1e-3
NUM_EPOCHS = 30
L1_LAMBDA = 1e-5        
NUM_CLASSES = 300       # Matched to your 300 custom folders

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# ==========================================
# 1. MOBILEVITV3 ARCHITECTURE DEFINITION
# ==========================================

def conv_1x1_bn(inp, oup):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
        nn.SiLU()
    )

def conv_nxn_bn(inp, oup, kernel_size=3, stride=1):
    return nn.Sequential(
        nn.Conv2d(inp, oup, kernel_size, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        nn.SiLU()
    )

class PreNorm(nn.Module):
    def __init__(self, dim, fn):
        super().__init__()
        self.norm = nn.LayerNorm(dim)
        self.fn = fn
    def forward(self, x, **kwargs):
        return self.fn(self.norm(x), **kwargs)

class FeedForward(nn.Module):
    def __init__(self, dim, hidden_dim, dropout=0.):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.SiLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout)
        )
    def forward(self, x):
        return self.net(x)

# Hardware-Fused Attention Kernel for Maximum Speed
class Attention(nn.Module):
    def __init__(self, dim, heads=8, dim_head=64, dropout=0.):
        super().__init__()
        inner_dim = dim_head * heads
        project_out = not (heads == 1 and dim_head == dim)

        self.heads = heads
        self.dropout_rate = dropout 

        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)

        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, dim),
            nn.Dropout(dropout)
        ) if project_out else nn.Identity()

    def forward(self, x):
        qkv = self.to_qkv(x).chunk(3, dim=-1)
        q, k, v = map(lambda t: rearrange(t, 'b p n (h d) -> b p h n d', h=self.heads), qkv)

        # PyTorch native SDPA triggers FlashAttention
        out = F.scaled_dot_product_attention(
            q, k, v, 
            dropout_p=self.dropout_rate if self.training else 0.0
        )

        out = rearrange(out, 'b p h n d -> b p n (h d)')
        return self.to_out(out)

class Transformer(nn.Module):
    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout=0.):
        super().__init__()
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
                PreNorm(dim, Attention(dim, heads, dim_head, dropout)),
                PreNorm(dim, FeedForward(dim, mlp_dim, dropout))
            ]))
    def forward(self, x):
        for attn, ff in self.layers:
            x = attn(x) + x
            x = ff(x) + x
        return x

class MV2Block(nn.Module):
    def __init__(self, inp, oup, stride=1, expansion=4):
        super().__init__()
        self.stride = stride
        hidden_dim = int(inp * expansion)
        self.use_res_connect = self.stride == 1 and inp == oup

        if expansion == 1:
            self.conv = nn.Sequential(
                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
                nn.BatchNorm2d(hidden_dim),
                nn.SiLU(),
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup),
            )
        else:
            self.conv = nn.Sequential(
                nn.Conv2d(inp, hidden_dim, 1, 1, 0, bias=False),
                nn.BatchNorm2d(hidden_dim),
                nn.SiLU(),
                nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 1, groups=hidden_dim, bias=False),
                nn.BatchNorm2d(hidden_dim),
                nn.SiLU(),
                nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
                nn.BatchNorm2d(oup),
            )

    def forward(self, x):
        if self.use_res_connect:
            return x + self.conv(x)
        else:
            return self.conv(x)

class MobileViTBlock(nn.Module):
    def __init__(self, dim, depth, channel, kernel_size, patch_size, mlp_dim, dropout=0.):
        super().__init__()
        self.ph, self.pw = patch_size

        self.conv1 = conv_nxn_bn(channel, channel, kernel_size)
        self.conv2 = conv_1x1_bn(channel, dim)

        self.transformer = Transformer(dim, depth, 1, 32, mlp_dim, dropout)

        self.conv3 = conv_1x1_bn(dim, channel)
        self.conv4 = conv_nxn_bn(2 * channel, channel, kernel_size)
    
    def forward(self, x):
        y = x.clone()

        x = self.conv1(x)
        x = self.conv2(x)
        
        _, _, h, w = x.shape
        pad_h = (self.ph - h % self.ph) % self.ph
        pad_w = (self.pw - w % self.pw) % self.pw
        
        if pad_h > 0 or pad_w > 0:
            x = nn.functional.pad(x, (0, pad_w, 0, pad_h))
        
        _, _, h_pad, w_pad = x.shape
        x = rearrange(x, 'b d (h ph) (w pw) -> b (ph pw) (h w) d', ph=self.ph, pw=self.pw)
        x = self.transformer(x)
        x = rearrange(x, 'b (ph pw) (h w) d -> b d (h ph) (w pw)', h=h_pad//self.ph, w=w_pad//self.pw, ph=self.ph, pw=self.pw)

        if pad_h > 0 or pad_w > 0:
            x = x[:, :, :h, :w]

        x = self.conv3(x)
        x = torch.cat((x, y), 1)
        x = self.conv4(x)
        return x

class MobileViTv3_Small(nn.Module):
    def __init__(self, image_size=(224, 224), num_classes=300): # Updated for 300 Classes
        super().__init__()
        ih, iw = image_size
        ph, pw = 2, 2
        
        dims = [144, 192, 240]
        channels = [16, 32, 64, 64, 96, 96, 128, 128, 160, 160, 640]
        
        self.conv1 = conv_nxn_bn(3, channels[0], stride=2)

        self.mv2 = nn.ModuleList([])
        self.mv2.append(MV2Block(channels[0], channels[1], 1, 4))
        self.mv2.append(MV2Block(channels[1], channels[2], 2, 4))
        self.mv2.append(MV2Block(channels[2], channels[3], 1, 4))
        self.mv2.append(MV2Block(channels[3], channels[4], 2, 4))
        
        self.mvit = nn.ModuleList([])
        self.mvit.append(MobileViTBlock(dims[0], 2, channels[5], 3, (ph, pw), int(dims[0]*2)))
        
        self.mv2_2 = nn.ModuleList([])
        self.mv2_2.append(MV2Block(channels[5], channels[6], 2, 4))
        
        self.mvit_2 = nn.ModuleList([])
        self.mvit_2.append(MobileViTBlock(dims[1], 4, channels[7], 3, (ph, pw), int(dims[1]*2)))
        
        self.mv2_3 = nn.ModuleList([])
        self.mv2_3.append(MV2Block(channels[7], channels[8], 2, 4))
        
        self.mvit_3 = nn.ModuleList([])
        self.mvit_3.append(MobileViTBlock(dims[2], 3, channels[9], 3, (ph, pw), int(dims[2]*2)))
        
        self.conv2 = conv_1x1_bn(channels[9], channels[10])
        self.pool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(channels[10], num_classes)

    def forward(self, x):
        x = self.conv1(x)
        for conv in self.mv2: x = conv(x)
        for m in self.mvit: x = m(x)
        for conv in self.mv2_2: x = conv(x)
        for m in self.mvit_2: x = m(x)
        for conv in self.mv2_3: x = conv(x)
        for m in self.mvit_3: x = m(x)
        x = self.conv2(x)
        x = self.pool(x).view(-1, x.shape[1])
        return self.fc(x)

# ==========================================
# 2. EDEN EXPERIMENT & PROFILING
# ==========================================

def run_experiment():
    torch.backends.cudnn.benchmark = True
    torch.cuda.empty_cache()
    gc.collect()
    
    # Initialize custom architecture (Training from scratch)
    model = MobileViTv3_Small(image_size=(224, 224), num_classes=NUM_CLASSES)
    model = model.to(DEVICE)
    
    optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)

    dummy_input = torch.randn(1, 3, 224, 224).to(DEVICE)
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        total_flops = FlopCountAnalysis(model, dummy_input).total()
    total_params = sum(p.numel() for p in model.parameters())
    
    # --- Data Loading: Minimal CPU Processing ---
    cpu_transform = transforms.Compose([
        transforms.Resize((224, 224)),
        transforms.ToTensor()
    ])
    
    # Directly loads from the 300 custom class folders
    train_set = ImageFolder(root=DATA_DIR, transform=cpu_transform)
    loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=4, pin_memory=True)

    criterion = nn.CrossEntropyLoss()
    scaler = torch.cuda.amp.GradScaler() 
    
    # --- 3. Profiling Initialization (SILENCED) ---
    cc_tracker = EmissionsTracker(measure_power_secs=1, save_to_file=False, log_level="critical")
    ct_tracker = CarbonTracker(epochs=NUM_EPOCHS, monitor_epochs=NUM_EPOCHS, update_interval=1)
    
    cc_tracker.start()
    all_logs = []
    total_iterations_counter = 0
    session_start_time = time.time()

    prev_cum_gpu_j, prev_cum_cpu_j, prev_cum_ram_j = 0.0, 0.0, 0.0

    # ImageNet Standard Normalizer pushed directly to the GPU
    gpu_normalizer = transforms.Normalize(
        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
    ).to(DEVICE)

    print(f"\n[EDEN PROFILING STARTED] | Model: Custom MobileViTv3-Small | Classes: {NUM_CLASSES}")
    print(f"Dataset: Custom ImageNet ({len(train_set)} images) | Saving quietly to CSV...\n")

    for epoch in range(NUM_EPOCHS):
        ct_tracker.epoch_start()
        torch.cuda.reset_peak_memory_stats() 
        epoch_start_time = time.time()
        model.train()
        
        running_loss = 0.0
        all_preds, all_labels = [], []
        epoch_grad_norms = []
        
        optimizer.zero_grad()
        pbar = tqdm(loader, desc=f"Epoch {epoch+1}/{NUM_EPOCHS}", unit="batch", leave=False)
        
        for i, (images, labels) in enumerate(pbar):
            # GPU-Accelerated Normalization
            images, labels = images.to(DEVICE), labels.to(DEVICE)
            images = gpu_normalizer(images)

            with torch.cuda.amp.autocast():
                outputs = model(images)
                loss = criterion(outputs, labels)
                
                # Active Sparse Training (L1 Penalty) applied natively
                l1_penalty = sum(p.abs().sum() for p in model.parameters())
                total_loss = loss + (L1_LAMBDA * l1_penalty)
                scaled_loss = total_loss / ACCUMULATION_STEPS

            scaler.scale(scaled_loss).backward()
            
            grad_norm = 0.0
            for p in model.parameters():
                if p.grad is not None:
                    grad_norm += p.grad.data.norm(2).item() ** 2
            epoch_grad_norms.append(grad_norm ** 0.5)

            if (i + 1) % ACCUMULATION_STEPS == 0:
                scaler.step(optimizer)
                scaler.update()
                optimizer.zero_grad()

            running_loss += loss.item() * ACCUMULATION_STEPS
            
            _, preds = torch.max(outputs, 1)
            all_preds.extend(preds.cpu().numpy())
            all_labels.extend(labels.cpu().numpy())
            total_iterations_counter += 1
            
            pbar.set_postfix(loss=f"{(loss.item()*ACCUMULATION_STEPS):.4f}")

        # --- A. Evaluation ---
        ct_tracker.epoch_end()
        epoch_end_time = time.time()
        epoch_duration = epoch_end_time - epoch_start_time
        avg_it_per_sec = len(loader) / epoch_duration
        
        acc = accuracy_score(all_labels, all_preds)
        p, r, f1, _ = precision_recall_fscore_support(all_labels, all_preds, average='macro', zero_division=0)
        
        model.eval()
        with torch.no_grad():
            sample_img = torch.randn(1, 3, 224, 224).to(DEVICE)
            _ = model(sample_img) 
            torch.cuda.synchronize()
            
            starter, ender = torch.cuda.Event(enable_timing=True), torch.cuda.Event(enable_timing=True)
            starter.record()
            _ = model(sample_img) 
            ender.record()
            torch.cuda.synchronize()
            lat_ms = starter.elapsed_time(ender)

        # --- B. Energy & Power Calculations ---
        emissions_data = cc_tracker._prepare_emissions_data()
        
        cum_gpu_j = emissions_data.gpu_energy * 3.6e6
        cum_cpu_j = emissions_data.cpu_energy * 3.6e6
        cum_ram_j = emissions_data.ram_energy * 3.6e6
        cum_total_j = cum_gpu_j + cum_cpu_j + cum_ram_j
        
        epoch_gpu_j = cum_gpu_j - prev_cum_gpu_j
        epoch_cpu_j = cum_cpu_j - prev_cum_cpu_j
        epoch_ram_j = cum_ram_j - prev_cum_ram_j
        epoch_total_j = epoch_gpu_j + epoch_cpu_j + epoch_ram_j
        
        prev_cum_gpu_j, prev_cum_cpu_j, prev_cum_ram_j = cum_gpu_j, cum_cpu_j, cum_ram_j

        avg_gpu_w = epoch_gpu_j / epoch_duration if epoch_duration > 0 else 0
        avg_cpu_w = epoch_cpu_j / epoch_duration if epoch_duration > 0 else 0
        avg_ram_w = epoch_ram_j / epoch_duration if epoch_duration > 0 else 0

        vram_peak = torch.cuda.max_memory_allocated(DEVICE) / (1024**3)

        # --- C. Minimal Terminal Update ---
        print(f"Epoch {epoch+1}/{NUM_EPOCHS} | Acc: {acc:.4f} | Loss: {running_loss/len(loader):.4f} | Energy: {epoch_total_j:.1f}J | Latency: {lat_ms:.2f}ms")

        # --- D. Unified Verified CSV Logging ---
        log_entry = {
            "epoch": epoch + 1,
            "loss": running_loss / len(loader),
            "accuracy": acc, "f1_score": f1, "precision": p, "recall": r,
            "epoch_energy_gpu_j": epoch_gpu_j, "epoch_energy_cpu_j": epoch_cpu_j,
            "epoch_energy_ram_j": epoch_ram_j, "epoch_total_energy_j": epoch_total_j,
            "cumulative_total_energy_j": cum_total_j, "carbon_emissions_kg": emissions_data.emissions,
            "avg_power_gpu_w": avg_gpu_w, "avg_power_cpu_w": avg_cpu_w, "avg_power_ram_w": avg_ram_w,
            "vram_peak_gb": vram_peak, "latency_ms": lat_ms, "avg_grad_norm": np.mean(epoch_grad_norms),
            "it_per_sec": avg_it_per_sec, "total_iterations": total_iterations_counter,
            "epoch_duration_sec": epoch_duration, "cumulative_time_sec": time.time() - session_start_time
        }
        all_logs.append(log_entry)
        pd.DataFrame(all_logs).to_csv(LOG_FILE, index=False)

    cc_tracker.stop()
    
    torch.save(model.state_dict(), MODEL_SAVE_PATH)
    print(f"\n[FINISH] Verified Optimization Complete. Model and CSV Saved.")

if __name__ == "__main__":
    run_experiment()