import torch import torch.nn as nn import torch.optim as optim from torch.utils.data import DataLoader from torchvision.datasets import ImageFolder import torchvision.transforms as transforms import torchvision.models as models from torchvision.models import EfficientNet_V2_S_Weights 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 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_unfrozen_custom_imagenet_efficientNet.csv" MODEL_SAVE_PATH = "eden_unfrozen_efficientnet_v2_custom_imagenet.pth" BATCH_SIZE = 32 ACCUMULATION_STEPS = 4 LEARNING_RATE = 1e-3 NUM_EPOCHS = 30 UNFREEZE_EPOCH = 5 L1_LAMBDA = 1e-5 NUM_CLASSES = 300 # Matched to your 300 custom folders DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu") def run_experiment(): torch.backends.cudnn.benchmark = True torch.cuda.empty_cache() gc.collect() # --- 1. Transfer Learning Setup (Stage 1: Frozen) --- weights = EfficientNet_V2_S_Weights.DEFAULT model = models.efficientnet_v2_s(weights=weights) for param in model.features.parameters(): param.requires_grad = False model.classifier[1] = nn.Linear(model.classifier[1].in_features, NUM_CLASSES) model = model.to(DEVICE) 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()) # --- 2. Dataset Setup --- transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) ]) # Directly loads from the 300 custom class folders train_set = ImageFolder(root=DATA_DIR, transform=transform) loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=4, pin_memory=True) optimizer = optim.Adam(model.classifier.parameters(), lr=LEARNING_RATE) criterion = nn.CrossEntropyLoss() scaler = torch.cuda.amp.GradScaler() # --- 3. Profiling Initialization (SILENCED) --- # Passing log_level="critical" forces CodeCarbon to stay off the terminal 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 prev_acc = 0.0 print(f"\n[EDEN PROFILING STARTED] | Model: EfficientNetV2-S | Classes: {NUM_CLASSES}") print(f"Dataset: Custom ImageNet ({len(train_set)} images) | Saving quietly to CSV...\n") for epoch in range(NUM_EPOCHS): # --- Stage 2: Progressive Unfreezing --- if epoch + 1 == UNFREEZE_EPOCH: print(f"\n[Epoch {epoch+1}] Unfreezing Backbone for Fine-Tuning...") for param in model.parameters(): param.requires_grad = True optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE * 0.1) 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() # tqdm progress bar left on to track batch speed pbar = tqdm(loader, desc=f"Epoch {epoch+1}/{NUM_EPOCHS}", unit="batch", leave=False) for i, (images, labels) in enumerate(pbar): images, labels = images.to(DEVICE), labels.to(DEVICE) with torch.cuda.amp.autocast(): outputs = model(images) loss = criterion(outputs, labels) trainable_params = [p for p in model.parameters() if p.requires_grad] l1_penalty = sum(p.abs().sum() for p in trainable_params) 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.requires_grad and 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) acc_gain = acc - prev_acc eag = acc_gain / epoch_total_j if epoch_total_j > 0 else 0 prev_acc = acc # --- 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), "eag_metric": eag, "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()