File size: 10,702 Bytes
7c85344 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 | import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
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 pickle
import os
import time
import logging
import warnings
import gc
# --- Environment Optimization ---
warnings.filterwarnings("ignore", category=UserWarning)
logging.getLogger("codecarbon").setLevel(logging.ERROR)
# --- Configurations ---
DATA_DIR = r"C:\Users\shanm\Dataset Download\cifar-10-batches-py"
LOG_FILE = "eden_unfrozen_cifar10_efficientNet.csv"
MODEL_SAVE_PATH = "eden_unfrozen_efficientnet_v2_cifar10.pth"
BATCH_SIZE = 32
ACCUMULATION_STEPS = 4
LEARNING_RATE = 1e-3
NUM_EPOCHS = 20
UNFREEZE_EPOCH = 5 # Epoch to unlock the full network
L1_LAMBDA = 1e-5
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# --- Dataset Loader (RAM Cached for 64GB System) ---
class CIFAR10Binary(Dataset):
def __init__(self, root, train=True, transform=None):
self.data = []
self.labels = []
self.transform = transform
if train:
for i in range(1, 6):
file_path = os.path.join(root, f'data_batch_{i}')
with open(file_path, 'rb') as f:
entry = pickle.load(f, encoding='latin1')
self.data.append(entry['data'])
self.labels.extend(entry['labels'])
self.data = np.vstack(self.data).reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
else:
file_path = os.path.join(root, 'test_batch')
with open(file_path, 'rb') as f:
entry = pickle.load(f, encoding='latin1')
self.data = entry['data'].reshape(-1, 3, 32, 32).transpose(0, 2, 3, 1)
self.labels = entry['labels']
def __len__(self): return len(self.data)
def __getitem__(self, idx):
img, target = self.data[idx], self.labels[idx]
if self.transform:
img = self.transform(img)
return img, target
# --- Main Profiling Engine ---
def run_experiment():
torch.cuda.empty_cache()
gc.collect()
# 1. Transfer Learning Initialization (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 # Freeze features initially
model.classifier[1] = nn.Linear(model.classifier[1].in_features, 10) # 10 Classes for CIFAR-10
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())
transform = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
train_set = CIFAR10Binary(DATA_DIR, train=True, transform=transform)
loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True, num_workers=4, pin_memory=True)
# Initially only optimize the classifier
optimizer = optim.Adam(model.classifier.parameters(), lr=LEARNING_RATE)
criterion = nn.CrossEntropyLoss()
scaler = torch.cuda.amp.GradScaler()
cc_tracker = EmissionsTracker(measure_power_secs=1, save_to_file=False)
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"\nEDEN PROFILING STARTED | DEVICE: {torch.cuda.get_device_name(0)}")
print(f"Dataset: CIFAR-10 | Params: {total_params:,} | FLOPs: {total_flops:.2e}\n")
for epoch in range(NUM_EPOCHS):
# --- Stage 2: Progressive Unfreezing ---
if epoch + 1 == UNFREEZE_EPOCH:
print("\n[STAGE 2 INITIATED] Unfreezing Backbone for Fine-Tuning...")
for param in model.parameters():
param.requires_grad = True
# Re-initialize optimizer with a lower learning rate for the whole model
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()
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)
# Active Sparse Training (L1 Penalty) applied to currently trainable parameters
trainable_params = [p for p in model.parameters() if p.requires_grad]
l1_penalty = sum(p.abs().sum() for p in trainable_params)
# Calculate total loss for backprop, but DO NOT log it
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()
# Track ONLY the clean classification loss for the CSV
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)
# Rigorous Inference Latency (With Warm-up)
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. Terminal Update ---
print(f"Epoch {epoch+1} Summary:")
print(f" > Acc: {acc:.4f} | F1: {f1:.4f} | Loss: {running_loss/len(loader):.4f}")
print(f" > Epoch Energy: {epoch_total_j:.1f}J | EAG: {eag:.8f}")
print(f" > Avg Power: GPU {avg_gpu_w:.1f}W | VRAM: {vram_peak:.2f}GB | Latency: {lat_ms:.2f}ms")
print("-" * 65)
# --- 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()
# --- E. Save Optimized Model ---
torch.save(model.state_dict(), MODEL_SAVE_PATH)
print(f"\n[FINISH] Verified Optimization Complete.")
if __name__ == "__main__":
run_experiment() |