Practice: Tuning Hyperparameters for a Model

Practical application of main concepts for optimizing deep learning models is the focus here. These concepts involve understanding the importance of weight initialization, how learning rate schedules can help convergence, and techniques like grid search and random search for finding good hyperparameter values. Learners will gain hands-on experience applying these ideas to tune a deep learning model.

Tuning hyperparameters is often more art than science, requiring experimentation. However, a systematic approach significantly increases your chances of finding a configuration that leads to better model performance and generalization.

Setting Up the Experiment

For this exercise, we'll use a common scenario: image classification using a simple Convolutional Neural Network (CNN) on the CIFAR-10 dataset. CIFAR-10 consists of 60,000 32x32 color images in 10 classes. We'll assume you have a basic PyTorch environment set up and are familiar with defining models, loading data, and writing training loops.

Our goal isn't to build the absolute best CIFAR-10 classifier, but rather to demonstrate the process of hyperparameter tuning.

First, let's define a simple CNN architecture using PyTorch. This will be our base model:

import torch
import torch.nn as nn
import torch.nn.functional as F

class SimpleCNN(nn.Module):
    def __init__(self, dropout_rate=0.5):
        super().__init__()
        # Convolutional layers
        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1) # Input: 3x32x32 -> Output: 16x32x32
        self.pool = nn.MaxPool2d(2, 2) # Output: 16x16x16
        self.conv2 = nn.Conv2d(16, 32, kernel_size=3, padding=1) # Output: 32x16x16
        # After pooling: 32x8x8

        # Fully connected layers
        self.fc1 = nn.Linear(32 * 8 * 8, 128) # Flattened size: 32*8*8 = 2048
        self.dropout = nn.Dropout(dropout_rate) # Apply dropout
        self.fc2 = nn.Linear(128, 10) # 10 output classes

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = torch.flatten(x, 1) # Flatten all dimensions except batch
        x = F.relu(self.fc1(x))
        x = self.dropout(x)
        x = self.fc2(x)
        return x

# Note: Weight initialization (like He init) is often handled by default
# in PyTorch layers, but could be explicitly set here if needed.

We also need standard data loading and transformation pipelines for CIFAR-10. We'll assume you have functions load_cifar10_data(batch_size) that return PyTorch DataLoader instances for training and validation sets. Remember to include normalization.

Identifying Hyperparameters and Search Space

Based on the chapter content, several hyperparameters are candidates for tuning:

1. Learning Rate ($\alpha$): Perhaps the most important hyperparameter. Affects convergence speed and final performance. We'll explore values typically ranging from $1e-4$ to $1e-2$. A logarithmic scale is often effective for searching learning rates.
2. Optimizer: We could compare Adam vs SGD with Momentum. Adam is often a good default, but SGD+Momentum can sometimes achieve slightly better generalization with careful tuning. For simplicity here, let's stick with Adam but tune its learning rate.
3. Weight Decay (L2 Regularization Strength, $\lambda$): Controls the penalty on large weights to prevent overfitting. Common values range from $0$ (no decay) to $1e-3$.
4. Dropout Rate: The probability of dropping neurons during training. Included in our SimpleCNN. Values typically range from $0.1$ to $0.5$.
5. Batch Size: Affects gradient estimation noise and training speed. Also interacts with the learning rate. Common values are powers of 2, like 32, 64, 128, 256.

We'll use Random Search for efficiency. Let's define the search space:

• Learning Rate: Sampled uniformly from a log scale between $10^{-4}$ and $10^{-2}$.
• Weight Decay: Sampled uniformly from a log scale between $10^{-5}$ and $10^{-3}$.
• Dropout Rate: Sampled uniformly between $0.1$ and $0.5$.
• Batch Size: Chosen randomly from the set {64, 128, 256}.

Implementing the Tuning Loop

The core idea is to run multiple training experiments (trials), each with a randomly sampled set of hyperparameters from our defined space. We train for a fixed, relatively small number of epochs (e.g., 10-15) to get a quick signal, record the validation performance, and then compare the results across trials.

Here's an outline of the tuning loop:

import random
import numpy as np
import torch.optim as optim

# Assume SimpleCNN, load_cifar10_data are defined
# Assume train_one_epoch() and evaluate() functions exist

num_trials = 20 # Number of random configurations to try
num_epochs_per_trial = 10 # Train for a short duration

results = []

for trial in range(num_trials):
    print(f"--- Trial {trial+1}/{num_trials} ---")

    # 1. Sample Hyperparameters
    lr = 10**np.random.uniform(-4, -2) # Log-uniform sampling for LR
    weight_decay = 10**np.random.uniform(-5, -3) # Log-uniform for weight decay
    dropout_rate = random.uniform(0.1, 0.5)
    batch_size = random.choice([64, 128, 256])

    print(f"Sampled: lr={lr:.6f}, wd={weight_decay:.6f}, dropout={dropout_rate:.4f}, batch_size={batch_size}")

    # 2. Setup Dataloaders, Model, Optimizer
    train_loader, val_loader = load_cifar10_data(batch_size=batch_size)
    model = SimpleCNN(dropout_rate=dropout_rate)
    # Consider using CUDA if available
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)

    optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay)
    criterion = nn.CrossEntropyLoss()

    best_val_accuracy = 0.0

    # 3. Train for fixed epochs
    for epoch in range(num_epochs_per_trial):
        # train_one_epoch(model, train_loader, criterion, optimizer, device)
        # val_loss, val_accuracy = evaluate(model, val_loader, criterion, device)

        # Dummy training/evaluation for structure illustration
        print(f"  Epoch {epoch+1}/{num_epochs_per_trial} - Simulating training...")
        # In a real run, update best_val_accuracy based on evaluate() results
        # For this example, let's simulate a result
        simulated_val_accuracy = 0.3 + trial*0.01 + epoch*0.02 + random.uniform(-0.05, 0.05) # Placeholder
        best_val_accuracy = max(best_val_accuracy, simulated_val_accuracy)

    print(f"Trial {trial+1} finished. Best Validation Accuracy: {best_val_accuracy:.4f}")

    # 4. Log Results
    results.append({
        'trial': trial + 1,
        'lr': lr,
        'weight_decay': weight_decay,
        'dropout_rate': dropout_rate,
        'batch_size': batch_size,
        'best_val_accuracy': best_val_accuracy
    })

# 5. Analyze Results (see next section)
print("\n--- Tuning Complete ---")
# Sort results by validation accuracy
results.sort(key=lambda x: x['best_val_accuracy'], reverse=True)

print("Top 5 configurations:")
for i in range(min(5, len(results))):
    print(f"Rank {i+1}: Acc={results[i]['best_val_accuracy']:.4f}, "
          f"LR={results[i]['lr']:.6f}, WD={results[i]['weight_decay']:.6f}, "
          f"Dropout={results[i]['dropout_rate']:.4f}, BS={results[i]['batch_size']}")

Note: The train_one_epoch and evaluate functions are standard PyTorch training components and are omitted here for brevity. You would implement them as usual.

Analyzing the Results

After running the tuning loop, the results list contains the performance for each hyperparameter configuration. Simply sorting by validation accuracy gives you the best-performing sets found during the search.

Visualizing the relationship between hyperparameters and performance can provide insights. For instance, let's plot validation accuracy against the learning rate (on a log scale):

Validation accuracy achieved by different randomly sampled learning rates after 10 training epochs. Values between roughly $10^{-3}$ and $3 \times 10^{-3}$ seem to perform best in this simulated run.

Similar plots can be made for weight decay and dropout rate. You might observe, for example, that very low or very high dropout rates hurt performance, or that a moderate amount of weight decay is beneficial. Analyzing the top-performing trials can help you understand which hyperparameter values (or ranges) are most promising.

Next Steps

• Refined Search: Based on the initial random search, you might identify promising ranges (e.g., learning rate between $5e-4$ and $3e-3$). You could then perform a second, more focused search (either random or grid) within these narrower ranges.
• Longer Training: The best hyperparameters found during a short run might not be optimal for full training. Once you have a few promising candidates, train them for more epochs with early stopping based on validation performance.
• Interaction Effects: Remember hyperparameters interact. The optimal learning rate might change depending on the batch size or the optimizer used. Random search explores some of these interactions. More advanced techniques like Bayesian Optimization explicitly model them.
• LR Schedules: We kept the learning rate constant here. Incorporating LR scheduling (e.g., StepLR, CosineAnnealingLR in PyTorch) adds more hyperparameters (decay factor, step size) but can often improve final results. You could include schedule choice or its parameters in your search space.

This hands-on practice demonstrates a fundamental workflow for improving deep learning models. While automated tools for hyperparameter optimization exist (e.g., Optuna, Ray Tune), understanding the manual process provides valuable intuition for setting up search spaces and interpreting results effectively. Experimentation is essential, so try adapting this process to your own models and datasets.