My Journey into Deep Learning with Fast.ai

Mar 25, 2025

Over the past few months, I've embarked on a fascinating exploration into the world of deep learning, guided primarily by the accessible and powerful Fast.ai library, and enriched by the practical insights from the book Deep Learning for Coders with fastai and PyTorch by Jeremy Howard and Sylvain Gugger.

This blog post summarizes key insights, practical experiences, and reflections from working on various projects and assignments involving deep learning and computer vision.

Setting up for Success: Fast.ai and Jupyter Notebooks

Fast.ai, built upon PyTorch, allowed me to dive quickly into deep learning without needing a PhD in machine learning. Coupled with Jupyter notebooks, the iterative nature of model development became seamless and intuitive, as I was able to experiment interactively with code and visualize immediate outcomes.

Image Classification: From Scraping Data to Understanding Models

Data Collection

My first practical experience involved classifying images of airplanes, automobiles, birds, cats, and dogs. Using DuckDuckGo to scrape images, I quickly built a dataset using Fast.ai's intuitive DataBlock API:

dls = DataBlock(
    blocks=(ImageBlock, CategoryBlock),
    get_items=get_image_files,
    splitter=RandomSplitter(0.2, seed=42),
    get_y=parent_label,
    item_tfms=Resize(192)
).dataloaders(path)

This simplified the data preparation process dramatically, making iterative experimentation much easier.

Choosing the Right Loss Function: Focal Loss

Understanding that imbalanced data and "easy" examples might skew the training, I chose the Focal Loss function:

$
\text{Focal Loss} = -(1 - p_t)^\gamma \log(p_t)
$

This helped focus the model training on more challenging examples, enhancing overall performance.

Model Visualization and Interpretation

Understanding the model performance involved visual tools:

  • Confusion Matrices: Clearly showed misclassifications, highlighting confusion between visually similar classes like cats and dogs.
  • t-SNE Visualizations: Allowed me to inspect how well my model differentiated between categories in a two-dimensional space.

These tools were essential for debugging and improving my models.

Fingerprint Recognition Project

Beyond simple image classification, I developed a fingerprint recognition system featuring:

  • GUI Application for enrolling fingerprints with associated names.
  • ROC Curve Analysis: To optimize the threshold for fingerprint identification, minimizing false positives and negatives.

This project reinforced the real-world applicability of deep learning combined with traditional computer vision methods.

CPU vs GPU: Understanding Performance

An integral learning was about performance optimization. I compared training and fine-tuning runtimes across CPU and GPU environments. Contrary to expectation, some smaller batch sizes ran quicker on the CPU, reinforcing the importance of empirical testing.

GPU monitoring tools like nvtop gave me insights into memory usage and GPU efficiency, helping optimize the batch sizes effectively.

Complementing Deep Learning with Classical Computer Vision

Assignments involving MATLAB brought complementary insights:

  • Street Sign Detection: Utilized Gaussian blurring, edge detection, and cross-correlation.
  • Pantograph Cable Tracking: Applied Hough transforms to track movements effectively.

These classical methods showcased alternative effective approaches where deep learning might not always be feasible.

Key Reflections

My journey with Fast.ai illuminated several important lessons:

  • Accessibility and Power of Fast.ai: Deep learning doesn't require extensive theoretical backgrounds initially; practical tools and libraries allow immediate productivity.
  • Importance of Visualization: Tools like confusion matrices and t-SNE aren't just diagnostics�they are critical in iterative development.
  • Empirical Optimization: Real-world performance insights are essential; never assume performance without empirical validation.
  • Balance of Approaches: Classical image processing techniques are still valuable in modern computer vision projects.

Looking Forward

I plan to expand my deep learning toolkit further by:

  • Experimenting with advanced architectures like ResNet and EfficientNet.
  • Applying more sophisticated augmentation strategies such as MixUp and CutMix.
  • Deepening my understanding of model deployment and scalability in production environments.

Stay tuned for more updates on this exciting journey!

References

Ipeglin.no