Applied Deep Learning & Computer Vision Foundations

A curated collection of modular, production-focused deep learning implementations and computer vision labs. This repository covers end-to-end workflows: automated data scraping, dataset verification, transfer learning backbones, pixel-level semantic segmentation, categorical tabular embeddings, collaborative filtering, error diagnosis, and standalone model serialization.

---

Modality & Technical Coverage

Lab / Domain	Core Architecture	Dataset / Source	Key Techniques
01. Binary Vision	ResNet-18	DuckDuckGo Web Scraping	Lighting invariance, DataBlock API, Single-image inference
02. Semantic Segmentation	U-Net + ResNet-34	CamVid Tiny	Pixel-level classification, custom argmax metric, foreground accuracy
03. Tabular Classification	MLP + Categorical Embeddings	Adult Census Income	Embedding high-cardinality data, 1-Cycle policy, F1-Score
04. Recommendation Engine	Matrix Factorization	MovieLens 100k Sample	Latent factor embeddings, Sigmoid output bounding (y_range)
05. Systems & Foundations	Theoretical & Design Guide	Literature & Industry Notes	Drivetrain approach, NLP factuality, CV domain drift
06. Vision Pipeline & Cleaning	ResNet-18 + Active Cleanup	Web-Scraped Multi-Class	Transform benchmarks, GUI data sanitization, standalone .pkl export

---

Repository Structure

  deep-learning-cv-foundations/
  ├── assets/                                 # Visuals, plots, and architecture diagrams
  ├── notebooks/
  │   ├── 01_binary_image_classifier_resnet18.ipynb
  │   ├── 02_semantic_segmentation_camvid_unet.ipynb
  │   ├── 03_tabular_classification_adult_census.ipynb
  │   ├── 04_collaborative_filtering_movie_recommender.ipynb
  │   ├── 05_deep_learning_foundations_and_system_design.ipynb
  │   └── 06_multiclass_bear_classifier_and_data_cleaning.ipynb
  ├── requirements.txt                        # Core dependencies
  └── README.md

---

Notebook Index & Detailed Summaries

01. Binary Image Classification (ResNet-18)

• Task: Binary Image Classification (Bird vs. Forest).
• Backbone: ResNet-18 (Pretrained on ImageNet).
• Highlights:
  • Automated image retrieval pipeline querying multi-condition search terms (sun, shade) to enforce lighting invariance.
  • FastAI DataBlock construction with directory-derived labeling (parent_label) and corrupted image pruning (verify_images).
  • Transfer learning fine-tuning loop and standalone single-instance inference with softmax confidence scores.

02. Pixel-Level Semantic Segmentation (U-Net)

• Task: Pixel-Level Semantic Segmentation (CamVid Tiny).
• Backbone: U-Net with ResNet-34 Encoder Backbone.
• Highlights:
  • Lambda-based segmentation mask file association mapping input scenes to pixel-level ground truth classes.
  • Custom evaluation metric implementation (acc_seg) computing argmax pixel tensor matches.
  • Multi-metric performance tracking across validation loss, global pixel accuracy, and foreground-specific accuracy.

03. Tabular Deep Learning & Categorical Embeddings

• Task: Binary Income Classification ($>$50\text{K}$ vs. $\le$50\text{K}$).
• Architecture: Tabular Neural Network with Categorical Embeddings.
• Highlights:
  • Automated continuous feature scaling (Normalize), missing value imputation (FillMissing), and high-cardinality discrete encoding (Categorify).
  • Learning rate scheduling optimized via the 1-Cycle policy (fit_one_cycle).
  • Balanced performance evaluation using both accuracy and F1Score (precision-recall harmonic mean).

04. Collaborative Filtering Recommendation Engine

• Task: Continuous User-Movie Rating Prediction (MovieLens Sample).
• Architecture: Embedding-based Latent Factor Matrix Factorization.
• Highlights:
  • Ingestion and mapping of sparse user-item interaction pairs into dense continuous latent spaces.
  • Bounded output layer scaling (y_range=(0.5, 5.5)) using scaled sigmoids to constrain predictions to realistic rating ranges without gradient clipping.
  • Fine-tuning with discriminative learning rates to capture latent user taste vectors.

05. Deep Learning Foundations & System Design

• Focus: Deep Learning Theory, Modality Comparison & Production System Design.
• Highlights:
  • Computer Vision Taxonomy: Recognition vs. Object Detection vs. Semantic Segmentation and handling out-of-domain distribution shift.
  • NLP System Dynamics: Hallucination mitigation, generator vs. detector arms races, and speed vs. accuracy trade-offs.
  • The Drivetrain Approach: A 4-step framework for causal data products (Objective $\rightarrow$ Levers $\rightarrow$ Data $\rightarrow$ Models) and counterfactual two-model recommendation evaluation.
  • Framework Architecture: Modularity breakdown of fastai (Learner abstractions), fastcore (idiomatic Python power tools), and fastbook.

06. Multi-Class Vision Pipeline & Data Cleaning

• Task: 4-Class Classification (grizzly, black, teddy, polar).
• Backbone: ResNet-18 with Image Cleaning and Model Export.
• Highlights:
  • Fault-tolerant image scraper with automated exponential backoff retries and multi-threaded parallel downloads (n_workers=8).
  • Systematic transform benchmarking comparing Squish, Pad, RandomResizedCrop, and GPU batch augmentations (aug_transforms).
  • Diagnostics and model interpretation via confusion matrices and top loss extraction.
  • In-notebook data cleaning with ImageClassifierCleaner to delete noisy samples and re-route mislabeled images on disk.
  • Production serialization (learn.export) and standalone artifact loading (load_learner).

---

Technical Stack & Dependencies

• Core Frameworks: PyTorch, FastAI, Torchvision
• Data Processing & Utilities: NumPy, Pandas, FastCore, Pillow
• Data Gathering & UI: DuckDuckGo Search API (ddgs), ipywidgets

---

Quickstart & Environment Setup

# Clone the repository
git clone https://github.com/Subhrajyoti8520/deep-learning-cv-foundations.git
cd deep-learning-cv-foundations

# Create and activate virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install project dependencies
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# Install nbdev for notebook development
pip install nbdev

Launch Jupyter:

jupyter lab
# or
jupyter notebook

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.