This project is a from-scratch convolutional neural network built using PyTorch to classify architectural images. The model begins with simple binary classification (e.g., residential vs industrial) and is evolving toward a full multi-class architecture recognition system, including both cultural origins (e.g., Egyptian, Chinese, Greek) and stylistic classifications (e.g., Renaissance, Baroque).
What makes this project different is that it mirrors how humans learn — through meaningful memory. The model saves the most significant examples from each training run (best and worst) and reuses them during future training to simulate a kind of learning memory. This concept is known in machine learning as replay buffering or exemplar memory, but it was developed here from a natural human insight.
- CNN model built from scratch in PyTorch
- DuckDuckGo image scraping with multi-query support
- Training and validation dataset pipeline
- Memory replay buffer (save most & least accurate samples)
- Resume training with memory-injected batches
- Binary classifier (residential vs industrial)
- Multi-class cultural classifier (Chinese, Egyptian, European, Greek, Indian)
Rather than just train on more and more data, this model is designed to remember what mattered. Each epoch, it saves the 10 most confidently correct and 10 most confusing samples. These are injected into future training runs, ensuring that the model is learning not just more, but better.
This concept is based on how humans often return to defining moments in our learning journeys — both successes and failures — to anchor deeper understanding.
- Residential vs Industrial buildings using scraped data
- Categories like: Egyptian, Chinese, Greek, Indian, European
- Replay buffer extended to multi-class scenario
- Achieved ~79% accuracy on validation
- Classify architectural styles within cultural categories
- Examples: Renaissance, Baroque, Gothic, Brutalist, Neoclassical
- Use replay memory to auto-refine misclassifications
- Incorporate knowledge-based reasoning and image similarity
- Move closer to human-like adaptive learning
AI_Building_Classifier/
├── Data/
│ ├── Raw/ # Raw scraped images
│ ├── Training/ # Structured train/val folders
│ ├── Val/ # Structured train/val folders
├── models/ # Saved model checkpoints
├── replay_buffer/ # Stored key images from past runs
├── model.py # CNN architecture
├── train.py # Basic training loop
├── replay_train.py # Training with replay buffer
├── replay_multiclass_train.py # Multi-class training with memory
├── evaluate.py # Model evaluation script
├── download_image_cultures.py # Image scraper for cultural classes
├── dataset_splitter.py # Train/val image sorter
└── JOURNAL.md # Thought log and original ideas
| Class | Precision | Recall | F1-score |
|---|---|---|---|
| Chinese | 0.94 | 0.78 | 0.85 |
| Egyptian | 0.74 | 0.86 | 0.80 |
| European | 0.70 | 0.77 | 0.73 |
| Greek | 0.91 | 0.83 | 0.87 |
| Indian | 0.70 | 0.69 | 0.70 |
| Avg | 0.80 | 0.79 | 0.79 |
The model now incorporates a replay refinement loop, where it:
- Logs predictions on the validation set during training.
- Extracts high-confidence correct (
focus) and high-confidence incorrect (reinforce) samples. - Re-injects these samples during future training to focus on meaningful mistakes and strengths.
This mimics how humans review both best and worst examples to deepen understanding.
| Epoch | Train Acc | Val Acc | Replay Loss |
|---|---|---|---|
| 1 | 39.02% | 43.40% | 1.43 |
| 2 | 50.73% | 53.77% | 1.15 |
| 3 | 56.10% | 64.15% | 1.30 |
| 4 | 58.78% | 66.98% | 1.34 |
| 5 | 58.05% | 66.04% | 0.95 |
| 6 | 62.68% | 55.66% | 1.24 |
| 7 | 58.54% | 63.21% | 1.39 |
| 8 | 62.20% | 60.38% | 1.01 |
| 9 | 66.59% | 63.21% | 1.38 |
| 10 | 66.34% | 71.70% | 1.21 |
📝 Logs and plots are stored under
/checkpoints/and/plots/. Replay samples in/replay/. SeeJournal.mdfor full self-training idea flow.
This project is currently released under a source-available, non-commercial license.
You are free to:
- View, learn from, and experiment with the code for personal or educational use.
You may not:
- Use this project or its derivatives for commercial purposes (including SaaS, resale, client work, or monetized tools).
A future license may allow commercial usage under specific conditions.
See LICENSE.md for full details.
Some code co-developed with the help of ChatGPT-5.1.