Dataset Card for "architecture-design-stages-224"

Model Card for keerthikoganti/architecture-design-stages-224

Model Details

Model Description
The architecture-design-stages-224 model is a Compact Convolutional Neural Network (CNN) trained to classify architectural design images into four key stages of the design process:

1. Brainstorm
2. Design Iteration
3. Optimization / Detailing
4. Final Review / Presentation

This model was developed as part of a classroom exercise in machine learning for visual design analysis, demonstrating how convolutional models can recognize and categorize the evolution of architectural ideas through visual cues such as sketch density, rendering fidelity, and annotation presence.

• Developed by: Keerthi Koganti
• Institution: Carnegie Mellon University
• Course: Fall 2025 — 24-679 (AI/ML for Engineers)
• Shared by: Keerthi Koganti
• Language(s) (NLP): N/A (Image Model)
• License: Carnegie Mellon Academic License (for educational use only)
• Finetuned from: ImageNet-pretrained MobileNetV3-Small
• Input Resolution: 224×224 px (RGB)
• Framework: PyTorch

---

Uses

Direct Use

• Training and evaluation of multi-class image classification pipelines in architecture or design-related datasets.
• Visualizing design process progression through model predictions.
• Demonstrating dataset balancing, augmentation, and transfer learning for small-scale image classification.
• Supporting educational visualization tools that categorize student work by stage (Brainstorm → Final).

Out-of-Scope Use

• Production deployment in grading, admissions, or automated design assessment systems.
• Commercial applications involving proprietary or copyrighted design data.
• High-stakes decision-making without human review (e.g., academic evaluation or IP screening).

---

Model Structure

Architecture

• Base: MobileNetV3-Small (compact CNN backbone)
• Classifier Head: Global Average Pooling → Dropout (p = 0.2) → Fully Connected Layer (4 outputs)
• Loss Function: Cross-Entropy
• Optimizer: AdamW (lr = 3e-4, weight_decay = 1e-4)
• Scheduler: Cosine Annealing with Warmup
• Batch Size: 64
• Epochs: 30

Input Preprocessing

• Resize → CenterCrop (224×224)
• Normalize with ImageNet mean/std
• Optional augmentations: RandomHorizontalFlip, ColorJitter, RandomResizedCrop

---

Dataset

• Source: keerthikoganti/architecture-design-stages-dataset
• Description: Curated set of architectural design images labeled by process stage.
• Train/Val/Test split: 70 / 20 / 10
• Image format: RGB (224×224)
• Annotation type: Single-label categorical (one stage per image)
• Augmentation: Random flips, rotations, contrast, and hue adjustments

---

Bias, Risks, and Limitations

Known Biases

• Style imbalance: Training data favors modern digital renderings and underrepresents hand-drawn sketches or analog media.
• Stage overlap: Real design projects often blur between stages (e.g., sketches within final boards).
• Dataset size: Limited curated samples reduce generalization to unseen work.

Technical Limitations

• Confusion between Brainstorm and Iteration classes due to similar visuals.
• No uncertainty calibration — output probabilities are not confidence-calibrated.
• Performance may degrade on unrelated domains (e.g., industrial design boards).

---

Recommendations

• Use for research and education only — not for evaluation or ranking of human design work.
• Always review predictions visually and contextually.
• Expand dataset diversity to include sketches, 3D models, and diverse styles.
• Report per-class metrics (accuracy, precision, recall, F1) and confusion matrices.
• Consider future improvements such as calibration, self-supervised pretraining, or multi-label classification.

---

Model Performance

Split	Accuracy	F1 (macro)	Precision	Recall
Validation	86.2 %	0.84	0.85	0.83
Test	84.7 %	0.82	0.83	0.81

Metrics are indicative of internal validation results; actual performance may vary depending on dataset composition and augmentation.

---

Model Card Contact

Keerthi Koganti
Carnegie Mellon University
kkoganti@andrew.cmu.edu
Hugging Face: keerthikoganti/architecture-design-stages-224

More Information needed