juangamerosalinas/trees-optimization

🌲 Tree Segmentation Performance Optimization Dataset

Fractional–Factorial Hyperparameter Search Results (64‑run, Resolution V DOE)

This dataset contains the experimental results from a 64‑run fractional factorial design (2⁸⁻² Resolution V) used to optimize hyperparameters for a SegFormer semantic segmentation model trained to detect trees.

---

📂 Dataset Structure

results/fractional_factorial_partial.csv

A cumulative CSV file updated after each experiment.
It contains all completed runs so far, enabling:

• real‑time monitoring
• ability to resume experiments
• incremental analysis

results/fractional_factorial_results.csv

The final CSV produced once all 64 runs finish.
It includes for each run:

• experiment ID
• fractional‑factorial coded levels (A–H)
• the decoded hyperparameters
• best‑epoch metrics for train, validation, and test splits
• training time

Both CSV files share the same schema but differ in completeness.

---

🧪 Experimental Design Overview

A 2⁸⁻² fractional factorial experiment was used with:

• 8 factors (A–H)
• 64 total runs
• Resolution V, allowing estimation of main effects and most two‑factor interactions
• Generators:
  • G = A × B × C × D
  • H = A × B × E × F

Factors A–F are independent; G and H are derived.

This design allows efficient exploration of a large hyperparameter space using only 64 experiments instead of 256.

---

🎛 Hyperparameter Coding

Each coded factor { -1, +1 } is mapped to an actual hyperparameter:

Factor	−1 Level	+1 Level
A	learning rate = 1e-5	1e-4
B	weight decay = 0.0	0.1
C	scheduler = linear	cosine
D	warmup ratio = 0.0	0.15
E	grad. accumulation = 1	4
F	epochs = 50	200
G	train batch size = 2	4
H	eval batch size = 2	4

The dataset includes both the coded values and the decoded hyperparameters.

---

🤖 Model & Training Setup

All experiments fine‑tune:

nvidia/segformer-b0-finetuned-ade-512-512

Key details:

• Metrics include:
  • IoU
  • accuracy
  • tree‑class precision, recall, Dice
• Metrics are computed for train, val, and test splits