Skip to content

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

CSE5104 FL26 Project 1: Concrete Compressive Strength

Linear regression analysis of the UCI Concrete Compressive Strength dataset (1030 instances), with gradient descent implemented from scratch.

Data split: rows 501–630 (1-indexed) are the 130-sample test set; the remaining 900 rows are the training set. Test data is never used to update model parameters, and all feature transformations are fit on the training set only.

Setup

Requires Python 3.9+. No GPU needed; the full analysis runs on CPU in seconds.

python3 -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Using conda instead? A venv inside an activated conda env works fine, or:

conda create -n cse5104 python=3.12 --file requirements.txt
conda activate cse5104

Avoid installing into conda base — mixing pip and conda there is what breaks numpy/pandas ABI compatibility.

Reproducing the results

bash run_all.sh runs every analysis script below in order (well under a minute); bash run_all.sh --reports also rebuilds the Part 2/3 reports. To run scripts individually, run them from the repository root. Each script is independent. Figures go to assets/figures/, tables to assets/tables/.

# Q2/Q3: gradient descent, scaling & convergence speed
#   -> figures/gradient_descent_convergence.png
python scripts/gradient_descent_self_contained.py

# Q5: log transformation, all three loss curves
#   -> figures/log_transformation_convergence.png
python scripts/log_transformation.py

# Q5: summary table across the three optimal models
#   -> tables/summary_table.csv
python scripts/summary_table.py

# Q6: univariate vs. multivariate coefficients, omitted-variable decomposition
#   -> tables/univariate_vs_multivariate.csv, tables/feature_correlations.csv,
#      figures/feature_correlations.png
python scripts/univariate_vs_multivariate.py

# Q6: eight univariate models trained with our gradient descent
#   -> tables/univariate_gradient_descent.csv, figures/univariate_gradient_descent.png
python scripts/univariate_gradient_descent.py

# Q7: p-values across the three configurations
#   -> tables/pvalues.csv
python scripts/pvalues.py

# Part 3 support: conditioning, learning-rate sweep, GD vs. closed form
#   -> tables/convergence_diagnostics.csv, tables/learning_rate_sweep.csv,
#      figures/convergence_scaling.png, figures/convergence_log_transform.png
python scripts/convergence_diagnostics.py

# Part 3 support: closed-form coefficients for all three feature sets
#   -> tables/optimal_coefficients.csv
python scripts/optimal_coefficients.py

# Part 2 support: training-set statistics, client figures and tables
#   -> tables/descriptive_stats.csv, tables/model_reliability.csv,
#      tables/mix_adjustment_effects.csv, tables/water_cement_bands.csv, ...,
#      figures/client_*.png
python scripts/descriptive_stats.py
python scripts/client_insights.py

Reports (Parts 2 and 3)

The client report and the technical appendix are written in Markdown under reports/ and built into formatted Word documents (Calibri 11 pt, 1-inch margins, 1.15 spacing, numbered tables and figures, native Word equations):

conda install pandoc                  # or https://pandoc.org/installing.html
pip install -r reports/requirements.txt
python reports/build_reports.py       # -> reports/output/*.docx, prints word counts
python reports/build_reports.py --pdf # macOS + Microsoft Word: also export PDFs

The reports cite figures and tables from assets/, so run the analysis scripts first. Edit the .md files, not the generated .docx.

The two gradient descent scripts are self-contained: they depend only on the listed packages and can be run from any directory, or copied out of the repo alongside Concrete_Data.xls.

Exploratory interface

train.py and eval.py are a parameterized version of the same analysis, useful for testing other learning rates:

python -m scripts.train --transform standardized --lr 0.01 --iters 5000
python -m scripts.train --feature Cement --lr 1e-6 --iters 5000
python -m scripts.eval --transform log
python -m scripts.scaling_convergence

Hyperparameters

Each learning rate is the largest stable value for that feature scaling, found by testing powers of ten and confirmed against the theoretical limit 2/L from the Hessian's largest eigenvalue. All runs use 5000 iterations.

Model Learning rate Diverges above Condition number
Untransformed 3e-7 ~6.0e-7 1.2e10
Standardized 0.2 ~0.43 8.7e1
Log-transformed 5e-3 ~5.7e-3 1.6e7

Layout

data/                     UCI dataset and its original readme
scripts/                  one script per analysis (see above)
utils/                    grad_descent.py, metrics.py, data.py, transforms.py,
                          plotting.py, report_style.py (shared figure style)
models/                   linear.py (univariate + multivariate wrappers)
assets/
  CSE5104 ... .pdf        assignment
  figures/                generated figures (PNG, 300 dpi for the reports)
  tables/                 generated tables (CSV)
reports/
  part2_client_report.md       Part 2 source
  part3_technical_appendix.md  Part 3 source
  build_reports.py             Markdown -> Word (and PDF) build
  templates/reference.docx     generated Word style template
  output/                      built .docx / .pdf deliverables
run_all.sh                reproduce every figure and table

Dependencies

numpy, pandas, matplotlib, scikit-learn, statsmodels, xlrd

xlrd is required for pandas.read_excel() to open the .xls file. scikit-learn and statsmodels are used only for StandardScaler, the external optimal-solution baselines (LinearRegression), and p-values (statsmodels.OLS) — the gradient descent optimizer is implemented from scratch in scripts/gradient_descent_self_contained.py and utils/grad_descent.py.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages