How rahulrayy/cifar-double-descent works
This is a research project reproducing epoch-wise double descent on CIFAR-10 with a CIFAR-adapted ResNet-18. The main workflow injects training-label noise, trains across noise levels and seeds, evaluates test performance and memorization, and analyzes aggregated curves and peak statistics. The README also documents label-smoothing and mixup interventions. Some analysis and result-generation connections are documented but their implementation details are not shown in the excerpts, so those links are identified as documented rather than inferred from source.
Main entry points: train.py Training and evaluation · metrics.py Peak and seed metrics · data.py Noisy dataset · sweep.py Seed and noise sweep
Components 13
Data preparation
- CIFAR-10
core/data.py - Noisy dataset
core/data.py - Data loaders
core/data.py
Training experiments
- Seed and noise sweep
experiments/sweep.py - Training and evaluation
experiments/train.py - CIFAR ResNet-18
models/resnet.py - Smoothing and mixup
experiments/train.py
Analysis
- Plotting and analysis
notebooks/analysis.ipynb - Peak and seed metrics
core/metrics.py
Research outputs
- Run checkpoints and metrics
experiments/train.py - Peak summary table
results/summary_table.csv - Research figures
results/figures
Other components
- Researcher
Connections 12
- Researcher → Seed and noise sweep: starts sweep
README.md - Seed and noise sweep → Training and evaluation: launches runs
experiments/sweep.py - Training and evaluation → Data loaders: requests batches
experiments/train.py - Data loaders → Noisy dataset: creates datasets
core/data.py - Noisy dataset → CIFAR-10: downloads and wraps
core/data.py - Training and evaluation → CIFAR ResNet-18: constructs model
experiments/train.py - Training and evaluation → Smoothing and mixup: applies options
experiments/train.py - Training and evaluation → Run checkpoints and metrics: saves run state
experiments/train.py - Peak and seed metrics → Run checkpoints and metrics: loads metric arrays
core/metrics.py - Researcher → Plotting and analysis: reviews results
README.md - Plotting and analysis → Research figures: produces plots
README.md - Peak and seed metrics → Peak summary table: summarizes peaks
README.md