70 lines (49 loc) · 2.22 KB

Experiments Guide

The notebooks under experiments/ are organized around the outputs already generated by training and benchmarking.

Notebook Inventory

`experiments/comparison/`

span-vs-fsrcnn.ipynb
selective-vs-full-processing.ipynb
fp16-vs-int8.ipynb

`experiments/frame-selection/`

sparse-sampling-analysis.ipynb
scene-change-threshold.ipynb
blur-detection-comparison.ipynb

`experiments/quality-analysis/`

ocr-accuracy-analysis.ipynb
psnr-ssim-correlation.ipynb
subjective-quality-study.ipynb

`experiments/optimization/`

quantization-experiments.ipynb
gpu-optimization.ipynb
batch-size-tuning.ipynb

Typical input data

These notebooks usually read from benchmarks/results/ and outputs/:

Input	Produced by
`quality-metrics.csv`	`benchmarks/scripts/benchmark_quality.py`
`processing-times.csv`	`benchmarks/scripts/benchmark_processing_time.py`
OCR comparison CSV/JSON files	OCR-aware evaluation and benchmark runs
training logs	`models.training.finetune_education` outputs
exported images or videos	desktop runner or visual comparison tools

Recommended workflow

Run training or evaluation so you have fresh CSVs and summaries.
Run benchmark wrappers if you need desktop timing or video metrics.
Start Jupyter from the repo root or experiments/.
Update notebook paths only if you intentionally changed output locations.

Good pairings

Notebook area	Best supporting workflow
comparison	benchmarks.md and multiple model checkpoints
frame-selection	`scripts/pipeline.py` plus lecture-video extracts
quality-analysis	`models.training.evaluation` and OCR summary files
optimization	`tools/benchmark-runner/runner.py` and TFLite conversions

Practical tips

Keep generated CSVs in benchmarks/results/ when possible; many notebooks assume that layout.
Export final thesis figures to thesis/ or a dedicated output directory instead of saving only inside notebook cells.
Clear notebook outputs before committing if the files become noisy or oversized.

Related Docs