Experiment Tracking & Doc Writing Link to heading


Experiment Link to heading

  1. Use configs
  2. Save results
  3. Have one script to compare

1. Use configs: for all pipeline-level hyperparameters Link to heading

  • Base config: captures every optional feature in the pipeline, shared across experiments
  • Experiment config: modifies 1โ€“2 items from the base config (overrides)
  • Stored config: for each experiment run, make sure it can be traced back later

Recommend using *.yaml for better readability example

- name: '01.SOL'
  description: 'SOL'
  overrides: 
    symbol: "SOL"
    data_file_prefix: "20251104"

- name: '02.BTC'
  description: 'BTC'
  overrides: 
    symbol: "BTC"
    data_file_prefix: "20251104"

2. Save results: config, logs, prediction, metrics, (panels) Link to heading

results
โ”œโ”€โ”€ (20251105)
โ”‚   โ””โ”€โ”€ 20251105_030057.00.SOL
โ”œโ”€โ”€ 01.SOL
โ”‚   โ”œโ”€โ”€ all_feature_coefficients.csv
โ”‚   โ”œโ”€โ”€ all_feature_coefficients.parquet
โ”‚   โ”œโ”€โ”€ config.json
โ”‚   โ”œโ”€โ”€ config.yaml
โ”‚   โ”œโ”€โ”€ hyperparameter_history.json
โ”‚   โ”œโ”€โ”€ metrics.json
โ”‚   โ”œโ”€โ”€ pipeline_20251105_032010.log
โ”‚   โ””โ”€โ”€ predictions.parquet
โ””โ”€โ”€ 02.BTC
    โ”œโ”€โ”€ all_feature_coefficients.csv
    โ”œโ”€โ”€ all_feature_coefficients.parquet
    โ”œโ”€โ”€ config.json
    โ”œโ”€โ”€ config.yaml
    โ”œโ”€โ”€ hyperparameter_history.json
    โ”œโ”€โ”€ metrics.json
    โ”œโ”€โ”€ pipeline_20251105_035457.log
    โ””โ”€โ”€ predictions.parquet

3. Have one script to compare: *.ipynb Link to heading

  • set up: result dir, start_date, end_date (, baseline)

  • tell the difference in configs (experiment names)

  • show metrics of graphs & tables

  • export to *.html: ipynb + quarto

    quarto render "$notebook_path"
    
  • in distinct result dir, we have distinct groups for comparison

    • so the table & graphs only focus on current ablation analysis

Weekly Doc Writing Link to heading

use Typora to write *.md

  • toc: content table
  • slide content size 16:9 (for each page)
  • marp: export to ppt/html/pdf

Structure Link to heading

  • Finished jobs
    • track new dev & changes (complement of git)
  • Findings
    • any results of interest (whether structured or to-be-analysis)
    • doubts
  • Discussion
    • open questions (for GPT prompting and for team-members to know)
  • TODOs
    • always track & update that