Context du cours et crise de reproductibilité en apprentissage profond

Alors que nous passons des modèles simples et autonomes aux architectures complexes à plusieurs étapes nécessaires pour le Projet d'objectif 1, suivre manuellement les paramètres critiques dans des feuilles de calcul ou des fichiers locaux devient totalement insoutenable. Ce flux de travail complexe introduit des risques graves pour l'intégrité du développement.

1. Identifier le goulot d'étranglement de la reproductibilité

Le flux de travail de l'apprentissage profond implique nécessairement une forte variabilité en raison de nombreux facteurs (algorithmes d'optimisation, sous-ensembles de données, techniques de régularisation, différences d'environnement). Sans suivi systématique, reproduire un résultat spécifique — essentiel pour le débogage ou l'amélioration d'un modèle déployé — est souvent impossible.

Que faut-il suivre ?

Hyperparamètres : All configuration settings must be recorded (e.g., Learning Rate, Batch Size, Optimizer choice, Activation function).

État de l'environnement : Software dependencies, hardware used (GPU type, OS), and exact package versions must be fixed and recorded.

Objets et résultats : Pointers to the saved model weights, final metrics (Loss, Accuracy, F1 score), and training runtime must be stored.

The "Single Source of Truth" (SSOT)

Systematic experiment tracking establishes a central repository—a SSOT—where every choice made during model training is recorded automatically. This eliminates guesswork and ensures reliable auditability across all experimental runs.

TERMINALbash — tracking-env

> Ready. Click "Run Conceptual Trace" to see the workflow.

EXPERIMENT TRACE Live

Simulate the run to visualize the trace data captured.

Question 1

What is the root cause of the Deep Learning Reproducibility Crisis?

PyTorch's dependence on CUDA drivers.

The sheer number of untracked variables (code, data, hyperparameter, and environment).

The excessive memory usage of large models.

The computational cost of generating artifacts.

Question 2

In the context of MLOps, why is systematic experiment tracking essential for production?

It minimizes the total storage size of model artifacts.

It ensures that the model achieving the reported performance can be reliably reconstructed and deployed.

It speeds up the training phase of the model.

Question 3

Which element is necessary to reproduce a result but is most often forgotten in manual tracking?

The number of epochs run.

The specific versions of all Python libraries and the random seed used.

The name of the dataset used.

The time the training started.

Challenge: Tracking in Transition

Why the transition to formal tracking is non-negotiable.

You are managing 5 developers working on Milestone Project 1. Each developer reports their best model accuracy (88% to 91%) in Slack. No one can reliably tell you the exact combination of parameters or code used for the winning run.

Step 1

What immediate step must be implemented to halt the loss of critical information?

Solution:
Implement a mandatory requirement for every run to be registered with an automated tracking system before results are shared, capturing the full hyperparameter dictionary and Git hash.

Step 2

What benefit does structured tracking provide to the team that a shared spreadsheet cannot?

Solution:
Structured tracking allows automated comparison dashboards, visualizations of parameter importance, and centralized artifact storage, which is impossible with static spreadsheets.