Skip to content

[ML] DataScienceWithAgents hors 02-ML-Cours : converger les 17 carnets Python hors canon au CPython 3.13 #23802

Open

Description

@jsboige

Part of #23384.

Constat

Hors 02-ML-Cours (deja suivi par #23374), les sous-series de ML/DataScienceWithAgents portent encore 17 carnets sous kernelspec python3 dont l'estampille n'est pas le canon CPython 3.13. Mesure faite sur les blobs d'origin/main (ls-tree -r origin/main + git show origin/main:<path>), au commit e89519fe26, le 2026-10-31 — jamais sur l'arbre de travail d'un checkout partage.

Carnet Estampille Cellules Classe
03-Deep-Learning/3.2-Optimisateurs 3.11.9 11 local, CPU
05-Wavelet-Scattering/WS-00c-Scattering-from-scratch 3.11.9 12 local
05-Wavelet-Scattering/WS-01-Denoising-SOTA 3.11.17 14 local
05-Wavelet-Scattering/WS-03-Synthese-Scattering-vs-ResNet 3.12.8 9 local
Track2-GoogleADK/Day5-DS-Star/Lab12c-Agent-Handoff 3.11.9 8 local
Track2-GoogleADK/Day6-MLE-Star/Lab13-Web-Search-SOTA 3.12.13 12 local
Track2-GoogleADK/Day6-MLE-Star/Lab14-Ablation-Refinement 3.12.13 13 local
04-Vision/4.1-Conv-NumPy-Torch-Allclose 3.11.9 10 telechargement HTTP
03-Deep-Learning/3.9b-Compression-Pruning-from-scratch 3.10.22 23 CUDA
03-Deep-Learning/3.9e-Compression-SOTA 3.11.15 20 CUDA
04-Vision/4.2-ConvNet-Profonde-Residuelles 3.11.9 19 CUDA
04-Vision/4.3-TransferLearning-ResNet 3.11.16 10 CUDA + poids pre-entraines
05-Wavelet-Scattering/WS-02-Scattering-SOTA 3.11.9 12 CUDA
03-Deep-Learning/3.13b-Deux-Cartes-Reelles-DDP-FSDP-nccl 3.11.10 7 2 GPU + nccl
Track1-LangChain/.../Lab6-First-Agent 3.11.9 9 cle d'API
Track2-GoogleADK/Day5-DS-Star/Lab12b-Sequential-Orchestration 3.11.9 8 cle d'API
03-Deep-Learning/_archive/3.9f-Compression-Pruning-SOTA 3.11.9 13 archive — hors canon par convention

Sur les 116 carnets de la serie, 99 sont deja au canon ; ces 17-la ne le sont pas.

Pourquoi la classe compte plus que le total

Re-executer un carnet ne se justifie que si la re-execution ne degrade pas sa preuve (C.2, Stop & Repair). Trois barrieres reelles, nommees ici pour que la prochaine lane n'ait pas a les re-deduire :

  1. CUDA (6 carnets) — la montee de noyau change la pile torch, et les noyaux CUDA ne sont pas reproductibles d'une version a l'autre : une re-execution peut deplacer des chiffres sans qu'aucune cause de code ne soit en jeu. Ce n'est pas une interdiction, c'est un diagnostic a poser (C.4, cause (d)/(e)) avant de committer.
  2. Cle d'API (2 carnets) — une re-execution sans cle produit Execution sautee (API non configuree) et remplace une sortie substantielle par un repli : c'est le contournement de C.2 par la porte de secours, et le ratchet Output-collapse le signale.
  3. Telechargement HTTP (1 carnet) — 4.1 recupere une image par urllib.request : la sortie depend du contenu servi ce jour-la, il lui faut un instantane fige ou une justification ecrite.

Les 7 carnets locaux (dont 3.2-Optimisateurs) sont le pool directement servable.

Attendu

Une PR par carnet (ou par paire coherente), kernel python3 au canon CPython 3.13.15, re-executee par papermill, commitee avec ses sorties (C.2), et un diagnostic C.4 ecrit dans le body quand la re-execution deplace une valeur.

scripts/notebook_tools/check_kernel_drift.py origin/main doit rendre findings: [] sur le diff.

Tranches

  • tranche 1 — 03-Deep-Learning/3.2-Optimisateurs (local, CPU, entierement seede)
  • tranches suivantes — les 6 autres locaux, puis les barrieres une par une avec leur diagnostic

Voir #23384 (census des carnets Python hors canon).

Activity

  1. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/03-Deep-Learning/3.2-Optimisateurs.ipynb -- tranche 1 : convergence kernel python3 vers CPython 3.13.15, re-execution papermill avec sorties

  2. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- PR #23804 -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/03-Deep-Learning/3.2-Optimisateurs.ipynb

    Tranche 1 de la famille 03-Deep-Learning : le carnet passe au noyau 3.13.15, avec deux correctifs de source que la re-execution a fait apparaitre (garde de debordement sur la norme, qui supprime un RuntimeWarning porteur d'un chemin machine dans la sortie committee ; identifiant de cellule invalide au regard du motif nbformat).

    Sorties : 11/11 cellules executees, 0 erreur, aucun texte de sortie modifie hors la cellule dont la source change. Cinq gardes locales vertes.

    Les tranches suivantes de cette famille (WS-00c, WS-01, WS-03, Lab12c, Lab13, Lab14, puis 4.1 avec son probleme d'instantane HTTP) restent ouvertes et ne sont pas couvertes par ce claim.

  3. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-00c-Scattering-from-scratch.ipynb

    Tranche 2 de la famille : re-execution de WS-00c-Scattering-from-scratch vers le canon CPython 3.13.15. La tranche 1 (3.2-Optimisateurs, livree PR #23804) est rendue par son [DELIVERED].

    Note d'env pour les tranches WS suivantes : kymatio (toutes versions disponibles, master compris) importe sph_harm au chargement via son frontend 3D, symbole retire de scipy >= 1.15 ; le canon noyau porte scipy 1.18.1. Reparation locale en cours (alias documente vers sph_harm_y), les carnets WS n'exercent jamais le chemin 3D.

  4. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- PR #23812 -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-00c-Scattering-from-scratch.ipynb

    Tranche 2 : WS-00c au canon CPython 3.13.15, 12/12 cellules executees, 0 erreur, 0 source modifiee. Deux cellules de sortie triees une par une (verdict allclose inchange, classement scattering > pixels inchange, aucun chiffre epingle en prose).

    Reparation d'env livree au passage pour toute la serie WS : kymatio (toutes versions disponibles, master compris) importe scipy.special.sph_harm (retire en 1.15) au chargement de son frontend 3D ; alias documente vers sph_harm_y pose dans le user-site du noyau. Detail dans le body de la PR.

    Tranches restantes : WS-01, WS-03, Lab12c/13/14, 4.1 (instantane HTTP), puis barrieres CUDA/API avec diagnostic.

  5. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-01-Denoising-SOTA.ipynb

    Tranche 3 de la famille : re-execution de WS-01-Denoising-SOTA vers le canon CPython 3.13.15. Les tranches 1-2 sont livrees (#23804, #23812).

    Env : PyWavelets + scikit-image installes dans le noyau (regle F) ; pywt roundtrip et denoise_wavelet smoke-testes.

  6. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- PR #23814 -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-01-Denoising-SOTA.ipynb

    Tranche 3 : WS-01 au canon CPython 3.13.15, 14/14 cellules executees, 0 erreur, 0 source modifiee. La seule sortie qui bouge est la banniere de versions (numpy 2.4.6, matplotlib 3.11.1, scikit-image 0.26.0, pywt 1.8.0 inchange) ; toutes les cellules numeriques (PSNR des six methodes, sweep, figures) sont byte-identiques.

    Env : PyWavelets + scikit-image installes dans le noyau canon (regle F), smoke-testes avant execution.

    Tranches restantes : WS-03, Lab12c/13/14, 4.1 (instantane HTTP), barrieres CUDA/API.

  7. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-03-Synthese-Scattering-vs-ResNet.ipynb

    Tranche 4 de la famille : re-execution de WS-03 vers le canon CPython 3.13.15. Tranches 1-3 livrees (#23804, #23812, #23814).

  8. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- PR #23822 -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/05-Wavelet-Scattering/WS-03-Synthese-Scattering-vs-ResNet.ipynb

  9. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day5-DS-Star/Lab12c-Agent-Handoff.ipynb

  10. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- PR #23839 -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day5-DS-Star/Lab12c-Agent-Handoff.ipynb

  11. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day5-DS-Star/Lab12b-Sequential-Orchestration.ipynb

  12. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day5-DS-Star/Lab12b-Sequential-Orchestration.ipynb -- Lab12b 3.11.9->3.13.15 par PR #23844

  13. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day6-MLE-Star/Lab14-Ablation-Refinement.ipynb

  14. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [DELIVERED] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day6-MLE-Star/Lab14-Ablation-Refinement.ipynb -- Lab14 3.12.13->3.13.15, modele preserve, par PR #23848

  15. jsboige commented on Oct 31, 2026

    @jsboige
    OwnerAuthor

    [CLAIMED-AMEND] lane myia-po-2027:CoursIA -- paths: MyIA.AI.Notebooks/ML/DataScienceWithAgents/Track2-GoogleADK/Day6-MLE-Star/Lab13-Web-Search-SOTA.ipynb -- tranche 8 : Lab13 3.12.13->3.13.15 (arXiv vivant + LLM openai/gpt-5.6-sol)

  16. 24 remaining items

  17. added a commit that references this issue on Nov 2, 2026
  18. added 2 commits that reference this issue on Nov 2, 2026
  19. added 6 commits that reference this issue on Nov 2, 2026
  20. added 2 commits that reference this issue on Nov 2, 2026
  21. added 4 commits that reference this issue on Nov 2, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions