Skip to content

Release 0.5.0: saved models, auto plan, no pytorch-tabular, LightGBM, new README - #3

Merged
mdicio merged 15 commits into
devfrom
release-0.5.0
Nov 13, 2026
Merged

mdicio merged 15 commits into
devfrom
release-0.5.0

Conversation

@mdicio

@mdicio mdicio commented Nov 13, 2026

Copy link
Copy Markdown
Owner

Brings dev up to 0.5.0. Everything after PR #2 (plug-and-play), as one linear history:

  • 0.4.0: saved, ready-to-use best models (load_model, autodeep predict); models: auto with a time budget and machine profile; the nine deep tabular architectures ported off pytorch-tabular (pandas 3 supported, no pinned-back dependencies); TabICL full-context test predictions; CPU-cost warnings; competitive XGBoost/CatBoost grids; GPU fixes; benchmark results; TPE after a third of the trials.
  • 0.5.0: LightGBM (also in the auto plan), offset columns (GLM-boost), residual MLP, search_folds, autodeep.ensemble.Stacker.
  • README: demo clip of the web app, new screenshots, a minimal Python quick start, original description/features text restored.

Checked locally on the tip: 127 fast tests and the slow all-models smoke test pass (CPU); lint passes; fresh clone + uv sync + autodeep quickstart works in ~30 s; the README's Python API and autodeep predict examples run.

Full details in CHANGELOG.md.

🤖 Generated with Claude Code

mdicio and others added 15 commits November 12, 2026 01:34
…deep predict

- TabularPreprocessor: the data loader's typing/imputation/one-hot/scaling as one fitted
  object that replays on raw rows (output identical to the previous loader on 24 cases)
- Predictor (preprocessing + model + original labels): predict, predict_proba,
  predict_frame, evaluate, save/load as a folder (predictor.pt + info.json)
- AutoRunner keeps the best model per dataset by validation score, saves it to
  models/<dataset>/best (save_best=True); save_models=True saves every model;
  runner.best_model(); leaderboard 'selected' column
- All 17 models pickle: PyTorch Tabular without training data / Lightning trainer,
  torch nets as build args + weights, TabICL/pytabkit/XGBoost device moves, so
  GPU-trained models load on CPU-only machines
- CLI: autodeep predict, autodeep info; web UI: download model, predict a file

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
- autoplan: picks models (CPU/GPU priority order), max_evals and a water-filled
  per-model time limit that fit time_budget (default 15m), with reasons for every
  skipped model; no model may need more than half the budget; 8% slack for overrun
- machine: cores, RAM, GPU and a ~0.5 s CPU benchmark (warm-up for powersave
  governors, cached 30 days) that scales the cost-curve priors until history exists
- AutoRunner: models=None/'auto' is the default, time_budget, plan(); explicit
  max_evals/timeout_per_model win; a budget with explicit models sets the timeout
- CLI: run/estimate accept a data file + --target, --budget, autodeep system
- Web UI: Auto mode with budget and plan explanation, /api/plan
- sklearn MLP prior re-measured (was ~15x optimistic above 10k rows)
- README/CHANGELOG/examples; version 0.4.0

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
pytorch-tabular 1.2 is the only dependency pinning pandas<3. Its nine
architectures (CategoryEmbedding, TabNet, FT-Transformer, TabTransformer,
AutoInt, GANDALF, GATE, NODE, DANet) are now plain nn.Modules in
autodeep/modelsdefinition/tabular_nets/ (adapted from pytorch-tabular
1.2.0, MIT, licence kept) and are trained by AutoDeep's own TorchTrainer
loop through EmbeddingTorchTrainer. TabNet uses pytorch-tabnet's network
directly. Hyperparameter names and defaults are unchanged.

- Drop pytorch-tabular, requests, the pandas<3 and pytorch-tabnet<4.2
  caps, the Lightning callback, the .pt_tmp monkeypatch and the pickling
  workarounds. uv.lock now resolves pandas 3.1.1 (2.3.3 on Python 3.10).
- No einops or omegaconf; torch.jit.script (deprecated in torch 2.15)
  dropped from the entmoid helpers.
- TabTransformer's out_ff_* options and transformer_head_dim, silently
  ignored by pytorch-tabular 1.x, now take effect.
- pandas 3 copy-on-write: copy arrays before wrapping them as tensors.
- TorchTrainer seeds random/numpy too, has an init_network hook for
  data-aware initialisation and falls back to default_params.batch_size.

Checked: fast suite (116) and the slow all-models smoke test pass on
pandas 3.1.1. Against the old pytorch-tabular trainers, same seeds and
trials, diabetes test RMSE is within a few percent for every model
except TabTransformer (whose head is now configured as intended).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
Checked on the RTX 3060 (all 17 models x 3 problem types, fast suite with
AUTODEEP_DEVICE=cuda, and GPU-trained models loaded without a GPU).

- DANet failed on CUDA ("view size is not compatible with input
  tensor's size and stride") when the locality mask output is not
  contiguous. Use reshape there and in the two NODE views on computed
  tensors. The same code is in pytorch-tabular.
- An unpickled TorchTrainer rebuilt its network on CPU while its device
  property pointed at the GPU, so predict() failed outside
  Predictor.load(). The network now goes to the trainer's device (CPU
  if that fails).
- test_predictor: a GPU-trained XGBoost regressor predicts on CPU up to
  float32 rounding; compare regression outputs with a tolerance.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
- New trainer flags: refit_on_full_train (refit the best configuration
  on train + validation rows after the search) and compute_train_metrics.
  TabICL uses both: "fitting" only stores the context, so the test
  predictions get the whole training set for free, and its train metrics
  (context rows scored against themselves) are no longer reported.
- SearchResult carries the raw best params for the refit.
- The estimate warns when a GPU-hungry model (TabICL, TabM, 1D-CNN,
  ResNet) would take more than 5 minutes on CPU. Job warnings are now
  printed by `autodeep estimate` and logged at run start, not only shown
  in the web app.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…ML fix

- XGBoost searched 50-200 trees with learning rates down to 0.01 and
  early_stopping_rounds 100, so low-rate trials were badly underfit and
  early stopping never triggered; CatBoost only searched iterations.
  Both now cap at 2000 trees with early stopping (50 / 100 rounds) and
  search learning rate, depth and regularisation. Also removes a
  duplicated early_stopping_rounds key.
- AutoRunner.run() pins the resolved device for the run. With "auto",
  every trainer.device lookup ran nvidia-smi again (once per prediction
  batch for torch models), and a run could switch from GPU to CPU
  halfway once its own allocations shrank the free memory.
- Unpickled torch networks stay on CPU and predict_with() moves them to
  the trainer's device (replaces the move in __setstate__, which put the
  weights on the GPU even for load_model(device="cpu")).
- openml: datasets failed on pandas 3 + scikit-learn 1.9 in the pandas
  ARFF parser; retry with liac-arff.
- Silence XGBoost's per-call "mismatched devices" warning for GPU
  boosters predicting from pandas data.
- The ported architectures log ignored hyperparameters at debug level.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
Mechanical `ruff --select UP --target-version py310 --fix` over autodeep/
and tests/: Dict/List/Optional/Union -> dict/list/X | None, isinstance
tuples -> unions, quoted annotations. The pre-commit ruff hook now
selects UP as well. No behaviour change; fast suite passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…oxes

- hyperopt 0.3.0 no longer imports pkg_resources: require it and drop
  setuptools<81. The lock now has setuptools 84.2.1; fast suite and the
  all-models smoke test pass.
- Web UI: on model cards with long descriptions the flex layout shrank
  the checkbox to a dot (seen in a headless-Chrome screenshot).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…nshots

- benchmarks/README.md: larger_datasets.py on california_housing,
  adult and covertype_50k (5 trials, one split). TabICLv2 is best on
  two of three without tuning; CatBoost/XGBoost win on adult. Caveats
  and run times included.
- Estimator: GPU factors measured on an RTX 3060 (TabICLv2 54-78x ->
  50, TabM median 13x -> 12, RealMLP ~1-2x -> 1.5, was 5) and XGBoost /
  CatBoost CPU priors re-fitted for the new 2000-tree grids (~10x / ~5x).
- README: screenshots of the web app (finished run, new experiment),
  taken with headless Chrome on a neutral /tmp workspace.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…ries take effect

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…h diagnostics

hyperopt starts TPE after 20 random trials by default, so a 25-trial search
was mostly random search. The start-up phase is now a third of the budget
(5 to 10 trials).

benchmarks/larger_datasets.py gets a core and a deep suite, runs every
(dataset, model) pair separately so an interrupted run resumes, and stops
before a model when the GPU has less than 4 GB free.
benchmarks/search_diagnostics.py checks the searches of a run: TPE gain over
the start-up trials, trial of the best score, hyperparameters at the edge of
their range.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC637XcrFPHFykZKx7pbm
…stacker

- lgbm: LightGBM trainer (CPU, native categoricals), default grids, fast preset
- default_params.offset_column: boosting starts from another model's log-odds
  (base_margin / init_score); the column is not a feature
- autodeep.ensemble.Stacker: hill climbing / L2 logistic regression on
  logit-ranks of OOF predictions, scored by nested CV

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC638tC97Wg2RMQkjZK56
- resmlp: PyTorch MLP; with offset_column it starts at the offset model
  (zero-initialised head) and learns the residual
- default_params.search_folds: score each hyperopt trial on K stratified folds

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013EC638tC97Wg2RMQkjZK56
…BM in the auto plan

- README: demo GIF/MP4 of the web app, new screenshots (setup, live run,
  results), a minimal Python quick start, and the original Description /
  Features / How to Use / View Results / Acknowledgments text brought back
  and updated. Dropped the obsolete setuptools troubleshooting row and the
  mock-up results table image.
- models: auto now considers LightGBM (after CatBoost, min 10 trials).
- Version 0.5.0; CHANGELOG entries for 0.4.0 and 0.5.0 dated.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mdicio
mdicio merged commit 20cb605 into dev Nov 13, 2026
3 checks passed
@mdicio
mdicio deleted the release-0.5.0 branch November 13, 2026 14:10
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant