v0.7

Key Highlights

• Pretrained Vision Models: we’ve added 20 additional TorchVision pretrained models as image encoders, including: AlexNet, EfficientNet, MobileNet v3, and GoogleLeNet.
• Image Augmentation: Ludwig v0.7 also introduces image augmentation, artificially increasing the size of the training dataset by applying a randomized set of transformations to each batch of images during training.
• 50x Faster Fine-Tuning via Automatic Mixed Precision (AMP) Training, Cached Encoder Embeddings, Approximate Training Set evaluation, and automatic batch sizing by default to maximize throughput.
• New Distributed Training Strategies: Distributed Data Parallel (DDP) and Fully Sharded Data Parallel (FSDP)
• Ray 2.0, 2.1, 2.2 and 2.3 support
• A new Ludwig profiler for benchmarking various CPU/GPU performance metrics, as well as comparing different Ludwig model runs.
• Revamped Ludwig datasets API with an even larger number of datasets out of the box.
• API annotations within Ludwig for contributors and Python users
• Schemification of the entire Ludwig Config object for better validation and checks upfront.

What's Changed

• Fix ray nightly import by @jppgks in #2196
• Restructured split config and added datetime splitting by @tgaddair in #2132
• enh: Implements InferenceModule as a pipelined module with separate preprocessor, predictor, and postprocessor modules by @brightsparc in #2105
• Explicitly pass data credentials when reading binary files from a RayBackend by @jeffreyftang in #2198
• MlflowCallback: do not end run on_trainer_train_teardown by @jppgks in #2201
• Fail hyperopt with full import error when Ray not installed by @tgaddair in #2203
• Make convert_predictions() backend-aware by @hungcs in #2200
• feat: MVP for explanations using Integrated Gradients from captum by @jppgks in #2205
• [Torchscript] Adds GPU-enabled input types for Vector and Timeseries by @geoffreyangus in #2197
• feat: Added model type GBM (LightGBM tree learner), as an alternative to ECD by @jppgks in #2027
• [Torchscript] Parallelized Text/Sequence Preprocessing by @geoffreyangus in #2206
• feat: Adding feature type shared parameter capability for hyperopt by @arnavgarg1 in #2133
• Bump up version to 0.6.dev. by @justinxzhao in #2209
• Define FloatOrAuto and IntegerOrAuto schema fields, and use them. by @justinxzhao in #2219
• Define a dataclass for parameter metadata. by @justinxzhao in #2218
• Add explicit handling for zero-length image byte buffers to avoid cryptic errors by @jeffreyftang in #2210
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2231
• Create dataset util to form repeatable train/vali/test split by @amholler in #2159
• Bug fix: Use safe rename which works across filesystems when writing checkpoints by @dantreiman in #2225
• Add parameter metadata to the trainer schema. by @justinxzhao in #2224
• Add an explicit call to merge_wtih_defaults() when loading a config from a model directory. by @justinxzhao in #2226
• Fixes flaky test test_datetime_split[dask] by @dantreiman in #2232
• Fixes prediction saving for models with Set output by @geoffreyangus in #2211
• Make ExpectedImpact JSON serializable by @hungcs in #2233
• standardised quotation marks, added missing word by @Marvjowa in #2236
• Add boolean postprocessing to dataset type inference for automl by @magdyksaleh in #2193
• Update get_repeatable_train_val_test_split to handle non-stratified split w/ no existing split by @amholler in #2237
• Update R2 score to handle single sample computation by @arnavgarg1 in #2235
• Input/Output Feature Schema Refactor by @connor-mccorm in #2147
• Fix nan in entmax loss and flaky sparsemax/entmax loss tests by @dantreiman in #2238
• Fix preprocessing dataset split API backwards compatibility upgrade bug. by @justinxzhao in #2239
• Removing duplicates in constants from recent PRs by @arnavgarg1 in #2240
• Add attention scores of the vit encoder as an additional return value by @Dennis-Rall in #2192
• Unnest Audio Feature Preprocessing Config by @connor-mccorm in #2242
• Fixed handling of invalud number values to treat as missing values by @tgaddair in #2247
• Support saving numpy predictions to remote FS by @hungcs in #2245
• Use global constant for description.json by @hungcs in #2246
• Removed import warnings when LightGBM and Ray not requested by @tgaddair in #2249
• Adds ability to read images from numpy files and numpy arrays by @geoffreyangus in #2212
• Hyperopt steps per epoch not being computed correctly by @arnavgarg1 in #2175
• Fixed splitting when providing pre-split inputs by @tgaddair in #2248
• Added Backwards Compatibility for Audio Feature Preprocessing by @connor-mccorm in #2254
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2256
• Fix: Don't skip saving the model if the save path already exists. by @justinxzhao in #2264
• Load best weights outside of finally block, since load may throw an exception by @dantreiman in #2268
• Reduce number of distributed tests. by @justinxzhao in #2270
• [WIP] Adds inference_utils.py by @geoffreyangus in #2213
• Run github checks for pushes and merges to *-stable. by @justinxzhao in #2266
• Add ludwig logo and version to CLI help text. by @justinxzhao in #2258
• Add hyperopt_statistics.json constant by @hungcs in #2276
• fix: Make BaseTrainerConfig an abstract class by @ksbrar in #2273
• [Torchscript] Adds --device argument to export_torchscript CLI command by @geoffreyangus in #2275
• Use pytest tmpdir fixture wherever temporary directories are used in tests. by @justinxzhao in #2274
• adding configs used in benchmarking by @abidwael in #2263
• Fixes #2279 by @noahlh in #2284
• adding hardware usage and software packages tracker by @abidwael in #2195
• benchmarking utils by @abidwael in #2260
• dataclasses for summarizing benchmarking results by @abidwael in #2261
• Benchmarking core by @abidwael in #2262
• Fixed default eval_batch_size when setting batch_size=auto by @tgaddair in #2286
• Remove obsolete postprocess_inference_graph function. by @justinxzhao in #2267
• [Torchscript] Adds BERT tokenizer + partial HF tokenizer support by @geoffreyangus in #2272
• Support passing ground_truth as df for visualizations by @hungcs in #2281
• catching urllib3 exception by @abidwael in #2294
• Run pytest workflow on release branches. by @justinxzhao in #2291
• Save checkpoint if train_steps is smaller than batcher's steps_per_epoch by @dantreiman in #2298
• Fix typo in amazon review datasets: s/review_tile/review_title by @dantreiman in #2300
• Refactor non-distributed automl utils into a separate directory. by @justinxzhao in #2296
• Don't skip normalization in TabNet during inference on a single row. by @dantreiman in #2299
• Fix error in postproc_predictions calculation in model.evaluate() by @arnavgarg1 in #2304
• Test for parameter updates in Ludwig components by @jimthompson5802 in #2194
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2311
• Use warnings to suppress repeated logs for failed image reads by @arnavgarg1 in #2312
• Use ray dataset and drop type casting in binary_feature prediction post processing for speedup by @magdyksaleh in #2293
• Add size_bytes to DatasetInfo and DataSource by @jeffreyftang in #2306
• Fixes TensorDtype TypeError in Ray nightly by @geoffreyangus in #2320
• Add configuration section for global feature parameters by @arnavgarg1 in #2208
• Ensures unit tests are deleting artifacts during teardown by @geoffreyangus in #2310
• Fixes unit test that had empty Dask partitions after splitting by @geoffreyangus in #2313
• Serve json numpy encoding by @jeffkinnison in #2316
• fix: Mlflow config being injected in hyperopt config by @hungcs in #2321
• Update tests that use preprocessing to match new defaults config structure by @arnavgarg1 in #2323
• Bump test timeout to 60 minutes by @tgaddair in #2325
• Set a default value for size_bytes in DatasetInfo by @jeffreyftang in #2331
• Pin nightly versions to fix CI by @geoffreyangus in #2327
• Log number of failed image reads by @arnavgarg1 in #2317
• Add test with encoder dependencies for global defaults by @arnavgarg1 in #2342
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2334
• Add wine quality notebook to demonstrate using config defaults by @arnavgarg1 in #2333
• fix: GBM tests failing after new release from upstream dependency by @jppgks in #2347
• fix: restore overwrite of eval_batch_size on GBM schema by @jppgks in #2345
• Removes empty partitions after dropping rows and splitting datasets by @geoffreyangus in #2328
• fix: Properly serialize ParameterMetadata to JSON by @ksbrar in #2348
• Test for parameter updates in Ludwig Components - Part 2 by @jimthompson5802 in #2252
• refactor: Replace bespoke marshmallow fields that accept multiple types with a new 'combinatorial' OneOfField that accepts other fields as arguments. by @ksbrar in #2285
• Use Ray Datasets to read binary files in parallel by @tgaddair in #2241
• typos: Update README.md by @andife in #2358
• Respect the resource requests in RayPredictor by @magdyksaleh in #2359
• Resource tracker threading by @abidwael in #2352
• Allow writing init_config results to remote filesystems by @tgaddair in #2364
• Fixed export_mlflow command to not assume an existing registered_model_name by @tgaddair in #2369
• fix: Fixes to serialization, and update to allow set repo location. by @brightsparc in #2367
• Add amazon employee access challenge kaggle dataset by @justinxzhao in #2349
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2362
• Wrap read of cached training set metadata in try/except for robustness by @jeffreyftang in #2373
• Reduce dropout prob in test_conv1d_stack by @dantreiman in #2380
• fever: change broken download links by @jppgks in #2381
• Add default split config by @hungcs in #2379
• Fix CI: Skip failing ray GBM tests by @justinxzhao in #2391
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2389
• Triton ensemble export by @abidwael in #2251
• Fix: Random dataset splitting with 0.0 probability for optional validation or test sets. by @justinxzhao in #2382
• Print final training report as tabulated text. by @justinxzhao in #2383
• Add Ray 2.0 to CI by @tgaddair in #2337
• add GBM configs to benchmarking by @jppgks in #2395
• Optional artifact logging for MLFlow by @ShreyaR in #2255
• Simplify ludwig.benchmarking.benchmark API and add ludwig benchmark CLI by @abidwael in #2394
• rename kaggle_api_key to kaggle_key by @jppgks in #2384
• use new URL for yosemite dataset by @jppgks in #2385
• Encoder refactor V2 by @dantreiman in #2370
• re-enable GBM tests after new lightgbm-ray release by @jppgks in #2393
• Added option to log artifact location while creating mlflow experiment by @ShreyaR in #2397
• Treat dataset columns as object dtype during first pass of handle_missing_values by @jeffreyftang in #2398
• fix: ParameterMetadata JSON serialization bug by @ksbrar in #2399
• Adds registry to organize backward compatibility updates around versions and config sections by @dantreiman in #2335
• Include split column in explanation df by @connor-mccorm in #2405
• Fix AimCallback to model_name as Run.name by @alberttorosyan in #2413
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2410
• Hotfix: features eligible for shared params hyperopt by @arnavgarg1 in #2417
• Nest FC Params in Decoder by @connor-mccorm in #2400
• Hyperopt Backwards Compatibility by @connor-mccorm in #2419
• Investigating test_resnet_block_layer intermittent test failure by @dantreiman in #2414
• fix: Remove duplicate option from cell_type field schema by @ksbrar in #2428
• Test for parameter updates in Ludwig Combiners - Part 3 by @jimthompson5802 in #2332
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2430
• Hotfix: Proc column missing in output feature schema by @arnavgarg1 in #2435
• Nest hyperopt parameters into decoder object by @arnavgarg1 in #2436
• Fix: Make the twitter bots modeling example runnable by @justinxzhao in #2433
• Add MLG-ULB creditcard fraud dataset by @jppgks in #2425
• Bugfix: non-number inputs to GBM by @jppgks in #2418
• GBM: log intermediate progress by @jppgks in #2421
• Fix: Upgrade ludwig config before schema validation by @connor-mccorm in #2441
• Log warning for calibration if validation set is trivially small by @dantreiman in #2440
• Fixes calibration and adds example scripts by @dantreiman in #2431
• Add medical no-show appointments dataset by @jppgks in #2387
• Added conditional check for UNK token insertion into category feature vocab by @arnavgarg1 in #2429
• Ensure synthetic dataset unit tests to clean up extra files. by @justinxzhao in #2442
• Added feature specific parameter test for hyperopt by @arnavgarg1 in #2329
• Fixed version transformation to accept user configs without ludwig_version by @tgaddair in #2424
• Fix mulitple partition predict by @magdyksaleh in #2422
• Cache jsonschema validator to reduce memory pressure by @tgaddair in #2444
• [tests] Added more explicit lifecycle management to Ray clusters during tests by @tgaddair in #2447
• Fix: explicit keyword args for seaborn plot fn by @jppgks in #2454
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2453
• Extended hyperopt to support nested configuration block parameters by @tgaddair in #2445
• Consolidate missing value strategy to only include bfill and ffill by @arnavgarg1 in #2457
• fix: Switched Learning Rate to NonNegativeFloat Field by @connor-mccorm in #2446
• Support GitHub Codespaces by @jppgks in #2463
• Enh: quality-of-life improvements for export_torchscript by @geoffreyangus in #2459
• Disables batch_size: auto for CPU-only training by @geoffreyangus in #2455
• buxfix: triton model version as a string by @abidwael in #2461
• Updating images to Ray 2.0.0 and CUDA 11.3 by @abidwael in #2390
• Loss, Split, and Defaults Schema Additions by @connor-mccorm in #2439
• More precise resource usage tracking by @abidwael in #2363
• Summarizing performance metrics and resource usage results by @abidwael in #2372
• Better gbm defaults based on benchmarking results by @jppgks in #2466
• Infer single distinct value columns as category instead of binary by @arnavgarg1 in #2467
• fix: Add explicit schema in to_parquet() during saving predictions by @hungcs in #2420
• Publish docker images from release branches by @tgaddair in #2470
• Add backwards-compatibility logic for model progress tracker by @jeffreyftang in #2468
• Backwards compatibility for class_weights by @connor-mccorm in #2469
• Test for parameter updates in Ludwig Decoders - Part 4 by @jimthompson5802 in #2354
• Fixed backwards compatibility for training_set_metadata and bfill by @tgaddair in #2472
• Fixed backwards compatibility for models with level metadata in saved configs by @tgaddair in #2475
• Fix profiler: account for missing values when running in docker by @jppgks in #2477
• Add L-BFGS optimizer by @jppgks in #2478
• fix: Automatically assign title to OneOfOptionsField by @ksbrar in #2480
• fix: handle 'numerical' entries in preprocessing config during backwards compatibility upgrade by @jeffreyftang in #2484
• fix: mark update_class_weights_in_features transformation for version 0.6 by @jeffreyftang in #2481
• Fixed usage of checkpoints for AutoML in Ray 2.0 by @tgaddair in #2485
• [fix flaky test] Relax loss constraint for unit tests for lbfgs optimizer. by @justinxzhao in #2486
• Fixed stratified splitting with Dask by @tgaddair in #1883
• Replace custom Union marshmallow fields with Oneof fields, and default allow_none=True everywhere. by @justinxzhao in #2482
• Resource isolation for dataset preprocessing on ray backends by @magdyksaleh in #2404
• Pin transformers < 4.22 until issues resolved by @tgaddair in #2495
• Fix flaky ray nightly image test by @arnavgarg1 in #2493
• Added workflow to auto cherry-pick into release branches by @tgaddair in #2500
• Enable hyperopt to be launched from a ray client by @ShreyaR in #2501
• GBM: support hyperopt by @jppgks in #2490
• Fixes saved_weights_in_checkpoint docstring, mark as internal only by @dantreiman in #2506
• Fix test length of predictions by @tgaddair in #2507
• Fixed support for distributed datasets in create_auto_config by @tgaddair in #2508
• Config-first Datasets API (ludwig.datasets refactor) by @dantreiman in #2479
• Add in-memory dataset size calculation to dataset statistics by @arnavgarg1 in #2509
• Surfacing dataset statistics in hyperopt by @arnavgarg1 in #2515
• Adds multimodal benchmark datasets from AutoGluon paper by @dantreiman in #2512
• Adds goodbooks dataset by @dantreiman in #2514
• GBM: correctly compute early stopping by @jppgks in #2517
• Fixes mnist dataset image files not exporting by @dantreiman in #2520
• Fix get_best_model in hyperopt for Ray 1.12 by @arnavgarg1 in #2527
• Populate Parameter Metadata by @connor-mccorm in #2503
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2532
• Update README to be consistent with ludwig.ai home page. by @justinxzhao in #2530
• Add missing declarative ML image in README by @arnavgarg1 in #2533
• fix: Add missing titles/descriptions to various schemas by @ksbrar in #2516
• Cleanup: move to per-module loggers instead of the global logging object. by @justinxzhao in #2531
• Updated schedule logic for placement groups for ray backend by @magdyksaleh in #2523
• Nit: Parameter update tests grammar. by @justinxzhao in #2537
• Hyperopt: Log warning with num_extra_trials if all grid search parameters and num_samples > 1 by @arnavgarg1 in #2535
• Adds model configs to ludwig.datasets by @dantreiman in #2540
• ZScore Normalization Failure When Using Constant Value Number Feature by @arnavgarg1 in #2543
• Adds class names to calibration plot title, reformats Brier scores as grouped bar chart by @dantreiman in #2545
• Pin ray nightly version to avoid new test failures by @arnavgarg1 in #2548
• Added tests for init_config and render_config CLI commands by @tgaddair in #2551
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2554
• Ensure bfill/ffill leave no residual NaNs in the dataset during preprocessing by @arnavgarg1 in #2553
• Comprehensive configs: Explicitly list and save all parameter values for input and output features in configs. by @justinxzhao in #2460
• Fixing SettingWithCopyWarning when using get_repeatable_train_val_test_split by @abidwael in #2562
• Replace numerical with number in dataset zoo configs. by @justinxzhao in #2558
• Benchmarking toolkit wrap up by @abidwael in #2462
• Migrate to Raincloud plots for hyperopt report by @arnavgarg1 in #2561
• Remove global torchtext version-specific tokenizer availability warnings. by @justinxzhao in #2547
• Only create hyperopt pair plots when there is more than 1 parameter by @arnavgarg1 in #2560
• fix: Limit frequency array to top_n_classes in F1 viz by @hungcs in #2565
• int: unpin Dask version by @geoffreyangus in #2550
• Fixed typehint and removed unused utility function by @magdyksaleh in #2570
• AutoML: stratify imbalanced datasets by @jppgks in #2525
• Use Ray Air Checkpoint to sync files between trial workers by @tgaddair in #2577
• GBM bugfix: matching predictions LightGBM, hummingbird by @jppgks in #2574
• specify seed in RayDataset shuffling by @abidwael in #2566
• update logging message when early_stop: -1 by @abidwael in #2585
• update docker with torch wheel by @abidwael in #2584
• Refactors test_ray.py to minimize duplicate training jobs by @geoffreyangus in #2573
• Explanation API and feature importance for GBM by @jppgks in #2564
• Remove duplicate option by @connor-mccorm in #2593
• Quick fix: Don't show calibration validation set warnings unless calibration is actually enabled by @dantreiman in #2595
• Fixed issue when uploading output directory artifacts to remote filesystems by @tgaddair in #2598
• Add API Annotations to Ludwig by @arnavgarg1 in #2596
• Tweaks to the README (forward-ported from release-0.6) by @justinxzhao in #2603
• Extend test coverage for non-conventional booleans by @jppgks in #2601
• Fix assertions in training_determinism tests by @arnavgarg1 in #2606
• Ensure no ghost ray instances are running in tests by @arnavgarg1 in #2607
• Allow explicitly plumbing through nics by @tgaddair in #2605
• bug: fix relative import in optimizers.py by @ksbrar in #2600
• GBM: increase boosting_rounds_per_checkpoint to reduce evaluation overhead by @jppgks in #2612
• regression tests: add GBM model trained on v0.6.1 by @jppgks in #2611
• Relax test constraint to reduce flakiness in test_ray by @arnavgarg1 in #2610
• Add splitter that deterministically splits on an ID column by @tgaddair in #2615
• fix(explain): missing columns for fixed split by @jppgks in #2616
• Fixed hyperopt trial syncing to remote filesystems for Ray 2.0 by @tgaddair in #2617
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2622
• feat: adds max_batch_size to auto batch size functionality by @geoffreyangus in #2579
• Set commonly used parameters by @connor-mccorm in #2619
• Factor out defaults mixin change by @connor-mccorm in #2628
• Add type to custom combiner by @connor-mccorm in #2627
• Remove hyperopt from config when running train through cli by @arnavgarg1 in #2631
• Ensure resource availability for ray datasets workloads when running on cpu clusters by @arnavgarg1 in #2524
• Speed up horovod hyperopt tests and solve OOMs by @arnavgarg1 in #2599
• [explain] add API annotations by @jppgks in #2635
• Added storage backend API to allow injecting dynamic credentials by @tgaddair in #2630
• Update version to 0.7.dev by @justinxzhao in #2625
• Unpin Ray nightly in CI by @tgaddair in #2614
• Skip Horovod 0.26 installation, add packaging to requirements.txt by @arnavgarg1 in #2642
• [Annotations] Callbacks by @arnavgarg1 in #2641
• Fix automl by @connor-mccorm in #2639
• accepting dictionary as input to benchmarking.benchmark by @abidwael in #2626
• Fixed automl APIs to work with remote filesystems by @tgaddair in #2650
• Adds minimum split size, ensures random split is never smaller than minimum for local backend by @dantreiman in #2623
• Categorical passthrough encoder training failure fix by @abidwael in #2649
• Changes learning_curves to use "step" or "epoch" as x-axis label. by @dantreiman in #2578
• Remove Trainer type Param by @connor-mccorm in #2647
• Model performace in GitHub actions by @abidwael in #2568
• Fixed race condition in schema validation by @tgaddair in #2653
• Fixed --gpu_memory_limit in CLI to interpret as fraction of GPU memory by @tgaddair in #2658
• Stopgap solution for test_training_determinism by @connor-mccorm in #2665
• Added min and max to sample ratio by @connor-mccorm in #2655
• Set internal only flags by @connor-mccorm in #2659
• Add support for running pytest github action locally with act by @dantreiman in #2661
• Enforcing a 1 to 1 matching in names between Ludwig datasets and AutoGluon paper by @abidwael in #2666
• Added default arg to get_schema by @connor-mccorm in #2667
• remove duplicate news_popularity dataset by @abidwael in #2668
• Switch defaults to use mixins and improve test by @connor-mccorm in #2669
• Documents running local tests with act by @dantreiman in #2672
• Config Object by @connor-mccorm in #2426
• Unpin protobuf by @justinxzhao in #2673
• Check vocab size of category features, error out if only one category. Also adds error.py for custom error types. by @dantreiman in #2670
• Ordered Schema by @connor-mccorm in #2671
• Fix Regression Test Configs by @connor-mccorm in #2678
• Testing always() inside expansion in condition by @dantreiman in #2681
• Add protos to the Ludwig project: DatasetProfile messages and Whylogs messages. by @justinxzhao in #2674
• Allow Ray Tune callbacks to be passed into hyperopt and log model config by @jeffkinnison in #2640
• Check for nans before testing equality in test_training_determinism by @dantreiman in #2687
• Set saved_weights_in_checkpoint on encoder, not input feature by @dantreiman in #2690
• Use fully rendered config dictionary when accessing model.config by @tgaddair in #2685
• bug: Set additionalProperties to True for preprocessing schemas. by @ksbrar in #2620
• Bump support for torch 1.11.0 by @justinxzhao in #2691
• Fix validator for reduce_learning_rate_on_plateau by @carlogrisetti in #2692
• Use TensorArray to speed up writing predictions with Ray by @tgaddair in #2684
• Dataset size checks in preprocess_for_training by @dantreiman in #2688
• Remove Duplicate Schema Fields by @connor-mccorm in #2679
• Speed up tune_batch_size by using synthetic batches by @tgaddair in #2680
• Add bucketing_field Param to Trainer by @connor-mccorm in #2694
• Fix InputDataError to be serializeable by @tgaddair in #2695
• Adds PublicAPI annotation to api.py by @dantreiman in #2698
• Cleanup: move to per-module loggers instead of the global logging object. (2) by @justinxzhao in #2699
• Adds Ray implementation of IntegratedGradientsExplainer that distributes across cluster resources by @tgaddair in #2697
• Fixed bug with non-category outputs in RayIntegratedGradientsExplainer by @tgaddair in #2702
• Fix example values for max_batch_size in trainer parameter metadata by @connor-mccorm in #2705
• Fix incorrect internal_only flags on audio feature metadata by @connor-mccorm in #2704
• add customer churn datasets by @abidwael in #2703
• Add Kaggle test splits by @abidwael in #2675
• Fix ComparatorCombiner by @jppgks in #2689
• Actually print the torchinfo summary in print_model_summary() by @justinxzhao in #2696
• Add H&M fashion recommendation dataset by @jppgks in #2708
• Fix GBM ray nightly test by @jppgks in #2676
• Adds DeveloperAPI and PublicAPI annotations to AutoML by @dantreiman in #2701
• Remove obsolete v0 whylogs callback. by @justinxzhao in #2713
• fill_value / computed_fill_value fix by @connor-mccorm in #2714
• Add path to RayDataset by @tgaddair in #2716
• Fixed Horovod to be an optional import when doing Hyperopt by @tgaddair in #2717
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2722
• Adds annotation to download_one method in benchmarks by @dantreiman in #2712
• fix: Prevent shared parameter_metadata instances between defaults and _features. by @ksbrar in #2715
• Added ngram tokenizer by @tgaddair in #2723
• Revert "Add H&M fashion recommendation dataset (#2708)" by @jppgks in #2724
• Optimize search space for hyperopt tests to decrease test durations by @arnavgarg1 in #2730
• Add custom to_dask() to infer Dask metadata from Datasets schema. by @arnavgarg1 in #2728
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2735
• Bump Ludwig to Ray 2.0 by @arnavgarg1 in #2729
• Parameter Metadata Updates by @connor-mccorm in #2736
• Removes some vestigial code and replaces Tensorflow with PyTorch in comments by @dantreiman in #2731
• @DeveloperAPI annotations for backend module by @dantreiman in #2707
• int: Refactor test_ray.py to limit number of full train jobs by @geoffreyangus in #2637
• BaseTrainer: add empty barrier() by @jppgks in #2734
• Use whylogs to generate dataset profiles for pandas and dask dataframes. by @justinxzhao in #2710
• Add IntegerOptions marshmallow field by @ksbrar in #2739
• Downgrade to Ray 2.0 in CI to get green Ludwig CIs again. by @justinxzhao in #2742
• Adds @DeveloperAPI annotations to combiner classes by @dantreiman in #2744
• Use clearer error messages in ludwig serving, and enable serving to work with configs that have stratified splitting on target columns. by @justinxzhao in #2740
• Update Ray GPU Docker image to CUDA 11.6 by @tgaddair in #2747
• Fix #1735 by @herrmann in #2746
• Enable dataset window autosizing by @jeffkinnison in #2721
• Downgrade to PyTorch 1.12.1 in Docker to due to NCCL + CUDA compatibility by @tgaddair in #2750
• Replicate ludwig type inference, using the whylogs dataset profile. by @justinxzhao in #2743
• fix: Encountered unknown symbol 'foo' warning in Category feature preprocessing by @geoffreyangus in #2662
• Expand ~ in dataset download paths by @dantreiman in #2754
• Updates twitter bots example to new datasets API by @dantreiman in #2753
• fix: refactor IntegerOptions field by @ksbrar in #2755
• Added ray datasets repartitioning in cases of multiple train workers by @ShreyaR in #2756
• fix: Fix metadata object-to-JSON serialization for oneOf fields and add full schema serialization test. by @ksbrar in #2758
• refactor: Add ProtectedString field (alias of StringOptions that only allows one string) by @ksbrar in #2757
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2761
• Updates ludwig docker readme by @dantreiman in #2760
• Annotates ludwig.datasets API by @dantreiman in #2751
• Annotate MLFlow callback, and utility functions by @dantreiman in #2749
• Drishi sarcasmdataset 1 by @drishi in #2725
• Add local_rank to BaseTrainer by @tgaddair in #2766
• Public datasets by @connor-mccorm in #2752
• Fix typo by @connor-mccorm in #2767
• Correctly infer bool and object types in autoML by @arnavgarg1 in #2765
• feat: Hyperopt schema v0, part 1: Move output feature metrics from feature classes to feature configs. by @ksbrar in #2759
• Fix by @connor-mccorm in #2769
• Add ray version to runners by @arnavgarg1 in #2771
• Annotate Ludwig encoders and decoders by @arnavgarg1 in #2773
• Move preprocess callbacks inside model.preprocess by @jeffreyftang in #2772
• Fix benchmark tests, update latest metrics, and use the local backend for GBM benchmark tests by @abidwael in #2748
• Ensure correct output reduction for text encoders like MT5 and add warning messages when not supported by @arnavgarg1 in #2774
• CVE-2007-4559 Patch by @TrellixVulnTeam in #2770
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2776
• Fix double counting of training loss by @arnavgarg1 in #2775
• feat: Hyperopt schema v0, part 2: Make BaseMarshmallowConfig abstract by @ksbrar in #2779
• feat: Hyperopt schema v0, part 3: Enable optional min/max support for FloatTupleMarshmallowField fields by @ksbrar in #2780
• feat: Hyperopt schema v0, part 4: Add and use new hyperopt registry, search algorithm instantiation by @ksbrar in #2781
• Added exponential retry for mlflow, remote dataset loading by @ShreyaR in #2738
• Add synthetic test data integration test utils, and use them for loss value decrease tests. by @justinxzhao in #2789
• feat: Hyperopt schema v0, part 5: Add basic search algorithm, scheduler, executor, and hyperopt schemas. by @ksbrar in #2784
• Add benchmark as a pytest marker to avoid warnings. by @justinxzhao in #2786
• feat: Hyperopt schema v0, part 6: Enable new hyperopt schema by @ksbrar in #2785
• Add sentencepiece as a requirement, which is necessary for some hf models like mt5. by @justinxzhao in #2782
• [Annotations] Ludwig data modules by @arnavgarg1 in #2793
• [Annotations] Add DeveloperAPI annotations to Ludwig utils - Part 1 by @arnavgarg1 in #2794
• [Annotations] Annotations for Ludwig's utils - Part 2 by @arnavgarg1 in #2797
• [Annotations] Add annotations for schema module (part 1) - Model Config, Split, Trainer, Optimizers, Utils by @arnavgarg1 in #2798
• [Annotations] Annotate Schema Part 2: decoders, encoders, defaults, combiners, loss and preprocessing by @arnavgarg1 in #2799
• Add new data utility functions for buffers and files, and rename registry by @arnavgarg1 in #2796
• [Annotations] Ludwig Schema - Part 3: Features, Hyperopt and Metadata by @arnavgarg1 in #2800
• [Annotations] Add annotations for Ludwig's data utils (file readers) by @arnavgarg1 in #2795
• Proceed with model training even if saving preprocessed data fails. by @justinxzhao in #2783
• Improve warnings about backwards compatibility and dataset splitting. by @justinxzhao in #2788
• Generate structural change warnings and log_once functionality by @arnavgarg1 in #2801
• Broadcast progress tracker dict to all workers by @arnavgarg1 in #2804
• Start fresh training run if files for resuming training are missing by @arnavgarg1 in #2787
• LIghtGBMRayTrainer repartition datasets with fewer blocks than Ray actors by @jeffkinnison in #2806
• Add InterQuartileTransformer normalization strategy for Number Features by @arnavgarg1 in #2805
• Add negative sampling to ludwig.data by @jppgks in #2711
• Rectify output features in dataset config by @abidwael in #2768
• int: Add JSON markup to support unique input feature names. by @ksbrar in #2792
• int: Replace StringOptions usage with ProtectedString in split schemas by @ksbrar in #2808
• int: Replace StringOptions with ProtectedString for combiner schema type fields by @ksbrar in #2809
• refactor: Replace StringOptions with ProtectedString for encoder/decoder schema type fields by @ksbrar in #2810
• Upload Datasets to Remote Location by @connor-mccorm in #2764
• [Annotations] Annotate AutoML utils by @arnavgarg1 in #2812
• [Annotations] Ludwig Visualizations by @arnavgarg1 in #2813
• [Annotations] Logging Level Registry by @arnavgarg1 in #2814
• refactor: Replace StringOptions with ProtectedString for loss/hyperopt schema type fields by @ksbrar in #2816
• Define custom Ludwig types and replace Dict[str, Any] type hints with them. by @justinxzhao in #2556
• Config Object Bug Fix by @connor-mccorm in #2817
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2803
• AutoML libraries that use DatasetProfile instead of DatasetInfo by @justinxzhao in #2802
• Remove Sentencepiece by @connor-mccorm in #2821
• fix: account for max_batch_size config param in batch size tuning on cpu by @geoffreyangus in #2693
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2823
• refactor: Add filtering based on model_type for feature, combiner, and model type schemas by @ksbrar in #2815
• [TorchScript] Add user-defined HF Bert tokenizers by @geoffreyangus in #2733
• [Annotations] Move feature registries into accessor functions by @arnavgarg1 in #2818
• [Annotations] Encoder and Decoder Registries by @arnavgarg1 in #2819
• Speed Up Ray Image Tests by @geoffreyangus in #2828
• fix: Restrict allowed top-level config keys by @ksbrar in #2826
• Moves image decoding out of Ray Datasets to Dask Dataframe by @geoffreyangus in #2737
• Improve type hints and remove dead code for DatasetLoader module by @arnavgarg1 in #2833
• Update stratified split with a more specific exception for underpopulated classes by @jeffkinnison in #2831
• Add Ludwig contributors to README by @arnavgarg1 in #2835
• Fix key error in AutoML model select by @ShreyaR in #2824
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2836
• Drop incomplete batches for Ray and Pandas to prevent Batchnorm computation errors by @arnavgarg1 in #2778
• Catch and surface Runtime exceptions during preprocessing by @arnavgarg1 in #2839
• fix: Mark width and height as internal_only for image encoders by @ksbrar in #2842
• Select best batch size to maximize training throughput by @tgaddair in #2843
• Make batch_size=auto more consistent by using median of 5 steps by @tgaddair in #2846
• Make trainable=False default for all pretrained models by @tgaddair in #2844
• fix: Add back missing split fields by @ksbrar in #2848
• Pin scikit-learn<1.2.0 by @tgaddair in #2850
• text_encoder: RoBERTa max_sequence_length by @rudolfolah in #2852
• Fix TorchText version in tokenizers ahead of torch 1.13.0 upgrade by @geoffreyangus in #2838
• Fix trainable=False to freeze all params for HF encoders by @tgaddair in #2855
• Add support for automatic mixed precision (AMP) training by @tgaddair in #2857
• Evaluate training set in the training loop by @tgaddair in #2856
• Extend parameter guidance documentation for regularization, and add explicit maxes to Non-Negative floats by @justinxzhao in #2849
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2860
• Fixes for the roberta encoder: explicitly set max sequence length, and fix output shape computation by @justinxzhao in #2861
• Enables Set output feature on Ray by @geoffreyangus in #2791
• Add go module for dataset profile protos. by @justinxzhao in #2834
• fix: Upgrade expected_impact for trainable to MEDIUM on all encoders. by @ksbrar in #2865
• support stratified split with low cardinality features by @abidwael in #2863
• fix: load spacy model for lemmatization in EnglishLemmatizeFilterTokenizer to work by @abidwael in #2868
• Token-level explanations by @jppgks in #2864
• Replace learning rate: auto with feature type and encoder-based heuristics by @abidwael in #2854
• Set RayBackend Config to use single worker for tests by @arnavgarg1 in #2853
• Remove _to_tensors_fn from Ray Datasets by @geoffreyangus in #2866
• Remove ludwig-dev Dockerfile by @arnavgarg1 in #2873
• Support Ray GPU image with Torch 1.13 and CUDA 11.6 by @arnavgarg1 in #2869
• Use native LightGBM for intermittent eval during training by @jeffkinnison in #2829
• Set default validation metrics based on the output feature type. by @justinxzhao in #2820
• Auto resize images for ViTEncoder when use_pretrained is True or False by @arnavgarg1 in #2862
• TLE Backwards Compatibility Fixes by @jppgks in #2875
• Do not drop batch size dimension for single inputs by @jppgks in #2878
• Save GBM after training if not previously saved by @jeffkinnison in #2880
• Fix TLE - Pt. 2 by @connor-mccorm in #2881
• Tle fix by @connor-mccorm in #2883
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2885
• Convert schema metadata to YAML by @tgaddair in #2884
• Automatically infer vector_size for vector features when not provided by @tgaddair in #2888
• Support MLFlowCallback logging to an existing run by @tgaddair in #2892
• Fix dataset synthesizer by @connor-mccorm in #2894
• Add a clear error message about invalid column names in GBM datasets by @jeffkinnison in #2879
• Explicitly track all metrics related to the best evaluation in the progress tracker. by @justinxzhao in #2827
• Added DistributedStrategy interface with support for DDP by @tgaddair in #2890
• Adopt PyTorch official LRScheduler API by @tgaddair in #2877
• Annotate Confusion Matrix with updated cmap by @arnavgarg1 in #2899
• Dynamically resize confusion matrix and f1 plots by @arnavgarg1 in #2900
• Update backward compatibility tests for LR progress tracker changes made in #2877. by @justinxzhao in #2904
• fix: Fix vague initializer JSON schema titles. by @ksbrar in #2909
• Support Distributed Training And Ray Tune with Ray 2.1 by @arnavgarg1 in #2709
• Expand vision models to support pre-trained models by @jimthompson5802 in #2408
• Add ECD Descriptions by @connor-mccorm in #2897
• Simplify titanic example to read config in-line, and skip saving processed input. by @justinxzhao in #2912
• Adds quick fix for pretrained models not loading by modifying state_dict keys on load. by @dantreiman in #2911
• fix: Schema split conditions should pass in [TYPE] and not string by @hungcs in #2917
• Refactor metrics and metric tables and support adding more in-training metrics. by @justinxzhao in #2901
• Updated AutoML configs for latest schema and added validation tests by @tgaddair in #2921
• Adds backwards compatibility for legacy image encoders by @tgaddair in #2916
• Pin Torch to >=1.13.0 by @connor-mccorm in #2914
• Hyperopt invalid GBM config by @jppgks in #2926
• Store mlflow tracking URI to ensure consistency across processes by @tgaddair in #2927
• Update automl heuristics for fine-tuning and multi-modal tasks by @tgaddair in #2922
• Bump torch version for benchmark tests by @connor-mccorm in #2929
• Fix signing key by @tgaddair in #2928
• Adds safe_move_directory to fs_utils by @arnavgarg1 in #2931
• Added separate AutoML APIs for feature inference and config generation by @tgaddair in #2932
• Dynamic resizing for Confusion Matrix, Brier, F1 Plot, etc. by @arnavgarg1 in #2936
• Raise RuntimeError only for category output features with vocab size 1 by @arnavgarg1 in #2923
• Bump min python to 3.8 by @tgaddair in #2930
• Evaluate training set in the training loop (GBM) by @jppgks in #2907
• [automl] Exclude text fields with low avg words by @tgaddair in #2941
• [pre-commit.ci] pre-commit suggestions by @pre-commit-ci in #2944
• Fix pre-commit by removing manually specified blacken-docs dep. by @justinxzhao in #2949
• Rotate Brier Plot X-axis labels to 45 degree angle by @arnavgarg1 in #2948
• Retry HuggingFace pretrained model download on failure by @jeffkinnison in #2951
• Disable AUROC for CATEGORY features. by @justinxzhao in #2950
• Deactivate GBM random forest boosting type by @jeffkinnison in #2954
• Make batch_size=auto the default by @tgaddair in #2845
• Twitter bots test small improvements by @dantreiman in #2955
• Disable bagging when using GOSS GBM boosting type by @jeffkinnison in #2956
• Add missing standardize_image key to metadata by @jppgks in #2946
• Integrated Gradients: reset sample_ratio to 1.0 if set by @jppgks in #2945
• Increase CI pytest time out to 75 minutes by @jimthompson5802 in #2958
• Add sacremoses as a dependency for transformer_xl encoder by @arnavgarg1 in #2961
• Move all config validation to its own standalone module, config_validation. by @justinxzhao in #2959
• Fixes longformer encoder by passing in pretrained_kwargs correctly by @arnavgarg1 in #2963
• Expected Impact Calibration by @connor-mccorm in #2960
• Update Camembert by @geoffreyangus in #2966
• fix: Fix epochs suggested range by @ksbrar in #2965
• fix: enable binary dense encoder by @abidwael in #2957
• GBM DART boosting type incopatible with early stopping by @jeffkinnison in #2964
• Improving metadata config descriptions by @w4nderlust in #2933
• Fix ludwig-gpu image by @tgaddair in #2974
• Skip test_ray_outputs by @arnavgarg1 in #2935
• Enable custom HF BERT models with default tokenizer config by @geoffreyangus in #2973
• Update CamemBERT in schema by @geoffreyangus in #2975
• Set reduce_output to sum for XLM encoder by @arnavgarg1 in #2972
• Skipped mercedes_benz_greener.ecd.yaml benchmark test by @tgaddair in #2980
• Add sentencepiece as a requirement for MT5 text encoder by @arnavgarg1 in #2967
• Disable CTRL Encoder by @connor-mccorm in #2976
• MT5 reduce_output can't be cls_pooled - set to sum by default by @arnavgarg1 in #2981
• Populate hyperopt defaults using schema by @arnavgarg1 in #2968
• Revert "Add sentencepiece as a requirement for MT5 text encoder (#2967)" and disable MT5 Encoder by @arnavgarg1 in #2982
• Change default reduce_output strategy to sum for CamemBERT by @arnavgarg1 in #2984
• Set max_failures for Tuner to 0 by @geoffreyangus in #2987
• Fix TLE OOM for BERT-like models by @jppgks in #2990
• Reorder Advanced Parameters by @connor-mccorm in #2979
• [Hyperopt] Modify _get_best_model_path to grab it from the Checkpoint object with ExperimentAnalysis by @arnavgarg1 in #2985
• GBM: disable goss boosting type by @jppgks in #2986
• Adds HuggingFace pretrained encoder unit tests by @geoffreyangus in #2962
• [Hyperopt] Set default num_samples based on parameter space by @arnavgarg1 in #2997
• LR Scheduler Adjustments by @connor-mccorm in #2996
• fix: Force populate combiner registry inside of get_schema function. by @ksbrar in #2970
• fix: Fix validation and serialization for Boolean and OneOfOptionsField fields by @ksbrar in #2992
• Ray 2.2 compatibility by @arnavgarg1 in #2910
• Compute fixed text embeddings (e.g., BERT) during preprocessing by @tgaddair in #2867
• Use iloc to fetch first audio value. by @justinxzhao in #3006
• Fix Internal Only Param by @connor-mccorm in #3008
• Ludwig Dataclass by @connor-mccorm in #3005
• Cap batch_size=auto at 128 for CPU training by @tgaddair in #3007
• Added ghost batch norm option for concat combiner by @tgaddair in #3001
• Refactored norm layer and added additional norm at the start of the FCStack by @tgaddair in #3011
• Fix assignment that undoes tensor move to CPU by @jeffkinnison in #3012
• [Explain] Detach inputs before numpy processing by @jppgks in #3014
• Handle CUDA OOMs in explanations with retry and batch size halving by @tgaddair in #3015
• fix: Remove ecd_ray_legacy model type alias. by @ksbrar in #3013
• Explain fixes by @jppgks in #3016
• Remove null GBM trainer config options by @jeffkinnison in #2989
• Disable reuse_actors in hyperopt by @arnavgarg1 in #3017
• Skip Sarcos dataset during benchmark tests by @arnavgarg1 in #3020
• Explain: improve docstring about IntegratedGradient baseline for number features by @jppgks in #3018
• Upgrade isort to fix pre-commit. by @justinxzhao in #3027
• Limit batch size tuning to ≤20% of dataset size by @geoffreyangus in #3003
• [schema] Mark skip internal only by @jppgks in #3022
• Add specificity metric for binary features by @jppgks in #3025
• Added FSDP distributed strategy by @tgaddair in #3026
• Move on_batch_end callback to omit eval from batch duration during benchmarking by @geoffreyangus in #2898
• Set 0.7.beta by @justinxzhao in #3028
• Added missing file for fsdp by @tgaddair in #3033
• Cleaning up seed / random_seed usage discrepancy by @w4nderlust in #3021
• Filter Competitions by @connor-mccorm in #3032
• Hyperopt Quick Fix by @connor-mccorm in #3034
• Expected Impact and Ordering for GBM Params by @connor-mccorm in #3038
• Transformer Encoder - Representation Parameter Fix by @connor-mccorm in #2999
• Enables a new GitHub Action for slow tests by @geoffreyangus in #3029
• Skip BOHB test when using hyperopt with ray + horovod by @arnavgarg1 in #3036
• Fix gradient clipping typo by @geoffreyangus in #3039
• Fix checkpoint loading for HuggingFace encoders by @geoffreyangus in #3010
• Schema Polishing by @connor-mccorm in #3041
• Address some warnings when running hyperopt tests by @arnavgarg1 in #3040
• Removed log spam from distributed loader by @tgaddair in #3042
• [Hyperopt] Fix get_best_model_path by @arnavgarg1 in #3043
• Bump Ludwig images to Ray 2.2.0 by @geoffreyangus in #3023
• Revert "Bump Ludwig images to Ray 2.2.0 (#3023)" by @geoffreyangus in #3044
• add httpx as required by starlette>=0.21.0 by @jppgks in #3047
• Raise exceptions from async batch producer thread on the main thread by @tgaddair in #3050
• Set zscore normalization as the default normalization strategy for number features by @arnavgarg1 in #3051
• Fix TLE: safe divide by zero + normalize at sequence level by @jppgks in #3046
• Fix fill_with_mode when using Dask by @arnavgarg1 in #3054
• Refactored ModelConfig object into a Marshmallow schema by @tgaddair in #2906
• Fix LR reduce on plateau interaction with base LR decay by @tgaddair in #3056
• Update schema to correctly reflect supported missing value strategies for different feature types by @arnavgarg1 in #3053
• Improve observability when using cached datasets during preprocessing by @arnavgarg1 in #3058
• Quick fix for cached logging by @arnavgarg1 in #3060
• Remove passthrough encoder from sequence and text features encoder registry by @jeffkinnison in #3061
• Remove RNN invalid cell types from the schema by @jeffkinnison in #3062
• Round confusion matrix numbers to 3 decimal places by @arnavgarg1 in #3065
• Fixed handling of {} hyperopt config section by @tgaddair in #3064
• Deflake test_tune_batch_size_lr_cpu by @justinxzhao in #3067
• Feature: Data Augmentation for Image Input Features by @jimthompson5802 in #2925
• Update transformer hidden_size / num_heads error message by @jeffkinnison in #3066
• Add MPS device support by @tgaddair in #3072
• Require env var LUDWIG_USE_MPS to enable MPS by @tgaddair in #3074
• Generate proc_column only after all preprocessing parameters are merged in to prevent incorrect cached dataset reads by @arnavgarg1 in #3069
• Remove duplicate validation field validation. by @justinxzhao in #3070
• Remove previous ModelConfig implementation and refactor to use __post_init__ by @tgaddair in #3083
• Set RunConfig verbosity to 0 by @tgaddair in #3085
• Fix default image on image read failure by @geoffreyangus in #3073
• Add Precision Recall curves to Ludwig by @arnavgarg1 in #3084
• Log number of rows dropped by DROP_ROWS strategy by @geoffreyangus in #3087
• Allow providing a Ludwig dataset as a URI of the form ludwig://<dataset> by @tgaddair in #3082
• Fixed augmentation schema check by @tgaddair in #3090
• XLNet: disable "uni" attention type by @jppgks in #3097
• Only show drop row logging if rows are dropped by @arnavgarg1 in #3094
• Upgrade torchmetrics to 0.11.1. Add ROC metrics for category features. Add sequence accuracy, char error rate, and perplexity metrics for text features. by @justinxzhao in #3035
• Disallow certain config parameters from accepting null as a value by @abidwael in #3079
• Deflake the lbgfs optimizer test by @justinxzhao in #3100
• Use window_size_bytes: auto to specify automatic windowing by @jeffkinnison in #3076
• Use proc col hash for checksum computation by @arnavgarg1 in #3095
• Fixed ethos_binary dataset to threshold the label at 0.5 by @tgaddair in #3102
• Add -1 as a valid negative class for binary type inference by @arnavgarg1 in #3101
• GBM: remove distributed=False from RayDMatrix by @jppgks in #3099
• Switch combiner num_fc_layers to expected impact 3 by @connor-mccorm in #3103
• Add a registry of additional config checks to check inter-parameter incompatibilities. by @justinxzhao in #3024
• Adds config parameters to replace outliers via a missing_value_strategy by @tgaddair in #3080
• Fixed serialization and deserialization of augmentation configuration by @jimthompson5802 in #3096
• Unregister CTRL and MT5 encoders since they have tensor placement and sentencepiece segfault issues by @arnavgarg1 in #3106
• Pin torch nightly to Feb 13, 2023 by @arnavgarg1 in #3110
• Resize confusion matrix properly by @arnavgarg1 in #3109
• Fold all validation into ModelConfig. by @justinxzhao in #3104
• fix: add pytest to hashfiles to be more selective about caching by @abidwael in #3113
• Disable XLM Text Encoder because of host memory pressure issues by @arnavgarg1 in #3108
• Bump ludwig docker image ray220 by @geoffreyangus in #3111
• Remove XLM encoder from slow encoders test by @geoffreyangus in #3114
• Added additional dropout to concat by @tgaddair in #3116
• Update field descriptions for ludwig-docs by @tgaddair in #3123
• Transformer divisibility error validation by @jeffkinnison in #3105
• fix: Updated Learning Rate decay_rate to use corresponding Metadata. by @martindavis in #3128
• refactor: Use TypeSelection to power optimizer field by @ksbrar in #3071
• refactor: Add separate ECD and GBM defaults schemas by @ksbrar in #3124
• Skip batch norm when chunk size is 1 in GhostBatchNorm by @arnavgarg1 in #3119
• Fix tests to be compliant with latest version of whylogs by @justinxzhao in #3131
• Make positive class weight a float. by @justinxzhao in #3133
• feat: Raise deprecation warnings for unknown parameters by @ksbrar in #3118
• Align sequence encoder descriptions with ludwig-docs by @tgaddair in #3135
• [Explain] always return global and row-level explanations by @jppgks in #3132
• Update benchmark tests by @abidwael in #3115
• [Hyperopt] Load checkpoints directly from the object store in Ray 2.2 by @arnavgarg1 in #3037
• Decouple loss schema from implementation by @tgaddair in #3141
• Adds non-slow HF unit test to validate constant value by @geoffreyangus in #3130
• Updated decoder and loss schemas, removed dep from schema -> loss_modules by @tgaddair in #3140
• Init minimal config by @tgaddair in #3143
• Updated HF long descriptions by @tgaddair in #3144
• Disallow GPU tensors in GBM training and eval by @jeffkinnison in #3139
• Not default to adding None as accepted tuple for FloatRangeTupleDataclassField by @abidwael in #3146
• Ray 2.3 Compatibility by @arnavgarg1 in #3009
• Unpin pyarrow by @arnavgarg1 in #3149
• Update Ludwig version to 0.7 by @arnavgarg1 in #3148

New Contributors

• @Marvjowa made their first contribution in #2236
• @Dennis-Rall made their first contribution in #2192
• @abidwael made their first contribution in #2263
• @noahlh made their first contribution in #2284
• @jeffkinnison made their first contribution in #2316
• @andife made their first contribution in #2358
• @alberttorosyan made their first contribution in #2413
• @herrmann made their first contribution in #2746
• @drishi made their first contribution in #2725
• @TrellixVulnTeam made their first contribution in #2770
• @rudolfolah made their first contribution in #2852
• @martindavis made their first contribution in #3128

Full Changelog: v0.6.4...v0.7