We're happy to announce the AutoGluon 0.4.1 release. 0.4.1 contains minor enhancements to Tabular, Text, Image, and Multimodal modules, along with many quality of life improvements and fixes.
This release is **non-breaking** when upgrading from v0.4.0. As always, only load previously trained models using the same version of AutoGluon that they were originally trained on. Loading models trained in different versions of AutoGluon is not supported.
This release contains [**55** commits from **10** contributors](https://github.com/awslabs/autogluon/graphs/contributors?from=2022-03-10&to=2022-05-23&type=c)!
See the full commit change-log here: https://github.com/awslabs/autogluon/compare/v0.4.0...v0.4.1
Special thanks to yiqings, leandroimail, huibinshen who were first time contributors to AutoGluon this release!
Full Contributor List (ordered by of commits):
- Innixma, zhiqiangdon, yinweisu, sxjscience, yiqings, gradientsky, willsmithorg, canerturkmen, leandroimail, huibinshen.
This version supports Python versions 3.7 to 3.9.
Changes
AutoMM
New features
- Added `optimization.efficient_finetune` flag to support multiple efficient finetuning algorithms. (1666) sxjscience
- Supported options:
- `bit_fit`: ["BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models"](https://arxiv.org/abs/2106.10199)
- `norm_fit`: An extension of the algorithm in ["Training BatchNorm and Only BatchNorm: On the Expressive Power of Random Features in CNNs"](https://arxiv.org/abs/2003.00152) and BitFit. We finetune both the parameters in the norm layers as long as the biases.
- Enabled knowledge distillation for AutoMM (1670) zhiqiangdon
- Distillation API for `AutoMMPredictor` reuses the `.fit()` function:
python
from autogluon.text.automm import AutoMMPredictor
teacher_predictor = AutoMMPredictor(label="label_column").fit(train_data)
student_predictor = AutoMMPredictor(label="label_column").fit(
train_data,
hyperparameters=student_and_distiller_hparams,
teacher_predictor=teacher_predictor,
)
- Option to turn on returning feature column information (1711) zhiqiangdon
- The feature column information is turned on for feature column distillation; for other cases it is turned off by default to reduce dataloader‘s latency.
- Added a `requires_column_info` flag in data processors and a utility function to turn this flag on or off.
- FT-Transformer implementation for tabular data in AutoMM (1646) yiqings
- Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, Artem Babenko, "Revisiting Deep Learning Models for Tabular Data" 2022. ([arxiv](https://arxiv.org/abs/2106.11959), [official implementation](https://github.com/Yura52/tabular-dl-revisiting-models))
- Make CLIP support multiple images per sample (1606) zhiqiangdon
- Added multiple images support for CLIP. Improved data loader robustness: added missing images handling to prevent training crashes.
- Added the choice of using a zero image if an image is missing.
- Avoid using `eos` as the sep token for CLIP. (1710) zhiqiangdon
- Update fusion transformer in AutoMM (1712) yiqings
- Support constant learning rate in `polynomial_decay` scheduler.
- Update `[CLS]` token in numerical/categorical transformer.
- Added more image augmentations: `verticalflip`, `colorjitter`, `randomaffine` (1719) Linuxdex, sxjscience
- Added prompts for the percentage of missing images during image column detection. (1623) zhiqiangdon
- Support `average_precision` in AutoMM (1697) sxjscience
- Convert `roc_auc` / `average_precision` to `log_loss` for torchmetrics (1715) zhiqiangdon
- `torchmetrics.AUROC` requires both positive and negative examples are available in a mini-batch. When training a large model, the per gpu batch size is probably small, leading to an incorrect `roc_auc` score. Conversion from `roc_auc` to `log_loss` improves training stablility.
- Added `pytorch-lightning` 1.6 support (1716) sxjscience
Checkpointing and Model Outputs Changes
- Updated the names of top-k checkpoint average methods and support customizing model names for terminal input (1668) zhiqiangdon
- Following paper: https://arxiv.org/pdf/2203.05482.pdf to update top-k checkpoint average names: `union_soup` -> `uniform_soup` and `best_soup` -> `best`.
- Update function names (`customize_config_names` -> `customize_model_names` and `verify_config_names` -> `verify_model_names`) to make it easier to understand them.
- Support customizing model names for the terminal input.
- Implemented the GreedySoup algorithm proposed in [paper](https://arxiv.org/pdf/2203.05482.pdf). Added `union_soup`, `greedy_soup`, `best_soup` flags and changed the default value correspondingly. (#1613) sxjscience
- Updated the `standalone` flag in `automm.predictor.save()` to save the pertained model for offline deployment (1575) yiqings
- An efficient implementation to save the donwloaded models from transformers for the offline deployment. Revised logic is in 1572, and discussed in 1572 (comment).
- Simplified checkpoint template (1636) zhiqiangdon
- Stopped using pytorch lightning's model checkpoint template in saving `AutoMMPredictor`'s final model checkpoint.
- Improved the logic of continuous training. We pass the `ckpt_path` argument to pytorch lightning's trainer only when `resume=True`.
- Unified AutoMM's model output format and support customizing model names (1643) zhiqiangdon
- Now each model's output is dictionary with the model prefix as the first level key. The format is uniform between single model and fusion model.
- Now users can customize model names by using the internal registered names (`timm_image`, `hf_text`, `clip`, `numerical_mlp`, `categorical_mlp`, and `fusion_mlp`) as prefixes. This is helpful when users want to simultaneously use two models of the same type, e.g., `hf_text`. They can just use names `hf_text_0` and `hf_text_1`.
- Support `standalone` feature in `TextPredictor` (1651) yiqings
- Fixed saving and loading tokenizers and text processors (1656) zhiqiangdon
- Saved pre-trained huggingface tokenizers separately from the data processors.
- This change is backwards-compatibile with checkpoints saved by verison `0.4.0`.
- Change load from a classmethod to staticmethod to avoid incorrect usage. (1697) sxjscience
- Added `AutoMMModelCheckpoint` to avoid evaluating the models to obtain the scores (1716) sxjscience
- checkpoint will save the best_k_models into a yaml file so that it can be loaded later to determine the path to model checkpoints.
- Extract column features from AutoMM's model outputs (1718) zhiqiangdon
- Add one util function to extract column features for both image and text.
- Support extracting column features for models `timm_image`, `hf_text`, and `clip`.
- Make AutoMM dataloader return feature column information (1710) zhiqiangdon
Bug fixes
- Fixed calling `save_pretrained_configs` in `AutoMMPrediction.save(standalone=True)` when no fusion model exists ([here](https://github.com/awslabs/autogluon/blob/5a323641072431091d2be5e6dbef5a87b646a408/text/src/autogluon/text/automm/utils.py#L644 )) (1651) yiqings
- Fixed error raising for setting key that does not exist in the configuration (1613) sxjscience
- Fixed warning message about bf16. (1625) sxjscience
- Fixed the corner case of calculating the gradient accumulation step (1633) sxjscience
- Fixes for top-k averaging in the multi-gpu setting (1707) zhiqiangdon
Tabular
- Limited RF `max_leaf_nodes` to 15000 (previously uncapped) (1717) Innixma
- Previously, for very large datasets RF/XT memory and disk usage would quickly become unreasonable. This ensures that at a certain point RF and XT will no longer become larger given more rows of training data. Benchmark results showed that the change is an improvement, particularly for the `high_quality` preset.
- Limit KNN to 32 CPUs to avoid OpenBLAS error (1722) Innixma
- Issue 1020. When training K-nearest-neighbors (KNN) models, sometimes a rare error can occur that crashes the entire process:
BLAS : Program is Terminated. Because you tried to allocate too many memory regions.
Segmentation fault: 11
This error occurred when the machine had many CPU cores (>64 vCPUs) due to too many threads being created at once. By limiting to 32 cores used, the error is avoided.
- Improved memory warning thresholds (1626) Innixma
- Added `get_results` and `model_base_kwargs` (1618) Innixma
- Added `get_results` to searchers, useful for debugging and for future extensions to HPO functionality.
Added new way to init a `BaggedEnsembleModel` that avoids having to init the base model prior to initing the bagged ensemble model.
- Update resource logic in models (1689) Innixma
- Previous implementation would crash if user specified `auto` for resources, fixed in this PR.
- Added `get_minimum_resources` to explicitly define minimum resource requirements within a method.
- Updated feature importance default `subsample_size` 1000 -> 5000, `num_shuffle_sets 3` -> 5 (1708) Innixma
- This will improve the quality of the feature importance values by default, especially the 99% confidence bounds. The change increases the time taken by ~8x, but this is acceptable because of the numerous inference speed optimizations done since these defaults were first introduced.
- Added notice to ensure serializable custom metrics (1705) Innixma
Bug fixes
- Fixed `evaluate` when `weight_evaluation=True` (1612) Innixma
- Previously, AutoGluon would crash if the user specified `predictor.evaluate(...)` or `predictor.evaluate_predictions(...)` when `self.weight_evaluation==True`.
- Fixed RuntimeError: dictionary changed size during iteration (1684, 1685) leandroimail
- Fixed CatBoost custom metric & F1 support (1690) Innixma
- Fixed HPO not working for bagged models if the bagged model is loaded from disk (1702) Innixma
- Fixed Feature importance erroring if `self.model_best` is `None` (can happen if no Weighted Ensemble is fit) (1702) Innixma
Documentation
- updated the text tutorial of cutomizing hyperparameters (1620) zhiqiangdon
- Added customizeable backbones from the Huggingface model zoo and how to use local backbones.
- Improved implementations and docstrings of `save_pretrained_models` and `convert_checkpoint_name`. (1656) zhiqiangdon
- Added cheat sheet to website (1605) yinweisu
- Doc fix to use correct predictor when calling leaderboard (1652) Innixma
Miscellaneous changes
- [security] updated `pillow` to `9.0.1`+ (1615) gradientsky
- [security] updated `ray` to `1.10.0`+ (1616) yinweisu
- Tabular regression tests improvements (1555) willsmithorg
- Regression testing of model list and scores in tabular on small synthetic datasets (for speed).
- Tests about 20 different calls to `TabularPredictor` on both regression and classification tasks, multiple presets etc.
- When a test fails it dumps out the config change required to make it pass, for ease of updating.
- Disabled image/text predictor when gpu is not available in `TabularPredictor` (1676) yinweisu
- Resources are validated before bagging is started. Image/text predictor model would require minimum of 1 gpu.
- Use class property to set keys in model classes. In this way, if we customize the prefix key, other keys are automatically updated. (1669) zhiqiangdon
Various bugfixes, documentation and CI improvements
- yinweisu (1605, 1611, 1631, 1638, 1691)
- zhiqiangdon (1721)
- Innixma (1608, 1701)
- sxjscience (1714)