v0.4.3

What's Changed

• Remove hardcoded registry file by @lfoppiano in #199

Full Changelog: v0.4.2...v0.4.3

v0.4.2

What's Changed

• fix: ensure learning rate schedule handles tensor steps correctly by @lfoppiano in #204
• Make the documentation working again! by @lfoppiano in #205
• Handle empty or missing directory for embeddings by @lfoppiano in #206

Full Changelog: v0.4.1...v0.4.2

v0.4.1

Breaking changes

• TensorFlow 2.17 / tf_keras 2.17: DeLFT now requires TensorFlow 2.17.1 and the standalone tf_keras 2.17.0 package. All Keras imports updated from tensorflow.keras to tf_keras. Pre-trained model weights from 0.3.x are not compatible and must be retrained. (#180, #181)
• Python 3.10+ required: Python 3.8 and 3.9 are no longer supported. (#180)
• CUDA 12.1 required for GPU: On Linux, torch is no longer in the base install to avoid CUDA conflicts. Use pip install "delft[gpu]" with the PyTorch cu121 index. (#195)
• LMDB embedding format changed: Embeddings stored as raw float32 bytes instead of pickle. Enables Java interoperability (GROBID). Convert with: python -m delft.utilities.convert_lmdb_embeddings --input <old> --output <new> (#180)
• ELMo support removed: Use transformer-based or static embeddings instead. (#190, #192)

New features

• Weights & Biases integration for experiment tracking (--wandb flag) (#172)
• Distributed training support via SLURM scripts (#191)
• Configurable num_workers parameter for data loading (defaults to CPU count - 1) (#183)
• BPE preprocessor auto-detection (#184)
• Safetensors format support for local models
• Licenses/copyrights classifier application
• LMDB format validation and conversion utility (#180)
• CI release pipeline with automated PyPI publishing (#196)
• Contributing guidelines (CONTRIBUTING.md) (#200)
• Code reformatting and lint enforcement with ruff (#197)

Bug fixes

• Fixed crash when accessing learning rate during training (#180)
• Fixed multiprocessing issues on macOS (#180)
• Fixed empty embeddings handling with additional validation checks (#180)
• Fixed resource-registry.json now bundled as package data (no more hardcoded path) (#201, #202)
• Fixed word2vec embedding URL (now uses HuggingFace for GloVe instead of Stanford) (#180)
• Fixed compatibility with newer versions of transformers library (#180)

Other changes

• Migrated build config from setup.py to pyproject.toml (#195)
• Retrained all models with TensorFlow 2.17 (#191)
• Updated dependencies (transformers 4.48, torch 2.5.1, numpy 1.26.4, scikit-learn 1.6.1, pandas 2.2.3)
• Improved GitHub Actions CI with multi-platform support (#171)
• Improved LMDB embedding concurrency with per-thread handles (#180)

Full Changelog: v0.3.4...v0.4.1

Version 0.3.4

• support multiple GPU training/inference (--multi-gpu parameter)
• support safetensors model weights format
• support private HuggingFace models
• in application scripts: add max-epoch parameter, learning rate parameter
• add grobid model for funding and acknowledgement information
• more parameter information printed when training
• some dependency updates

Version 0.3.3

with PyPI:

pip install delft==0.3.3

• support for incremental training
• fix SciBERT tokenier initialization from HuggingFace model
• updated HuggingFace transformers library to 4.25.1 and tensorflow to 2.9.3
• review the support of BPE tokenizers in the case of pre-tokenized input with the updated transformers library for most transformer models using it (tested with Roberta/GPT2, CamemBERT, bart-base, albert-base-v2, and XLM)
• addition of some model variants for sequence labeling (BERT_FEATURES, BERT_ChainCRF_FEATURES)

Version 0.3.2

with PyPI:

pip install delft==0.3.2

• Print model parameters at creation and load time
• Dataset recognition
• Model updates
• Set feature channel embeddings trainable

Full Changelog: v0.3.1...v0.3.2

Version 0.3.1

• fix a problem with CRF tensorflow-addons when batch size is 1

Version 0.3.0

• Migration of DeLFT to TensorFlow 2.7
• Support of HuggingFace transformer library (auto* library)
• New architectures and updated models
• General usage of optimizer with learning rate decay
• Updated docs now via readthedoc
• Improved ELMo embeddings
• Transformers wrapper to limit usage of Hugging Face hub only necessary, model with transformer layer fully portable without hub access

Version 0.2.6

• add automatic download of embeddings if not locally available
• enable embedding preload script for docker image

Version 0.2.5

• fix serialization of models with feature preprocessor (PR #110)
• update grobid models with features
• some other models and score updates
• add "software was used" classification model for software citations
• update tensorflow dependency