- ef895f3: Include GraSH for hyperparameter tuning (thanks @fniesel). Paper, original implementation.
- 3ba5ef8: Update versions of requirements (thanks @fniesel).
- PR #265: implementation of regularization by N3 norm (thanks @Kenkoko)
- PR #263: fixed bug in application of penalty to models with reciprocal relations (thanks @Kenkoko)
- 5127cf2: Fix differences in TransE scoring implementations
- PR #224: Take floating point issues into account for tie calculation in entity-ranking (thanks @sfschouten)
- 9a4f69a: Refactor time measurement with new timer class
- PR #191: Fix loading of pretrained embeddings with reciprocal relation models
- 27e8a32: improve validation time by allowing bulk KvsAll index lookup and improved history computation
- PR #154: store checkpoint containing the initialized model for reproducibility
- 9e88117: Add Transformer model and learning rate warmup (thanks nluedema)
- PR #176: Add TransH model (thanks Mayo42)
- PR #164: Allow to easily add custom training/evaluation/search jobs
- PR #159: Add a plugin mechanism (thanks @sfschouten)
- PR #157: Add CoDEx datasets and pretrained models (thanks @tsafavi)
- PR #155: Faster reading of triple files
- d275419, 87c5463: Support parameter groups with group-specific optimizer args
- PR #152: Added training loss evaluation job
- PR #147: Support both minimization and maximization of metrics
- PR #144: Support to tune subbatch size automatically
- PR #143: Allow processing of large batches in smaller subbatches to save memory
- PR #140: Calculate penalty for entities only once, if subject-embedder == object-embedder
- PR #138: Revision of hooks, fix of embedding normalization
- PR #135: Revised sampling API, faster negative sampling with shared samples
- PR #112: Initialize embeddings from a packaged model
- PR #113: Reduce memory consumption and loading times of large datasets
- Various smaller improvements and bug fixes
- PR #110: Support for different tie-breaking methods in evaluation (thanks Nzteb)
- 1d26e63: Add head/tail evaluation per relation type
- dfd0aac: Added squared error loss (thanks Nzteb)
- PR #104: Fix incorrect relation type measurement (thanks STayinloves)
- PR #101: Revise embedder penalty API (thanks Nzteb)
- PR #94: Support for packaged models (thanks AdrianKS)
- Improved seeding of workers when a fixed NumPy seed is used
- Various smaller improvements and bug fixes
- Added more mappings from entity IDs to names for Freebase datasets (in entity_strings.del file)
- Improved shared negative sampling (WOR sampling, exclude positive triples from negative sample)
- PR #86: Support (s,?,o) queries for KvsAll training (thanks vonVogelstein)
- cf64dd2: Fast dataset/index loading via cached pickle files
- 4bc86b1: Add support for chunking a batch when training with negative sampling
- 14dc926: Add ability to dump configs in various ways
- PR #64: Initial support for frequency-based negative sampling (thanks AdrianKS)
- PR #77: Simpler use of command-line interface (thanks cthoyt)
- 76a0077: Added RotatE
- 7235e99: Added option to add a constant offset before computing BCE loss
- 67de6c5: Added CP
- a5ee441: Added SimplE
- PR #71: Faster and more memory-efficient training with negative sampling (thanks AdrianKS)
- Initial release