2.2407.0

• Introduction of Logistic Regression Classifier: Added a new classification kernel leveraging Logistic Regression for efficient text categorization without the need for fine-tuning.

• Support for Multiple Pooling Strategies: Implemented various pooling strategies, including MEAN, MEAN_MASKED, MAX, MAX_MASKED, CLS, SUM, and ATTENTION_WEIGHTED for flexible embedding generation.

• Template and Instruct Models: Introduced support for instruct templates with models like intfloat/multilingual-e5-large-instruct to enhance performance by utilizing structured templates.

• Model Export and HuggingFace Integration: Simplified the process of saving and publishing models to HuggingFace with automatic model cards and additional metadata such as tags and languages.

• Inference Server: Added a dockerized inference server with an HTTP API to facilitate deployment. This includes new scripts for starting the server both in a docker container and on a host machine.

• Improved Documentation: Updated and expanded documentation, including examples for training models, classification kernels, pooling strategies, model export, and inference.

1.2406.5

Centroid Classifier Refactor:

• Normalization Improvements: Introduced _normalize method for efficient tensor normalization using torch.nn.functional.normalize.
• Training Enhancements:
  • train method now calculates centroids using mean embeddings for each label.
  • Centroids are stored and normalized upon training.
• Prediction Optimization:
  • Improved predict and predict_one methods to utilize normalized centroids.
  • Replaced cosine similarity calculations with dot product for faster computations.

Interface Changes:

• Updated get_embeddings method to yield torch.Tensor instead of numpy.ndarray.
• Removed redundant code and streamlined embedding extraction process.

Embedding Model Initialization:

• Ensured the embedding model is set to evaluation mode immediately after loading to improve inference efficiency (self._model.eval()).