v0.3.3

[0.3.3] - 2024-10-29

Added

• Add BiQwen2 model

Changed

• Modified ColQwen and BiQwen to prevent the useless forward pass in the last layer of the original model (classification head)
• Bumped "breaking" dependencies on MTEB and Transformers version and made the corresponding changes in the code
• Casted Image dtype in ColPali due to breaking 4.46 transformers update
• Added a "num_image_tokens" kwarg to the ColQwen2Processor to allow for different image resolutions

Fixed

• Fix wrong variable name for ColPaliProcessor's prefixes

Full Changelog: v0.3.2...v0.3.3

v0.3.2: The interpretability update

Description

✨ This release brings the interpretability module to colpali-engine and adds support for generating similarity maps with the ColQwen2 model.

🛠️ We’ve also made several code improvements and added tests for ColQwen2 to ensure better performance and reliability.

Features

Added

• Restore, refactor, and improve interpretability module for generating similarity maps

Changed

• Remove dummy image from ColPaliProcessor.process_queries

Fixed

• Fix the compute_hardnegs.py script

Tests

• Add missing model.eval() in tests
• Add tests for ColQwen2

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

v0.3.1: ColQwen2

[0.3.1] - 2024-09-27

Added

• Add module-level imports for collators
• Add sanity check in the run inference example script
• Add E2E test for ColPali
• Add Qwen2-VL support

Changed

• Improve code clarity the run inference example script
• Subset the example dataset in the run inference example script
• Rename scorer test to test_processing_utils
• Greatly simplify routing logic in Trainer selection and when feeding arguments to the model forward pass (refacto)
• Removed class ContrastiveNegativeTrainer which is now just integrated in ContrastiveTrainer. This should not affect the user-facing API.
• Bumped transformers version to 4.45.0 to get Qwen2-VL support

Fixed

• Import HardNegCollator at module-level if and only if datasets is available
• Remove the need for typer in the run inference example script
• Fix edge case when empty suffix "" given to processor
• Fix bug in HardNegCollator since 0.3.0

Full Changelog: v0.3.0...v0.3.1

v0.3.0: Extensive package refacto

Description

✨ This release is an extensive package refacto, making ColPali more modular and easier to use.

🚨 It is NOT backward-compatible with previous versions.

Features

Added

• Restructure the utils module
• Restructure the model training code
• Add custom Processor classes to easily process images and/or queries
• Enable module-level imports
• Add scoring to processor
• Add CustomRetrievalEvaluator
• Add missing typing
• Add tests for model, processor, scorer, and collator
• Lint Changelog
• Add missing docstrings
• Add "Ruff" and "Test" CI pipelines

Changed

• Restructure all modules to closely follow the transformers architecture
• Hugely simplify the collator implementation to make it model-agnostic
• ColPaliProcessor's process_queries doesn't need a mock image input anymore
• Clean pyproject.toml
• Loosen the required dependencies
• Replace black with the ruff linter

Removed

• Remove interpretability and eval_manager modules
• Remove unused utils
• Remove TextRetrieverCollator
• Remove HardNegDocmatixCollator

Fixed

• Fix wrong PIL import
• Fix dependency issues

Full Changelog: v0.2.2...v0.3.0

v0.2.2

Fix: Remove forced "cuda" usage in Retrieval Evaluator

v0.2.1

[0.2.1] - 2024-09-02

Patch query preprocessing helper function disalignement with training scheme.

Fixed

• Add 10 extra pad token by default to the query to act as reasoning buffers. This was added in the collator but not the external helper function for inference purposes.

v0.2.0

[0.2.0]

Large refactoring to adress several issues and add features. This release is not backward compatible with previous versions.
The models trained under this version will exhibit degraded performance if used with the previous version of the code and vice versa.

Branch

Added

• Added multiple training options for training with hard negatives. This leads to better model performance !
• Added options for restarting training from a checkpoint.

Changed

• Optionally load ColPali models from pre-initialized backbones of the same shape to remove any stochastic initialization when loading adapters. This fixes 11 and 17.

Fixed

• Set padding side to right in the tokenizer to fix misalignement issue between different query lengths in the same batch. Fixes 12
• Add 10 extra pad token by default to the query to act as reasoning buffers. This enables the above fix to be made without degrading performance and cleans up the old technique of using tokens.

v0.1.1: Reference release for the ColPali paper

Initial release

This release contains the code of reference for the ColPali arXiv paper [url]. In particular, it contains the model architecture, the loss function, and the trainer used for training ColPali.

To use this version of colpali-engine, install the package with:

pip install colpali-engine==0.1.1