0.0.5

Updates since 0.0.4:

We have made some big refactors since 0.0.4. We list some of them here:

• We separate a lot of processes into their own respective files and we will expand on some of them below.
• The notion of defining train or inference is gone. We separate these processes into their own respective .py files (e.g., ecg_bench/evaluate_elm.py, ecg_bench/train_elm.py).
• We also simplified and separated the ELM and its components. Simply define the --llm or --encoder flags for the model you are considering to run some processes on.
• Previously, many configurations were hardcoded arbitrarily throughout the code, making it difficult to find these hardcoded values. We still hardcode most of these values but unify them in ecg_bench/configs/constants.py. We found this to be much easier to work with. We will continuously work to abstract the code to get rid of these values (if necessary).

0.0.4

Updates since 0.0.3:

• Release stratified data splits for many of our ECG dataset on Hugging Face!
• Add naive encoder free method.
• Code cleanups (got rid of percentiles based sampling and unused configs)
• automatic learning rate scaler for effective batch size and distributed training
• Tensor broadcasting bug fix (we arange over the batch to properly allocate the corresponding signal latent for each batch)

0.0.3

Some updates:

• Added some more known bugs on seeding and non-deterministic cuda behavior
• Fixed bugs in padding for second stage.
• Differentiated scheduler between training transformers (LLM) in second or end-to-end stage vs training encoders in first stage
• A bit of re-organization to the structure of the configs/args
• Preprocessing pipeline cleanups

0.0.2

Some updates from the initial release:

1. Added a lot of documentation!
2. Made preprocessing code cleaner.
3. Added RAG training and inference
4. Added ECG Grounding Dataset

0.0.1

We present the first initial release of ECG-Bench. Enjoy!