Dev deeplearning Integration (nnLM and new models)

[mahmoud1yaser]

Notes

This pull request introduces major enhancements to the AFID detection workflow, adding support for a new nnLandmark (nnLM) detection mode, improving configuration flexibility, and optimizing inference for both GPU and CPU environments. The changes enable users to select between three detection strategies (prior-based, sliding-window without prior, and nnLM single-pass), provide more granular control over inference parameters, and restructure the Snakemake rules to support both parallel and sequential inference workflows.

Key changes include:

Detection Mode and Configuration Enhancements:

• Added new command-line arguments and configuration options in snakebids.yml to select detection mode (--detect_with_prior, --detect_without_prior, --detect_with_nnlm), and to configure nnLM-specific parameters (fold, plans, checkpoint, device), as well as sliding-window inference overlap and batch size.
• Introduced a new nnLM model resource URL and a detailed afids_inference section for per-AFID checkpoint configuration, patch size, device, and overlap.
• Added enable_sequential_inference config option to control whether inference is run sequentially (recommended for GPU) or in parallel (recommended for CPU).

Workflow and Rule Structure:

• Refactored the Snakemake workflow in Snakefile to dynamically include the appropriate rules (cnn.smk or nnlm.smk) based on the selected detection mode, and to automatically enable sequential inference for GPU.
• Updated the rule all input and descriptor logic to select the correct FCSV output based on the detection mode, ensuring output files are labeled appropriately.

CNN Inference Rule Improvements:

• Overhauled cnn.smk rules to support both sequential (single job for all AFIDs) and parallel (one job per AFID) inference for both prior-based and no-prior detection, including new gather rules to combine per-AFID outputs. [1] [2]
• Added logic to use the correct environment (pytorch.yaml) and improved parameter passing for model checkpoints and inference settings.

Environment and Dependency Updates:

• Added new Conda environment files for PyTorch (pytorch.yaml) and nnLM (nnlm.yaml), specifying compatible Python and library versions and including the nnLandmark package for nnLM inference. [1] [2]

Quality-of-Life Improvements:

• Suppressed extraneous warnings for unrecognized BIDS entities and explicitly set the BIDS spec version in the workflow.

These changes make the workflow more flexible, scalable, and ready for integration with the new nnLandmark model while maintaining backward compatibility with existing CNN-based modes.