This project provides a pipeline to train and evaluate models of C. elegans neural dynamics. The main.py script is the primary entry point for orchestrating the training process, comparing models with and without connectome-based constraints, and generating post-training visualizations.

Data folder is not provided in this repo as it'll be too chunky, you can download from original White and Zimmer paper or ask me for it.

1. Python Environment: Ensure you have a Python environment (e.g., Python 3.8+) set up.
2. Dependencies: Install the necessary Python packages. You can typically do this using a requirements.txt file (if provided) with pip:

   pip install -r requirements.txt

   Key libraries include torch, numpy, pandas, scikit-learn, matplotlib, plotly, and tqdm.
3. Data (not included in the repo to keep it slim):
   • Neural activity data (e.g., IDs.xlsx, traces.xlsx, tracesDif.xlsx, fps.xlsx) should be placed in a directory specified by data_dir in the config (default: data/activity/WT_NoStim/).
   • Connectome data (e.g., connectome_EJ.csv, Celeganconnect.xls) should be in a directory specified by connectome_path (default: data/connectome/White1986/).

Parameters for the training pipeline are primarily configured within the main() function in main.py.

Key parameters to adjust at the beginning of the main() function:

• num_runs: Number of independent training runs.
• num_epochs: Number of training epochs per run.
• training_volume: Number of frames in the training set.
• test_volume: Number of frames in the test set.
• train_gap_junctions: Boolean to determine if gap junction weights in the model should be trainable.

Further detailed hyperparameters are located in the config dictionary within main.py. Users can modify these values directly in the script before running. Important config settings include:

• sheet_name: Specifies which sheet from the neural activity Excel files to use.
• data_dir, connectome_path: Paths to the input data.
• latent_dim: The dimensionality of the latent space for PCA.
• sequence_length, test_seq_length: Length of training and testing sequences.
• n_epochs, learning_rate, weight_decay, lambda_cov, lambda_corr: Hyperparameters for the training process.
• beta_init, tau_init, noise_init, synaptic_gain_init: Initial parameters for the neural dynamics model.
• run_avg_loss_vis, run_avg_pc_vis: Booleans to enable/disable post-training average loss and PC trajectory visualizations.

To run the training pipeline, execute main.py from your terminal:

python minimal_code/main.py