This repository belongs to the publication ""

Environment setup

conda create -n <env_name> python=3.9

conda activate <env_name>

pip install -r requirements.txt

Data

The data contains of the following:

Natural movie stimulus

• stimulus frames: all frames of the stimulus, down-sampled to 300x350 pixels
• fixation file: file that specify which frame is loaded and the center of the frame
• response file: dictionary containing cell responses for all cells for all frames across all trials
• pre-computed STAs: STAs for each cell precomputed from the white noise stimulus
• training config: config file specifying the response file, the cells and the crop to be used

The expected file structure is the following:

---stimuli
    |--- img_file_0
    |--- ....
    |--- img_file_n
---responses
    |--- config file
    |--- response file dataset1
    |--- response file dataset2
---fixations
    |--- fixation file 
---stas
    |--- sta cell 0
    |--- ...
    |--- sta cell n

White noise stimulus

• training config file for white
• stimulus frames: all frames of the stimulus, down-sampled to 300x350 pixels
• response file: dictionary containing cell responses for all cells for all frames across all trials
• pre-computed STAs: STAs for each cell pre-computed from the white noise stimulus
• training config: config file specifying the response file, the cells and the crop to be used

---non_repeating_stimulus
    |--- trial_001
      |--- all_images.npy
    |--- ....
    |--- trial_n
       | --all_images.npy
---repeating_stimulus
    |--- all_images.npy
---responses (same directory as for nm stimulus)
    |--- config file
    |--- response file dataset1
    |--- response file dataset2
---stas (same directory as for nm stimulus)
    |--- sta cell 0
    |--- ...
    |--- sta cell n

Training

To train an LN model on natural movie, run the ln_model_factorized_marmoset_nm.py To train an LN model on white noise, run the ln_model_factorized.py

For finetuning, run the

• train_multiretinal_ln_wn_for_nm.py for models trained on white noise and to be finetuned on natural movies
• train_multiretinal_ln_nm_for_wn.py for models trained on natural movies to be finetuned on white noise

Evaluation & analysis

Models can be evaluated using evaluations/ln_model_performance.py Sizes are estimated using functions in evaluations/sizes_estimations.py The power spectra of the stimulus, filtered stimulus and filters is in poser_spectrum_analysis.py for rank 1 LN models and power_spectrum_3d.py for rank2 LN models. The analysis of LN model performance and size and is in notebooks/adaptation_paper_figers.ipybn