derekhoward/morphologycnn

MorphologyCNN

Install packages

Install neuron_morphology package: https://neuron-morphology.readthedocs.io/en/latest/install.html

Download data

Gouwens morphology

Download SWC files from ftp://download.brainlib.org:8811/biccn/zeng/pseq/morph/200526/ into gouwens-data/data/

Extracted features: Download extracted features (SpecimenMetadata.csv) from https://knowledge.brain-map.org/data/1HEYEW7GMUKWIQW37BO/specimens

Scala morphology

Download SWC files from https://download.brainimagelibrary.org/3a/88/3a88a7687ab66069/inhibitory/ into scala-data/inhibitory/

View SWC files

3D morphology viewer: https://neuroinformatics.nl/HBP/morphology-viewer/

Analysis steps

1. Run collect_training_images.ipynb to generate .png images from the Gouwens and Scala .swc files. This code will also output some metadata required for copying/moving the files into the correct folders for splitting training/validation/test of the model.

2. Run preprocess.py to downscale images to appropriate sizes

3. the only dataset subfolders you should run for first set of tests is create_t_type_subfolders.py

• if you run other scripts for creating data subfolders, you end up with 21 subtypes in the t-type directory?
• model training code vanilla_cnn_t_type.py requires 20 t-type classes for training to work.

4. Run vanilla_cnn_t_type.py

• saves model to model_path = './gouwens-data/models/vanilla_cnn_gouwens_t_type_modified_b2_lr1e-4_e4_rs7_256_vderek'
• should be able to reload either that model or the previously trained one on github for inference in next step

5. Run inference_t_type.py

• test out both versions of model tried.
• how do scores compare on Scala dataset?