Visual_Neuro_InSilico_Exp

This project serves the purpose of developping and testing computational tools in silico for future visual experiments in vivo.

Structure of the Repo

It has mulitple components, structure of project is the following:

NN Infrastructures
- Loading networks in Caffe backend net_utils
- Loading Generator in Generator
- Loading networks in Torch backend torch_net_utils. (Autograd for Hessian computation requires torch.)
A class of Experimental Objects defined in insilico_Exp
Many types of Zeroth Order Optimizers defined in
- The older one in Optimizer CMA-ES, Cholesky CMA-ES, GA, CholeskyCMAES_Sphere
- The series of optimizer based on the Hessian Aware ZO Optimization paper ZO_HessAware_Optimizers
- The optimizers based on Powell's method PowellOptimizers
Some objects target at estimating and analyzing Hessian for Black Box and non linear functions.
- pytorch_CNN_hessian
- pytorch_GAN_similarity_hessian
- A collection of functions to help analyze hessian matrix and spectrum hessian_analysis
- Analysis code for many experiments hessian_analysis_batch
- A non-gradient Hessian estimation method inspired by the HessAware ZOO paper ZO_Hessian_Estim
Hessian subfolder contains the experimental codes and infrastructures for the Hessian project. For a more structured code base, see this repo.
movie_timescale subfolder contains the code for measuring the response timescale for units across neural network, CNN or Transformers.
Cosine subfolder contains the code for doing Cosine evolution, i.e. evolution with an objective that involves a population.
optimizer_dev subfolder contains scripts for developping and testing some new kinds of optimizers, e.g. BO, QD,

This Repo can be deployed onto cluster to run at large scale, based on torch (preferred) or Caffe or TF backend.

Bash scripts (*.sh,*.bsub) store in cluster_scripts, these will normally call some python scripts with different parameters.
Specialized python scripts (*.py) with command line interface (CLI) will have RIS_cluster or cluster suffix