pvrnn_sa

PyTorch code about predictive-coding-inspired variational recurrent neural network model (PV-RNN)

1. System requirements
   -OS: Ubuntu (18.04)
   -Python (3.7.10)
   -Pytorch (1.8.1+cpu)

2. Installation guide
   If you have established Python environment, you may only have to install Pytorch according to instructions on the official website (https://pytorch.org/). As needed, please install python packages such as matplotlib, glob, numpy, yaml, os, sys, and time. The installs may take no longer than half an hour.

3. Demo and Instructions
   The folder structure is below.

“training/“
–”networks/”: Some python codes defining PV-RNN structures and functions (e.g., initialization, forward generation)
–“example/”: Codes for setting network hyperparameters and executing program (training).
・“example_of_training_rnn_part.py”: Main code for running program
・“network_config.yaml”: Hyperparameter setting
・“result_training/”, “trained_model/”: Results are saved here
・“target/”: Target data is placed here (target0000.txt ~ target0023: Self-produced condition, target0024.txt ~ target0047: Externally produced condition)
・“dataset.py”, “utilities.py”: Some supplemental functions read by “example_of_training_rnn_part.py”
・"plot_development.py": Program for generating figure

“test/”
–”networks/”: Some python codes defining PV-RNN structures and functions (e.g., initialization, forward generation)
–“example/”: Codes for setting network parameters and executing program (online inference test).
・“example_of_error_regression.py”: Main code for running program
・“network_config.yaml”: Hyperparameter setting
・“result_ER/”: Results are saved here
・“target/”: Target data (test data) is placed here
・“error_regression.py”, “dataset.py”, “utilities.py”: Some supplemental functions read by “example_of_error_regression.py”
・“generate*.npy”, “model*.npy”: Trained parameters read by “example_of_error_regression.py”
・"plot_timeseries.py": Program for generating figure

3-1. Learning experiment (It may take about one day and a half to train one neural network.)

    cd /training/example/
    python example_of_training_rnn_part.py
    #Trained model and reproduced timeseries will be saved every 5000 learning epoch as 
    #"/training/example/trained_model/model_*.pth" and "/training/example/result_training/generate_*.npy".
    
    #plot_development.py can be used to generate figures (pdf files) of developmental trajectory of sensory attenuation
    #saved in “/training/example/”.
    python plot_development.py

3-2. Test experiment (8 trials by one neural network may take about one hour.)

    #You have to copy “model_00200000.pth” and “generate_00200000.npy” (learning results) into "/test/example/"
    #, although I put an example of the baseline model by default.
    #If you want to set a trained model other than the defalt,
    cp /training/example/trained_model/model_00200000.pth /test/example/
    cp /training/example/result_training/generate_00200000.npy /test/example/
    
    #Test trial run
    cd /test/example/
    python example_of_error_regression.py
    #Timeseries for each trial will be saved in “/test/example/result_ER/sit01/seq01/ite050/window010/lr0090/” 
    #("seq01" can be "seq01"~"seq08", corresponding to different trials).
    
    #plot_timeseries.py can be used to generate figure (pdf file) of timeseries for each trial
    #saved in “/test/example/result_ER/sit01/seq01/ite050/window010/lr0090/” ("seq01" can be "seq01"~"seq08").
    python plot_timeseries.py