Introduction

The repo contains Cupy, C and CHP implementation of the Nwave Dynamics Model. We wrote notes about our research, please see: https://drive.google.com/file/d/16BQkJAv_0oqrUk0mwHLh7jCzGl8-j1bD/view?usp=sharing.

Cupy implementation aims to code the model in GPU environment. Our experiment is on GCP (Google Cloud Platform).

C implementation aims to be the first step that converting Cupy code to FPGA. Moreover, C code verifies the feasibility of message-passing style algorithm, which is better suited to alternative forms of parallelization.

CHP implementation is our ultimate goal. The CHP language is created by Prof. Rajit Manohar, and the language fits well in FPGA environment.

Directory Structure

• nwave/ : implementation of nwave algorithms.

  • dictlearner.py : update rule of feedforward weights
  • neurodynamics.py : update rule of activations
  • wave.py : combine dictlearner and neurodynamics
  • utils.py : store many functions that will be used in other files

• c_version/ : convert the program to c

  • makefile: to compile the c code
  • compute_act_laplacian.c: compute activations using laplacian matrix
  • compute_act_msg_passing: compute activations using message passing algorithm
  • compute_RC.c: compute receptive field using trained codebook
  • matrix_float.c: implement vector multiplication and matrix multiplication in pure c
  • msg_passing_algorithm.c: implement message passing algorithm in C
  • plot_RC.ipynb: plot receptive field given data from compute_RC.c
  • read_csv.c: read data from csv files
  • read_parameter.c: read hyperparameters in csv files
  • test_between_C_and_python.ipynb: compare the results between C version and python version
  • train_laplacian.c: implement training steps using laplacian matrix in C
  • train_msg_passing.c: implement training steps using message passing algorithm in C

• FPGA_version/ : convert the program to FPGA environment

  For ACT language, refer to https://avlsi.csl.yale.edu/act/doku.php?id=language:start.

  • nn7.act: CHP code of neuron activation process, using message-passing algorithm to better fit the parallelization
  • support.c: Since CHP doesn't have exponential function, we define the weights in support.c, and call the function in nn7.act. Moroever, we writie a function to read in the .csv file in support.c, and call the function in nn7.act
  • x.conf: configuration file
  • compile.sh: compile support.c in order to make it compatible to nn7.act file

• variables_record/ : record and analyze all the values during the computation

  • record_variable_values.py: record all the values. Index are from random sampling the batch; value are from the computations
  • analyze_variable_values.py: analyze and make histogram plots

• experiment.py : interface to experiment, including training, plotting receptive fields and plotting activations

• dataloader.py : load data

• configs.py : read all the hyperparameters

• compute_whole_activ.py: compute whole "activity.npy", which contains the activations of over 50,000 words

• compute_act_convolve.py: Don't compute whole activity file. Given some words, plot corresponding activations using spicy.convolve

• compute_act_laplacian.py: Don't compute whole activity file. Given some words, plot corresponding activations using multiplication of laplacian matrix and activations

• compute_act_msg_passing.py: Don't compute whole activity file. Given some words, compute and plot activations using message passing algorithm

• matrix.py: implement vector multiplication and matrix multiplication in pure python

• train_convolve.py: standalone python file of training process, and using convolution

• train_laplacian.py: standalone python file of training process, and using laplacian matrix

• train_msg_passing.py: standalone python file of training process, and using message passing algorithm

start

1. Run the code in GPU environment

To run the project locally, run python experiment.py --row [test number] --processor [CPU/GPU] , and all the results will be stored in the result folder.

To run the project on gcloud, these are the steps to start and stop an instance on GCP:

Install gcloud cli https://cloud.google.com/sdk/docs/install

Login to gcloud using netid: gcloud auth login

Set project: gcloud config set project jl2994-neuralnet-project-aa1f

Start instance: gcloud compute instances start --zone "us-east4-b" "rob-test"

SSH to instance: gcloud compute ssh --zone "us-east4-b" "rob-test" --tunnel-through-iap --project "jl2994-nerualnet-project-aa1f"

Stop instance gcloud compute instances stop --zone "us-east4-b" "rob-test"

2. Run the code in FPGA environment

First install ACT environment, refer to https://avlsi.csl.yale.edu/act/doku.php?id=install.

Under FPGA_version/ directory, run the following commands:

1. ./compile.sh

2. actsim -cnf=x.conf -Wlang_subst:off nn7.act 'grid<40,3,5>'