PUMA Functional Simulator

The functional simulator models different components of matrix operations execution with ReRam backend.

System requirements

Requirement	Version
Anaconda	4.5.11
Python	3.6.6 or higher
Tensorflow-gpu	tested on 1.10.0

Quick Start (for installing tensorflow-gpu in conda)

conda create --name tf-gpu python=3.6
source activate tf-gpu
conda install -c anaconda tensorflow-gpu

Download CIFAR-100 dataset and convert it to tfrecords files (train and test sets).

Usage

Data Loader data_loader.py in repo has been tested for CIFAR-100 dataset, should be extensible to other datasets with changing the parameters such as number of classes, training and testing example etc.

The training and testing scripts use vgg16 models by default - ``vgg16_puma.py```.

Note: this version runs on 1 GPU only.

Training

For training with PANTHER operations use ifpanther=True.

python train_puma.py --dataset=<my_path>

Testing

python test.py --dataset=<my_path>

Tensorboard (see results, training progress or debugging)

Tensorboard helps in visualization of several statistics collected during the training run (accuracy, loss etc).

tensorboard --logdir=<my_name>:<my_logpath>

Measuring runtime and memory consumption of operations in tensorboard

Uncomment Lines 185-187, Line 194 in train_puma.py (and comment Lines 188 and 195) to dump related metadata during training. Launch tensorboard, choose run-step on left-pane, choose compute time or memory from the checklist.

Authors

Aayush Ankit (Purdue University), Sai Rahul Chalamalasetti (Hewlett Packard Labs)