DEFFE: Data-Efficient Framework for Exploration

This repository has the sources for DEFFE framework, which is intended for design space exploration with exploration and machine learning based prediction capabilities. The technical details of DEFFE is explained in the [reference, see below][1]. The state of the art design space exploration tools evaluate the design points (samples) and identify the optimal design points based on the cost metrics. However, the evaluation of design points are time consuming and may require heavy computation for some problems. DEFFE will help for such problems.

DEFFE's machine learning model tries to learn from the evaluated design points. Once after learning for some time, DEFFE's machine learning model inference can predict the cost metrics with ~98% accuracy, which can be used for fast design space exploration. It needs less than 5% of the samples to get the near-accurate machine learning prediction model and can save huge time in the design space exploration.

DEFFE framework is implemented fully in python and it is configured through json file. The json file has configuration of problem with parameters, cost metrics, evaluate procedure, prediction method, sampling technique and exploration algorithm. It provides great flexibility to users to add their custom python modules for the above tasks.

Reference

[1]: Frank Liu, Narasinga Rao Miniskar, Dwaipayan Chakraborty and Jeffrey S. Vetter, "Deffe: a data-efficient framework for performance characterization in domain-specific computing", ACM International Conference on Computing Frontiers (CF 2020), May 2020, Catania, Italy, https://doi.org/10.1145/3387902.3392633

DEFFE framework component diagram

The above figures shows the main blocks and their corresponding python files in the DEFFE framework.

Hardware Dependencies

1. GPU to train the machine learning model
2. SLURM environment if to run evaluate (RISCV-GEM5) simulations
   • Installation of GEM5 is given in http://learning.gem5.org/book/part1/building.html

Software Dependencies

1. Python-3
2. Graphviz
3. Python packages
   • keras
   • tensorflow
   • tensorflow-gpu
   • torch
   • doepy
   • scikit-learn
   • xlsxwriter
   • matplotlib
   • pandas
   • pathlib
   • pydot
   • tqdm
   • multiprocess
   • torchsummary
   • jsoncomment

   Install all required packages from requirements.txt file in the repository.
   $ pip3 install -r requirements.txt
   
   Or
   
   $ python3 -m pip install keras tensorflow torch doepy scikit-learn xlsxwriter matplotlib pandas pathlib pydot tqdm torchsummary jsoncomment multiprocess

Docker support

Docker support is provided for DEFFE. Set proxy settings in Docker file and also in apt.conf file appropriately.

$ make -f Makefile.docker build ; 
$ make -f Makefile.docker run ;

How to run DEFFE?

An example DEFFE configuration for RISCV design space exploration along with their associated files are placed in example directory.

$ pip install -e .
$ cd example ; 
$ run_deffe -config config_small.json
$ cd .. ;

A bare minimal simple two parameters and one application test case is available in test directory.

$ pip install -e .
$ cd test; 
$ sh run_deffe.sh
$ cd .. ;

An example DEFFE configuration for RISCV design space exploration (without slurm) along with their associated files are placed in example directory.

$ pip install -e .
$ cd example ; 
$ run_deffe -config config_small.json -no-slurm
$ cd .. ;

The run_deffe.py file can show all command line options with the below command.

$ pip install -e .
$ run_deffe -h
$ cd .. ;

To run the exploration on the preloaded data

$ pip install -e .
$ cd example ;
$ run_deffe -config config_small.json \
   -only-preloaded-data-exploration -step-start 0 -step-end 1 \
   -epochs 100 -batch-size 256 
$ cd .. ;

To run full exploration (all samples at once training, means no transfer learning across samples) on the pre-evaluated/pre-loaded data

$ pip install -e .
$ cd example ;
$ run_deffe -config config_small.json \
  -only-preloaded-data-exploration -full-exploration -train-test-split 0.7 \
  -validation-split 0.23 -step-start 0 -step-end 1 -epochs 100 -batch-size 256 
$ cd .. ;

Installation with pip

pip install git+ssh://git@code.ornl.gov/miniskarnr/deffe.git

How to run experiments?

Run DEFFE with K-means with pre-evaluated complete dataset given to model train

* Dataset: examples/output_kmeans_deffe.csv
* Config file: examples/config_kmeans.json
* Loss functions: custom_mean_abs_log_loss (default) or custom_mean_abs_exp_loss 
*  (exponential loss function)
* Run directory: example/experiments/full_explore/log/kmeans   (For log loss function)
                 example/experiments/full_explore/exp/kmeans   (For exponential loss function)
* Command to run: 
    $ pip install -e .
    $ cd example/experiments/full_explore/log/kmeans 
    $ run_deffe \
    -config $DEFFE_DIR/example/config_kmeans.json \
    -only-preloaded-data-exploration -epochs 20000 -batch-size 4096 \
    -full-exploration \
    -train-test-split 0.7 -validation-split 0.23 -loss custom_mean_abs_log_loss

* Command to generate stats. It will load the same training and testing indices used for ML model 
    $ run_deffe \
    -model-extract-dir checkpoints -config $DEFFE_DIR/example/config_kmeans.json \
    -only-preloaded-data-exploration -train-test-split 0.7 \
    -validation-split 0.23 -load-train-test -loss custom_mean_abs_exp_loss \
    -model-stats-output test-output-exploss.csv

* Output files: 
    ** Intermediate checkpoint files directory: example/experiments/full_explore/log/kmeans/checkpoints
    ** Training and Test indexes used: step<int>-train-indices.npy, step<int>-val-indices.npy, 
    ** which have training and validation indexes used for training for that step
    ** Output statistics in file: test-output-exploss.csv in the format (Epoch, TrainLoss, 
    ** ValLoss, TestLoss, Step, TrainCount, ValCount)
* Try sample parameters:
    ** Input test-input.csv
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_kmeans.json \
       -input test-model.csv -icp kmeans.hdf5 -output output-prediction.csv -inference-only
    ** Output test-output.csv
    ** Input ../../../../output_kmeans_deffe.csv 
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_kmeans.json \
       -icp kmeans.hdf5  -input ../../../../output_kmeans_deffe.csv  \
       -output test-output-full.csv \ -inference-only
    ** Output test-output-full.csv

Run DEFFE with K-means with pre-evaluated data set but passed as set of samples and enabled transfer learning across samples

* Dataset: examples/output_kmeans_deffe.csv
* Config file: examples/config_kmeans_tl_samples.json
* Loss functions: custom_mean_abs_log_loss (default) or custom_mean_abs_exp_loss 
* (exponential loss function)
* Run directory: example/experiments/transfer_learning_samples/log/kmeans   (For log loss function)
                 example/experiments/transfer_learning_samples/exp/kmeans   (For exponential loss function)
* Command to run: 
    $ pip install -e .
    $ cd example/experiments/transfer_learning_samples/log/kmeans 
    $ run_deffe \
    -config $DEFFE_DIR/example/config_kmeans_tl_samples.json \
    -only-preloaded-data-exploration -epochs 1000 -batch-size 256 \
    -train-test-split 1.0 -validation-split 0.23 

* Command to generate stats. It will load the same training and testing indices used for ML model 
    $ run_deffe \
    -model-extract-dir checkpoints \
    -config $DEFFE_DIR/example/config_kmeans.json  \
    -only-preloaded-data-exploration \
    -train-test-split 1.0 -validation-split 0.23 \
    -load-train-test -loss custom_mean_abs_exp_loss \
    -model-stats-output test-output-exploss.csv
    $ run_deffe \
    -model-extract-dir checkpoints \
    -config $DEFFE_DIR/example/config_kmeans.json  \
    -only-preloaded-data-exploration \
    -train-test-split 1.0 -validation-split 0.23 \
    -load-train-test -loss custom_mean_abs_log_loss \
    -model-stats-output test-output-logloss.csv

* Output files: 
    ** Intermediate checkpoint files directory: 
    **    example/experiments/transfer_learning_samples/log/kmeans/checkpoints
    ** Training and Test indexes used: step<int>-train-indices.npy, step<int>-val-indices.npy, 
    ** which have training and validation indexes used for training for that step
    ** Output statistics in file: test-output.csv in the format (Epoch, TrainLoss, ValLoss, 
    ** TestLoss, Step, TrainCount, ValCount)
* Try sample parameters:
    ** Input test-input.csv
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_kmeans_tl_samples.json \
       -input test-model.csv \
       -icp kmeans.hdf5 -output output-prediction.csv -inference-only
    ** Output test-output.csv
    ** Input ../../../../output_kmeans_deffe.csv 
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_kmeans_tl_samples.json -icp kmeans.hdf5  \
       -input ../../../../output_kmeans_deffe.csv  -output test-output-full.csv \
       -inference-only
    ** Output test-output-full.csv

Run DEFFE with Matmul with pre-evaluated data set but passed as set of samples and enabled transfer learning across samples and also enabled transfer learning from kmeans.

* Dataset: examples/output_matmul_deffe.csv
* Frozen layers: 2 convolution layers
* Config file: examples/config_matmul_tl_samples.json
* Loss functions: custom_mean_abs_log_loss (default) or custom_mean_abs_exp_loss 
* (exponential loss function)
* Run directory: example/experiments/transfer_learning_samples_across_kernels/log/matmul 
*                               (For log loss function)
*                example/experiments/transfer_learning_samples_across_kernels/exp/matmul  
*                               (For exponential loss function)
* Command to run: 
    $ pip install -e .
    $ cd example/experiments/transfer_learning_samples_across_kernels/log/matmul 
    $ run_deffe \
    -config $DEFFE_DIR/example/config_matmul_tl_samples.json -icp ../../kmeans.hdf5 \
    -only-preloaded-data-exploration -epochs 1000 -batch-size 256 -train-test-split 1.0 \
    -validation-split 0.23 

* Command to generate stats. It will load the same training and testing indices used for ML model 
    $ run_deffe \
    -model-extract-dir checkpoints \
    -config $DEFFE_DIR/example/config_matmul.json  \
    -only-preloaded-data-exploration \
    -train-test-split 1.0 -validation-split 0.23 -load-train-test \
    -loss custom_mean_abs_exp_loss -model-stats-output test-output-exploss.csv

    $ run_deffe \
    -config $DEFFE_DIR/example/config_matmul.json  -only-preloaded-data-exploration \
    -train-test-split 1.0 -validation-split 0.23 -load-train-test \
    -loss custom_mean_abs_log_loss -model-stats-output test-output-logloss.csv

* Output files: 
    ** Intermediate checkpoint files directory: 
    **     example/experiments/transfer_learning_samples_across_kernels/log/matmul/checkpoints
    ** Training and Test indexes used: step<int>-train-indices.npy, step<int>-val-indices.npy, 
    ** which have training and validation indexes used for training for that step
    ** Output statistics in file: test-output.csv in the format (Epoch, TrainLoss, ValLoss,
    ** TestLoss, Step, TrainCount, ValCount)
* Try sample parameters:
    ** Input test-input.csv
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_matmul_tl_samples.json -icp matmul.hdf5 \
       -input test-input.csv -output test-output.csv -inference-only
    ** Output test-output.csv
    ** Input ../../../../output_matmul_deffe.csv 
    ** Command given below
       $ run_deffe \
       -config $DEFFE_DIR/example/config_matmul_tl_samples.json -icp matmul.hdf5  \
       -input ../../../../output_matmul_deffe.csv  -output test-output-full.csv \
       -inference-only
    ** Output test-output-full.csv

Bash scripts in the below directories to run above configurations

example/experiments/transfer_learning_samples/exp/kmeans/run.sh
example/experiments/transfer_learning_samples/log/kmeans/run.sh
example/experiments/transfer_learning_samples_across_kernels/exp/matmul/run.sh
example/experiments/transfer_learning_samples_across_kernels/log/matmul/run.sh
example/experiments/full_explore/exp/kmeans/run.sh
example/experiments/full_explore/log/kmeans/run.sh

How to extend DEFFE?

All classes in DEFFE have the below bare minimal methods.

import os
import argparse
import shlex

class DeffeExploration:
    def __init__(self, framework):
        self.config = framework.config.GetExploration()
        self.framework = framework

    # Initialize the members
    def Initialize(self):
        None

    def Run(self):
        None

def GetObject(framework):
    obj = DeffeExploration(framework)
    return obj

The "GetObject" method returns the object of the class. It should take the DEFFE framework object as an input to configure the class. The class will have "Initialize" and "Run" method.

How to configure JSON configuration?

An example JSON configuration file parameters are shown in short given below.

{
    "python_path" : ["."],
    "knobs" : 
        [
            ### Array of Knobs (Architecture / Application Knobs) ###
        ],
    "scenarios" : 
        [
            ### Array of Scenarios (Application scenarios) ###
        ],
    "costs" : 
        [ 
            ### Array of Cost metric names ###
        ],
    "model" : {
            ### ML-Model JSON key-value configuration parameters ###
    },
    "exploration" : {
            ### Exploration (DSE) module json key-value configuration parameters ###
    },
    "sampling" : {
            ### Intelligent sampling (Random) module JSON key-value configuration parameters ###
    },
    "evaluate" : {
            ### Evaluate (Sampling points evaluation) module json key-value configuration parameters ###
    },
    "extract" : {
            ### Extract (evaluation output extraction) module json key-value configuration parameters ###
    },
    "framework" : {
            ### Framework  module json key-value configuration parameters ###
    },
    "slurm" : {
            ### Slurm json key-value configuration parameters ###
    }
}

DEFFE Convenience Environment Variables

These environment variables aid in the writing of evaluate scripts.

Variable	Description
DEFFE_EXP_DIR	The path to the experiment directory from which run_deffe was called.
DEFFE_CONFIG_DIR	The path to the configuration directory where the json file specified by -config is located.

To use these environment variables in evaluate.sh include the following snipit.

: "${DEFFE_EXP_DIR:=$PWD}"
: "${DEFFE_CONFIG_DIR:=$PWD}"
echo "*********************** Evaluate.sh *********************"
echo "DEFFE_EXP_DIR    = $DEFFE_EXP_DIR"
echo "DEFFE_CONFIG_DIR = $DEFFE_CONFIG_DIR"
echo "*********************************************************"