NeuRI: Diversifying DNN Generation via Inductive Rule Inference

Welcome to the artifact repository of the NeuRI paper which is accepted by ESEC/FSE 2023.

Bug finding evidence (RQ3)

See links to real-world bug reports.

Get ready for running experiments!

Important

General test-bed requirements

OS: A Linux System with Docker Support;
Hardware: X86/X64 CPU; 16GB RAM; 1TB Storage; Good Network to GitHub and Docker Hub;

Note

Before you start, please make sure you have Docker installed.

To check the installation:

docker --version # Test docker availability
# Docker version 20.10.12, build e91ed5707e

# 💡 Get Docker image from Docker Hub
docker pull ganler/neuri-fse23-ae:latest
# 💡 [Alternative] load image from https://doi.org/10.5281/zenodo.8319975
# 1. Install and "unzip" the package
# 2. Run `docker load --input neuri-image-fse23.tar.gz`

# Run Docker image
docker run -it --name ${USER}-neuri ganler/neuri-fse23-ae:latest
# By using this command, you will "get into" the image like entering a virtual machine.
# The session will be kept under the name "${USER}-neuri"
# So that later on even if you exit the interactive bash, you can come back via:
# `docker start ${USER}-neuri && docker exec -it ${USER}-neuri /bin/bash`

# Inside the image, try to update any timely bug fixes
cd /artifact
git remote set-url origin https://github.com/ise-uiuc/neuri-artifact.git
git pull origin main

Evaluating Coverage (RQ1)

The overall process of NeuRI is shown as:

Note Pre-generated rules: For simplicity, we pre-generated the output of instrumentation and rule inference in ./data directory. To perform instrumentation from scratch, follow instructions here. To perform rule inference, follow the RQ2 section.

S1: Start fuzzing and cache test-cases

We will use ./fuzz.sh to generate the test-cases.

Note

Command usage of: ./fuzz.sh NSIZE METHOD MODEL BACKEND TIME

Arguments:

NSIZE: the number of operators in each generated graph.
METHOD: in ["neuri", "neuri-i", "neuri-r", "symbolic-cinit"].
MODEL: in ["tensorflow", "torch"].
BACKEND: in ["xla", "torchjit"].
TIME: fuzzing time in formats like 4h, 1m, 30s.

Outputs:

$(pwd)/gen/${MODEL}-${METHOD}-n${NSIZE}.models: the generated test-cases (models)

For PyTorch

Note

You may parallelize the scripts below if you have multiple cores, at the risk of potential result unstability. Nonetheless the results in the paper are generated without script parallelization.

source ./env_std.sh
./fuzz.sh 5 neuri          torch torchjit 4h
./fuzz.sh 5 neuri-i        torch torchjit 4h
./fuzz.sh 5 neuri-r        torch torchjit 4h
./fuzz.sh 5 symbolic-cinit torch torchjit 4h # NNSmith
./fuzz.sh 1  neuri         torch torchjit 4h
./fuzz.sh 9  neuri         torch torchjit 4h
./fuzz.sh 13 neuri         torch torchjit 4h

For TensorFlow

source ./env_std.sh
./fuzz.sh 5 neuri          tensorflow xla 4h
./fuzz.sh 5 neuri-i        tensorflow xla 4h
./fuzz.sh 5 neuri-r        tensorflow xla 4h
./fuzz.sh 5 symbolic-cinit tensorflow xla 4h # NNSmith

S2: Collect coverage

For PyTorch

source ./env_cov.sh
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-n5.models          --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-i-n5.models        --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-r-n5.models        --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-symbolic-cinit-n5.models --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-n1.models          --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-n9.models          --model_type torch --backend_type torchjit --parallel $(nproc)
python experiments/evaluate_models.py  --root $(pwd)/gen/torch-neuri-n13.models         --model_type torch --backend_type torchjit --parallel $(nproc)
# Compute coverage
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-n5.models    \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-i-n5.models  \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-r-n5.models  \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-symbolic-cinit-n5.models \
                                       --llvm-config-path $(which llvm-config-14)       \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-n1.models    \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-n9.models    \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)
python experiments/process_profraws.py --root $(pwd)/gen/torch-neuri-n13.models   \
                                       --llvm-config-path $(which llvm-config-14) \
                                       --instrumented-libs "$(pwd)/build/pytorch-cov/build/lib/libtorch_cpu.so" "$(pwd)/build/pytorch-cov/build/lib/libtorch.so" \
                                        --batch-size 1000 --parallel $(nproc)

For TensorFlow

source ./env_cov.sh
python3 experiments/evaluate_tf_models.py --root $(pwd)/gen/tensorflow-neuri-n5.models --parallel $(nproc)
python3 experiments/evaluate_tf_models.py --root $(pwd)/gen/tensorflow-neuri-i-n5.models --parallel $(nproc)
python3 experiments/evaluate_tf_models.py --root $(pwd)/gen/tensorflow-neuri-r-n5.models --parallel $(nproc)
python3 experiments/evaluate_tf_models.py --root $(pwd)/gen/tensorflow-symbolic-cinit-n5.models --parallel $(nproc)
# Compute coverage
python3 experiments/process_lcov.py --root $(pwd)/gen/tensorflow-neuri-n5.models --parallel $(nproc)
python3 experiments/process_lcov.py --root $(pwd)/gen/tensorflow-neuri-i-n5.models --parallel $(nproc)
python3 experiments/process_lcov.py --root $(pwd)/gen/tensorflow-neuri-r-n5.models --parallel $(nproc)
python3 experiments/process_lcov.py --root $(pwd)/gen/tensorflow-symbolic-cinit-n5.models --parallel $(nproc)

S3: Checkout the results

Table 1 & 2

# PyTorch
python experiments/genstat.py --root $(pwd)/gen/torch-neuri-n5
python experiments/genstat.py --root $(pwd)/gen/torch-neuri-i-n5
python experiments/genstat.py --root $(pwd)/gen/torch-neuri-r-n5
python experiments/genstat.py --root $(pwd)/gen/torch-symbolic-cinit-n5

# TensorFlow
python experiments/genstat.py --root $(pwd)/gen/tensorflow-neuri-n5
python experiments/genstat.py --root $(pwd)/gen/tensorflow-neuri-i-n5
python experiments/genstat.py --root $(pwd)/gen/tensorflow-neuri-r-n5
python experiments/genstat.py --root $(pwd)/gen/tensorflow-symbolic-cinit-n5

Check the terminal output for the results.

Figure 6 (a)

python experiments/viz_merged_cov.py --folders                  \
        $(pwd)/gen/tensorflow-symbolic-cinit-n5.models/coverage \
        $(pwd)/gen/tensorflow-neuri-r-n5.models/coverage        \
        $(pwd)/gen/tensorflow-neuri-i-n5.models/coverage        \
        $(pwd)/gen/tensorflow-neuri-n5.models/coverage          \
    --tags '\textsc{NNSmith}' '\textsc{NeuRI}$^r$' '\textsc{NeuRI}$^i$'  '\textsc{NeuRI}'

Check images under ./results/branch_cov-time.png for the results.

Figure 6 (b)

python experiments/viz_merged_cov.py --folders             \
        $(pwd)/gen/torch-symbolic-cinit-n5.models/coverage \
        $(pwd)/gen/torch-neuri-r-n5.models/coverage        \
        $(pwd)/gen/torch-neuri-i-n5.models/coverage        \
        $(pwd)/gen/torch-neuri-n5.models/coverage          \
    --tags '\textsc{NNSmith}' '\textsc{NeuRI}$^r$' '\textsc{NeuRI}$^i$'  '\textsc{NeuRI}'

Check images under ./results/branch_cov-time.png for the results.

Figure 6 (c)

python experiments/viz_merged_cov.py --folders             \
        $(pwd)/gen/torch-neuri-n1.models/coverage          \
        $(pwd)/gen/torch-neuri-n5.models/coverage          \
        $(pwd)/gen/torch-neuri-n9.models/coverage          \
        $(pwd)/gen/torch-neuri-n13.models/coverage         \
    --tags '\#Node 1' '\#Node 5' '\#Node 9' '\#Node 13'

Check images under ./results/branch_cov-time.png for the results.

Evaluating Rule Inference (RQ2)

S1: Tree & rule generation

source ./env_std.sh
python3 neuri/autoinf/inference/tree.py
python3 neuri/autoinf/inference/augmentation.py
python3 neuri/autoinf/inference/shape_solve.py
python3 neuri/autoinf/inference/predicate_solve.py
python3 neuri/autoinf/inference/nnsmith_reuse.py
python3 neuri/autoinf/inference/rule_validity.py
python3 neuri/autoinf/inference/rosette_solve.py

Rules will be stored in gen/.

S2: 1-Node test-cases generation

RULE_DIR=$(pwd)/gen ./fuzz.sh 1 neuri-i   torch      torchjit 4h
RULE_DIR=$(pwd)/gen ./fuzz.sh 1 neuri-i   tensorflow xla      4h

S3: Check the results

Table 3 & 4

python3 table3.py
python3 table4.py

Check the terminal output for the results.

Documentation for reusing NeuRI

Command-line usages
Basic concept of operator rules
Logging and errors
Test-as-documentation: detailed internal APIs usages can be learnt through our test-suite at tests/
Known issues

Learning More

Pre-print:
NeuRI is being merged into NNSmith

bwry / neuri-artifact Goto Github PK

neuri-artifact's Introduction

NeuRI: Diversifying DNN Generation via Inductive Rule Inference

Bug finding evidence (RQ3)

Get ready for running experiments!

Evaluating Coverage (RQ1)

S1: Start fuzzing and cache test-cases

For PyTorch

For TensorFlow

S2: Collect coverage

For PyTorch

For TensorFlow

S3: Checkout the results

Table 1 & 2

Figure 6 (a)

Figure 6 (b)

Figure 6 (c)

Evaluating Rule Inference (RQ2)

S1: Tree & rule generation

S2: 1-Node test-cases generation

S3: Check the results

Table 3 & 4

Documentation for reusing NeuRI

Learning More

neuri-artifact's People

Contributors

Recommend Projects

Recommend Topics

Recommend Org