VisuTVM

TVM Relay IR Visualization Tool (TVM 可视化工具)

Introduction

Visualize the TVM's Relay IR structure, and support the visualization of Pass optimized Relay IR.

👉note：In the example(b), the nodes with the same color are fused.

Usage

# add env path
export PYTHONPATH=$PYTHONPATH:${your-path}/VisuTVM

# visu relay ir(default: FuseOps)
python main.py -bp relay_ir/example_fo_bp.txt -ap relay_ir/example_fo_ap.txt -sn example

# visu relay ir with tensor info
python main.py -bp relay_ir/example_fo_bp.txt -ap relay_ir/example_fo_ap.txt -sn example -wi

# if you only have one relay ir file(before pass file or after pass file), 
# you can run the following command
# input unoptimized relay ir file(before pass file)
python main.py -ri relay_ir/mobilenet_v2_fo_bp.txt -sn mobilenet_v2

# input optimized relay ir file(after pass file) and specify the pass name
python main.py -ri relay_ir/resnet18_all_pass.txt -pn AllPass -sn resnet18 -wi

# create relay ir txt file(depend on TVM environment)
python examples/example.py --passname FuseOps

Installation

• Step1: install graphviz

• Step2: install graphviz's python API

  pip install graphviz

Supported Pass

• FuseOps
• RemoveUnusedFunctions(no case)
• ToBasicBlockNormalForm(no case)
• EliminateCommonSubexpr
• FoldConstant
• SimplifyInference
• CombineParallelConv2D
• CombineParallelDense
• CombineParallelBatchMatmul
• FoldScaleAxis
• SimplifyExpr
• CanonicalizeCast
• CanonicalizeOps(no case)
• FlattenAtrousConv
• FastMath
• ConvertLayout
• MergeComposite

Preview

🚀️ EliminateCommonSubexpr

🚀️ FoldConstant

🚀️ SimplifyInference

🚀️ CombineParallelConv2D

🚀️ FlattenAtrousConv

🚀️ All Pass with tensor info

Q&A

在使用过程中遇到可视化失败的Relay IR网络结构，可以在 issues 上提出你的问题，如果有任何好的想法，也可以进行交流哦👏👏👏