QTron - ONNX Viewer with Integrated Analysis

A lightweight VS Code extension for viewing and analyzing ONNX models with seamless onnx_tool integration.
Inspired by Netron and ONNX Viewer but optimized for autoaisys workflows.

Why was this developed?

Load Large models faster (we are impatient like that :P); Strip the weights.
Profile the ONNX model (lets make our lives simpler).
I would like to automate mundane tasks in my life. Thus automated the flow.

All this happens with just a click to view an ONNX file—yep, that’s what I’ve been dreaming about lately, and I finally cooked this extension!

NOTE: Loading a model for the first time might take a might more time, but post to which you can happily load the lightweight profiled onnx model.

📁 ONNX File →      🚀 strip_and_profile_model() →         👁️ Visualize
     ↓                            ↓                                 ↓
  Original              [Simplify + Analyze]                   Interactive
   Model                          ↓                              Viewer   
                       📦 Weights Stripped              [LightWeight Profiled Model]
                       Profiled for 
                       MAC, Memory, Params (.onnx)             
                       📄 Reports (.txt, .csv)

Qtron in Action

Qtron

QTron: Weight Stripped/Profiled

Weight Stripped/Profiled ONNX

Profiling Info Layer Wise Profiling Data

Simplified View Simplified ONNX

✨ Key Features

🎯 One-Click Analysis: Simply open any .onnx file in VS Code

Automatic model simplification for cleaner visualization
Integrated onnx_tool profiling with detailed metrics

📊 Comprehensive Analysis:

Layer-wise MAC (Multiply-Accumulate) operations counting
Parameter count and memory size analysis
Export to TXT and CSV formats for further processing
Shape-only ONNX files for efficient analysis
NEW: Advanced dynamic shape handling for models with variable inputs

⚡ Smart Defaults:

Results saved next to your ONNX file automatically
Graceful fallback when advanced tools unavailable
Intelligent handling of dynamic batch sizes and sequence lengths
Configurable through VS Code settings

🔧 Configuration

Access settings via VS Code preferences (qtron.*):

Setting	Description	Default
`qtron.pythonPath`	Python interpreter path	`"python"`
`qtron.enableSimplification`	Enable ONNX simplification	`true`
`qtron.enableOnnxToolProfiling`	Enable onnx_tool profiling	`true`
`qtron.onnxToolResultsPath`	Analysis results directory	Auto-detected
`qtron.enableDynamicShapeHandling`	Smart handling of dynamic input shapes	`true`

📁 Analysis Output

When onnx_tool profiling is enabled, you'll get organized results:

your_model_directory/
├── your_model.onnx                    # Original model
└── onnx_analysis_results/
    └── your_model/
        ├── your_model_profile.txt      # Human-readable summary
        ├── your_model_profile.csv      # Detailed metrics (Excel-ready)
        └── your_model_shapes_only.onnx # Shape-optimized model

🔄 Processing Workflow

QTron follows an optimized processing pipeline:

Integrated Workflow (onnx_tool enabled):

📁 ONNX File →      🚀 strip_and_profile_model()     → 👁️ Visualize
     ↓                        ↓                            ↓
  Original            [Simplify + Analyze]             Interactive
   Model                      ↓                           Viewer                              
                        📄 Reports (.txt, .csv)  [ LightWeight Profiled Model]
                        📦 Weights Stripped/
                        Profiled Model (.onnx)

Fallback Workflow (simplification only):

📁 ONNX File → 🔧 Simplify → 👁️ Visualize
     ↓              ↓          ↓
  Original      Simplified  Interactive
   Model         Model      Viewer

Optimization Benefits:

✅ Single-step processing eliminates duplicate simplification
✅ Automatic fallback handling for maximum compatibility
✅ Preserves original files while generating analysis

📦 Installation

Build:

git clone https://gitlab.qualcomm.com/sathyapr/qtron
cd onnx_viewer
./regen.sh

Installation

# If you are feeling lazy download the .vsix I have genearted
./extension/qtron-*.vsix

🚀 Usage

Open any .onnx file in VS Code
QTron automatically processes the model:
- Simplifies for better visualization
- Runs comprehensive analysis (if available)
- Opens interactive viewer
Find analysis results in onnx_analysis_results/ directory

💡 Contributing

Have suggestions or found a bug? We'd love to hear from you!

Open an issue on GitHub
Submit a pull request
Try to fix it on your own or ping me on Teams and buy me a coffee, we can debug together :P

QTron

satabios